An AI layer on top of your ERP
We design the intelligence layer that sits on the ERP and CRM you already own — reading your data continuously, forecasting what lands next, and arriving with a recommended action, its confidence, and the name of the person who approves it.
None of these numbers are ours. They are the published industry record, and they are the reason we do not sell you another platform — we build the layer that makes the one you have decide.
Average ERP cost overrun in manufacturing — the worst of any industry.
source ↗Discrete manufacturing ERP projects that fail to meet their original objectives.
source ↗ERP projects that exceed their original budget, per Panorama and Gartner.
source ↗Overruns caused by understaffing alone. Scope expansion adds 35%, data issues 34%.
source ↗Nobody at your Monday meeting needs another view of a closed month. They need to know what next month looks like, how confident that is, which lever moves it, and what it is worth to pull. A report is a rear-view mirror with a very good polish. A model is a windscreen.
What a report can do
What a decision layer does
Every vendor on your shortlist can show you what happened. These are the twelve moves the layer makes on its own, before anyone asks — each with the manual work it removes and a figure behind it. Not because the ERP is bad at its job, but because every one of these needs data the ERP does not hold and a decision it was never asked to make. That list is the whole argument.
Forecasts the next nine weeks per product family and territory, with the model’s own uncertainty attached.
ReplacesA forecast so aggregated it hides which lines are actually moving.
Bias improved to +2.1% — fourth consecutive week inside band. A single-line forecast hides the thing you most need to know.
Demand agent · Analyst and Operations
Prices each line on each quote from your live cost model — material, freight, tooling, currency — with that customer’s contracted terms and volume breaks already applied.
ReplacesQuoting from last quarter’s cost sheet and finding out at month end.
Anything below your margin floor cannot be sent at all — it routes for a second approval.
Pricing intelligence · Finance and Sales
Watches days of cover per item per location against its own reorder point and proposes the replenishment before the shelf empties.
ReplacesNoticing a stock-out from a customer’s phone call.
Dallas DC at 5 days against a 7-day reorder point — 3,400 units clears it and protects the San Antonio push.
Inventory agent · Supply and Logistics
Re-plans the week against your real constraint set — changeover matrices, certified operators, curing times, inbound ASNs.
ReplacesA planner rebuilding the schedule by hand every time a coil slips.
One changeover of 1.5 hours to save 14 hours of idle: 12.5 net hours, two customer commitments protected.
Production agent · Operations
Ranks overdue accounts by what is actually recoverable and drafts the approach for each one.
ReplacesA dunning letter to a distributor whose warehouse is full.
$216K overdue, 71% of it in three accounts — all three at the bottom of sell-through. It is an inventory problem wearing a cash problem’s clothes.
Finance agent · Finance
Finds the runs that belong together across customers and rebuilds them inside every delivery window.
ReplacesPlanning fourteen orders as fourteen loads because they belong to fourteen accounts, not because they would not travel together.
14 loads at 62% fill became 9 multi-stop runs at 89% — $8,400 in a week, 2,100 fewer miles, 1.9t less CO₂.
Route agent · Logistics
Translates a supplier slip into which customer commitments are exposed, in currency, days before it lands.
ReplacesFinding out a shortage mattered on the morning it mattered.
Two exceptions, one cause: a cover shortfall and an expansion plan competing for the same 3,400 units, held by different owners.
Compliance and Demand agents · CXO
Builds the quote, the proposal, the RFP response and the contract summary from your own live data and governed content.
ReplacesA week of an expensive person’s time spent on retrieval and formatting.
Industry benchmarks: 45% average RFP win rate, 10–20% improvement reported by AI-assisted teams, 30–40% faster turnaround.
Sales and Executive agents · Commercial
Chooses which plant, DC or 3PL fills each line on the order — weighing freight, capacity, inventory risk, promise date and the margin on that specific line, not just the nearest warehouse.
ReplacesOne sourcing rule applied to every line on the order, written three years ago.
Order management platforms optimise fulfilment across speed, shipping cost, labour, inventory risk and margin simultaneously. Your ERP allocates against a static rule and calls it sourcing.
Route and Inventory agents · Logistics and Sales
Answers “when can we have it?” line by line against real capacity — machine hours, tooling, certified labour, inbound materials — instead of against a stock figure.
ReplacesAvailable-to-promise arithmetic that ignores whether you can actually make it.
ATP checks stock; capable-to-promise checks operating hours, machine availability, labour, tooling and materials. Even when the answer is no, you get a defensible alternative date — and buyers respect a data-backed commitment more than an optimistic one.
Production and Demand agents · Operations and Sales
Watches every line from quote to cash — configured price, contracted terms, rebate accrual, freight recovery, deductions and the final invoice — and flags where they stop agreeing.
ReplacesFinding a pricing error at month end, in the variance.
Published benchmarks put quote-to-invoice mismatch at 5–15% and revenue leakage from disconnected quote-to-cash at 1–5% of EBITDA, with 30–40% of sales-ops capacity spent reconciling it. Automated quoting cuts cycle times 30–50% and pricing errors by over 90%.
Finance and Sales agents · Commercial
Spots phantom inventory — system stock with no sales — by comparing each store against its own comparable neighbours rather than against an average.
ReplacesA chain-wide stock count to find nine stores.
Six consecutive days of zero sales against positive system stock, in stores whose neighbours kept selling. Nine stores explain most of the availability gap; counting all of them costs four days of labour.
Inventory agent · Retail and Distribution
Nothing on this list executes on its own. The agent drafts, ranks and explains; a named human approves, and the whole thing is logged, explainable and reversible. That is not a limitation we apologise for — it is the reason these systems survive their second year.
Every line below is a decision the system reaches on its own and hands to a person with the reasoning attached, at whatever grain the decision needs — a plant, a territory, an account, a line item. Against each domain is the outcome range published for that class of work: third-party research, named, so you can check it rather than take it from us.
The business, one quarter ahead
Published supply-chain research puts AI-driven forecasting at 20–50% lower forecast error, with lost sales cut by up to 65% and administration cost down 25–40%.
McKinsey-derived supply-chain benchmarksEvery account with a next action attached
Automated quoting cuts quote cycle times 30–50% and pricing errors by over 90%. On bids the average win rate is 45%, and AI-assisted teams report 10–20% improvement with 30–40% faster turnaround.
CPQ and RFP benchmarks, 2026The data your ERP has never seen
This is the gap almost no ERP fills: syndicated market data — NielsenIQ, Circana, SPINS, retailer portals — lives outside every system of record. Meanwhile around 60% of CPG trade promotions return less than they cost, 72% in the United States, and you cannot see which without that data alongside your own.
Trade promotion effectiveness research, 2026Make what earns, keep the line running
Sites that land predictive maintenance properly see 30–60% less unplanned downtime on targeted equipment and 25–40% lower maintenance cost, with payback typically at 12–24 months.
Predictive maintenance benchmarks, 2026Right stock, right place, before the shelf empties
Out-of-stocks cost retail roughly 8% of sales and nearer 10% during promotions. AI forecasting is documented to cut inventory 20–50% and safety stock 20–30% without raising that risk, with warehousing cost down 5–10%.
On-shelf availability and forecasting researchSee the disruption while it is still a rumour
From this page’s own worked example: 14 loads at 62% trailer fill rebuilt into 9 multi-stop runs at 89% — $8,400 in one week, 2,100 fewer miles, every delivery window still met.
Worked example, Smart Route agentCash you can see coming
AI-driven collections are reported to cut DSO by 8–25 days and automate 60–80% of routine receivable touches inside 90 days; top performers run over 90% touchless. Quote-to-invoice mismatch runs 5–15%, and leakage from disconnected quote-to-cash at 1–5% of EBITDA.
AR and quote-to-cash benchmarks, 2026Buy better, and know what you signed
Early adopters report 15–30% cost savings and 40–60% less manual processing time; leading teams sustain 8–12% annual savings against total spend. On the risk side, AI vision inspection reaches 95–99% detection accuracy with around 37% fewer defects.
Procurement AI and quality benchmarks, 2026The week your best people lose to documents
Teams with a governed content library reuse about 66% of proposal content and report 25-hour responses falling under 5. Sixty-eight per cent of proposal teams now use AI somewhere in the process.
RFP and proposal research, 2026Every range above is published third-party research for that class of work. None of it is a forecast for your plant, and we will not quote you one before seeing your data — a vendor who does is guessing with your money. What these figures establish is that the problem is worth solving and roughly what solving it has been worth to others. The first two weeks of any engagement are a data reality check that measures your actual baseline; from then on the only number that matters is your delta against it.
This page is long on purpose — a CEO and a single-store owner need almost nothing in common. Jump to yours.
A CEO who opens a dashboard of two hundred KPIs learns nothing. Five material exceptions, each with its cause, its exposure, its owner and a recommended action, is a different instrument entirely.
Use cases we build first
What materially changed since you last looked, ranked by financial exposure rather than by position on a page. Everything else stays one click away and out of your morning.
Every figure links down to the operating event that moved it — a changeover, an expedite, a scheme, a supplier slip. Board questions get answered in the meeting.
Model a price change, a capacity addition, a channel shift or a lost distributor against the real constraint set, before it becomes a committed decision.
When commercial, plant and finance quote three different numbers, the argument is about data lineage. We make that argument resolvable in minutes rather than weeks.
Which categories, channels, territories and customers actually earn their capital — with the cost to serve included, not averaged away.
What was recommended, who approved it, what was expected, and what it delivered. Institutional memory that survives people leaving.
What the CXO dashboard opens on
What the system escalated to the CEO this week
| Metric | Definition | The question it actually answers | Cadence |
|---|---|---|---|
| Revenue at risk | Committed orders threatened by supply, credit or capacity. | What could go wrong before month-end, in currency. | Daily |
| Contribution by unit | Margin after cost to serve, by channel, territory and category. | Which parts of the business fund the rest. | Weekly |
| Decision latency | Hours from signal detected to action approved. | How fast the organisation actually moves, measured. | Weekly |
| Forecast reliability | Signed bias and error, by unit, over time. | Whether the numbers you plan on can be trusted. | Monthly |
| Capital efficiency | Return on inventory and capacity employed. | Whether growth is being bought or earned. | Monthly |
| Concentration risk | Revenue and margin share in the top customers and SKUs. | How exposed you are to losing one relationship. | Monthly |
The analyst is the most misused role in a manufacturer. Hired to find insight, spent on assembling reports other people could have self-served. We move them from report-builder to decision partner — which is also the fastest way to prove the platform paid for itself.
Use cases we build first
The analyst curates definitions — what "active outlet", "on-time" and "margin" mean — once, centrally. Every dashboard and every AI answer then agrees, because they resolve to the same model.
They set what counts as an exception, at what level, for whom, with what escalation. This is judgement work that cannot be automated and should not be guessed by a vendor.
The system surfaces the anomaly. The analyst explains why it happened and what to do — which is the part that actually needs a human brain.
Profiling results, contract breaches, reconciliation failures and freshness slips land on a queue they own, with lineage to trace any disagreement to its source.
Certified datasets, documented measures and a natural-language layer that refuses to answer what the data cannot support. Self-service without governance is just faster disagreement.
Every recurring request becomes a governed asset. The measure of success is that the queue shrinks and nobody notices the analyst is no longer in it.
What the analyst works from
The analyst queue, this morning
| Today | Why it is expensive | After | Cadence |
|---|---|---|---|
| Builds the same weekly deck | Skilled time on assembly, not thinking. | Deck generates itself; analyst writes the commentary. | Weekly |
| Fields ad-hoc requests | Queue grows with headcount, forever. | Governed self-service; queue becomes exceptions only. | Daily |
| Reconciles conflicting numbers | Days lost to whose spreadsheet is right. | One semantic model; lineage settles it in minutes. | Daily |
| Finds problems by looking | Depends on who happened to check. | Thresholds surface them; analyst diagnoses. | Live |
| Owns knowledge personally | It leaves when they leave. | Definitions and logic live in the model. | — |
| Cannot prove value | Reporting is a cost centre by default. | Decision latency and forecast accuracy are measurable. | Monthly |
Service, cost, margin and changeover pull against each other. A recommendation that does not show what it sacrificed is not a recommendation — it is an instruction you cannot argue with.
Use cases we build first
The proposed sequence with its trade-offs visible: what it protects, what it delays, what it costs, and the two alternatives it was chosen over.
Your real bottleneck — a specific furnace, a curing time, one certified operator — modelled explicitly instead of assumed away by textbook MRP.
A supplier slip translated into which customer commitments are exposed, in currency, days before it lands.
Available-to-promise that reflects reality: planned downtime, changeover, labour availability and quality yield, not nameplate capacity.
When supply is short, your rule decides who gets it — margin, contract, relationship, strategic priority — applied consistently rather than by whoever calls loudest.
Condition monitoring, vibration and micro-stop patterns turned into a maintenance recommendation with a date on it — plus the spares to have on hand, the technician certified to do it, and the production slot it should occupy. Reactive maintenance is the most expensive kind because you pay for it twice: once in the repair and once in the schedule it destroyed.
Commercial forecast, production plan and financial plan tied to one another, with the variance between them named rather than averaged away.
Availability, performance and quality separated — because a 62% OEE caused by micro-stops needs a completely different intervention from one caused by scrap.
Failure probability from vibration, temperature, cycle count and repair history, with cost-of-inaction weighed against cost-of-downtime.
SPC on the characteristics that matter, defect Pareto by line, shift and operator, and lot genealogy in both directions for recall.
What the operations dashboard opens on
What the plant was told before the shift started
| Metric | Definition | Why it earns its place | Cadence |
|---|---|---|---|
| Schedule adherence | Orders produced to plan ÷ orders planned. | Low adherence makes every downstream promise fiction, however good the plan was. | Daily |
| OTIF | Delivered on time and in full ÷ total orders. | The customer's definition of whether you are good at your job. | Daily |
| Changeover cost | Lost minutes × contribution rate per switch. | Turns "we should batch more" from an opinion into a number. | Daily |
| Available-to-promise | Sellable capacity after downtime, changeover and yield. | What sales can safely commit — the honest version, not nameplate. | Hourly |
| Expedite spend | Premium freight and overtime caused by replanning. | The hidden cost of poor forward visibility, made visible. | Weekly |
| Plan stability | How many times an order moved before it ran. | Churn in the plan is a leading indicator of chaos downstream. | Weekly |
| OEE | Availability × performance × quality. | Only useful decomposed — the headline number hides which of the three is hurting. | Hourly |
| Micro-stop frequency | Stops under five minutes, by cause. | Individually invisible, collectively the largest hidden capacity loss in most plants. | Live |
| First-pass yield | Units right first time ÷ units started. | Rework is capacity you already paid for and are now spending twice. | Hourly |
A margin decline is not a finance event. It is a changeover decision, an expedite, a discount, a route, a scheme — made weeks earlier by someone who never saw the financial consequence.
Use cases we build first
By product, customer, channel, route, plant and shift — down to the changeover that caused it. Variance with a name attached, not a bucket.
Inventory ageing, receivable ageing, payable timing and the cash impact of every planning decision before it is approved.
Accounts ranked by amount, probability of recovery and relationship risk, so effort goes where it converts rather than where it is loudest.
A forecast that reflects approved operating actions rather than last quarter's assumptions, updated when reality changes rather than when the calendar does.
Variance fires when it crosses your rule, routed by role and seniority, with escalation if unacknowledged. Nobody has to open a report to find out.
Duty drawback, export incentives, local-content and government concession claims assembled from production data as evidence packs, not reconstructed from memory.
What the executive dashboard opens on
Margin, decomposed to its operating cause
| Metric | Definition | The operating cause we tie it to | Cadence |
|---|---|---|---|
| Gross margin variance | Actual versus standard, decomposed. | Changeover, expedite freight, scrap, discount, mix shift. | Daily |
| Cash conversion cycle | Inventory days + receivable days − payable days. | Batch sizing, allocation policy, claim settlement speed. | Weekly |
| Cost to serve | Fully-loaded cost per outlet or per order. | Route design, drop size, visit frequency, order fragmentation. | Monthly |
| Inventory ageing | Stock value by age bucket and location. | Forecast error, sell-in targets, planogram non-compliance. | Weekly |
| Trade spend ROI | Incremental margin ÷ scheme cost. | Scheme design, timing against stock, outlet selection. | Monthly |
| Forecast bias | Signed error over time, not absolute error. | Persistent optimism in a territory or a product family. | Monthly |
Onboarding, compliance, maintenance and utilisation for vehicles, drivers and loads — modelled together, because a driver out of hours and a trailer out of service are the same problem wearing different clothes.
Use cases we build first
Asset register with VIN, plate, class, capacity, axle configuration, telematics device, insurance, registration, annual inspection date and depreciation schedule. Onboarded once, referenced everywhere.
The qualification file as a workflow, not a folder: licence class and endorsements, medical certificate expiry, MVR pull, testing history, training records and prior employment checks.
Weight and volume limits, axle loading, stacking rules, temperature, hazmat segregation, multi-drop sequencing, and the customer window that actually governs the plan.
Hours-of-service exposure, electronic DVIRs, IFTA mileage by jurisdiction, inspection outcomes and how each feeds your safety score — surfaced before the audit, not during it.
Preventive schedules by mileage, hours and calendar; defect trends by component; parts availability; and the cost of deferring service weighed against an out-of-service event.
Cost per mile and per drop, empty-mile percentage, dwell at each stop, fuel burn against benchmark, on-time delivery, and utilisation by asset and driver.
What the fleet dashboard opens on
Fleet exposure, before it becomes a violation
| Requirement | What it actually demands | Source |
|---|---|---|
| Hours of service | ELD-captured duty status. Violations can reach $19,277 per infraction for carriers, and HOS violations rose from 410,000 in 2023 to over 500,000 in 2025. | source ↗ |
| ELD records | Duty status, engine diagnostics and DVIR entries retained a minimum of six months under 49 CFR 395.8(k)(1). | source ↗ |
| Electronic DVIR | Explicitly authorised by FMCSA rule effective 23 March 2026. Must carry required signatures, route to recipients, and be retained three months. | source ↗ |
| Safety scoring | CSA overhauled: BASICs renamed to Compliance Categories, severity simplified to 1 or 2, 2,000+ codes consolidated to roughly 100 groups, Vehicle Maintenance split in two. | source ↗ |
| IFTA fuel tax | ELDs already capture GPS and state mileage, so returns can be auto-populated instead of reconstructed from receipts. | source ↗ |
| Audit readiness | Auditors examine qualification files, HOS with six months of ELD data, DVIRs for three months, testing records, maintenance and annual inspections, authority and insurance. | source ↗ |
US federal figures; equivalents exist in every jurisdiction we work in. We build the system that keeps the records straight — we are not your compliance counsel.
A rep who spends the morning deciding where to go, then promises a date the plant cannot hit, has had a bad day before 10am. Both problems are data problems.
Use cases we build first
The week is built from revenue potential, urgency, credit status, stock availability and real drive time — not from a beat plan written two years ago and never revisited.
Before the promise, not after. The rep sees whether production and stock can actually hit the date, so the commitment made in the room is one the factory can keep.
Grounded in margin, current stock position and this outlet's own buying history. Not the product with the biggest incentive attached this quarter.
Outstanding balance, ageing and limit visible at the point of order. Nobody discovers a credit block after the truck is loaded.
Yearly, quarterly, monthly, weekly and daily targets against live actuals, at rep, territory and product grain — with the gap explained, not just displayed.
Which reps convert, which lose at which stage, which need range training rather than pressure. Managed on evidence instead of instinct.
Balance territories by potential and drive time rather than history, model the revenue effect before you redraw, and keep the trend line intact through the change.
Quota set from territory potential rather than last year plus ten percent, and a forecast whose bias is measured by person so optimism becomes coachable.
What the sales dashboard opens on
What changed in the field this week
| Metric | Definition | Why it earns its place | Cadence |
|---|---|---|---|
| Selling time ratio | Hours in front of a customer ÷ total working hours. | The single most improvable input in field sales, and almost nobody measures it. | Daily |
| Strike rate | Orders placed ÷ outlets visited. | Separates a coverage problem from a conversion problem — different fixes entirely. | Daily |
| Lines per order | Distinct SKUs per order. | Range penetration. Rising volume on falling lines means concentration risk. | Daily |
| Promise accuracy | Delivered-on-date ÷ dates promised. | The number that decides whether a customer believes your rep next quarter. | Daily |
| Drop size trend | Average order value, same outlet, versus its own history. | Catches a quietly declining account before it churns. | Weekly |
| Cost of coverage | Fully-loaded cost to serve ÷ margin generated, per outlet. | Some accounts cost more to visit than they return. Worth knowing which. | Monthly |
A distributor who takes stock is not the same as a distributor who moves it. Plan production against sell-in and you will build exactly the wrong things, confidently, for years.
Use cases we build first
What actually left the distributor's warehouse, by SKU and by outlet. The gap against sell-in is the most valuable number a manufacturer can see.
Suggested reorder per SKU built from that distributor's own sell-through, seasonality and current cover — with the reasoning shown so it can be overridden intelligently.
Live availability, allocation position and realistic lead time. No more phoning three people to find out whether an order can be filled.
Trade spend tracked to the outlet, claim status visible to the distributor, disputes resolved from records instead of relationships.
Distributor, sub-dealer, outlet — modelled as a real hierarchy, so coverage and penetration mean something at every level.
Agreement, margin structure, warehouse setup, opening balance, portal access and their stock feed integrated — one flow, days not months.
What the distributor dashboard opens on
Sell-in against sell-through, by distributor
| Metric | Definition | Why it earns its place | Cadence |
|---|---|---|---|
| Sell-in vs sell-through gap | Primary dispatch minus secondary sales, indexed. | A widening gap means you are financing a warehouse, not growing a market. | Weekly |
| Days of cover | Distributor stock ÷ average daily secondary sales. | The only honest stock-out predictor. National inventory tells you nothing. | Daily |
| Numeric distribution | Outlets stocking the SKU ÷ outlets in the universe. | Separates a demand problem from an availability problem. | Weekly |
| Order-to-pattern deviation | This order versus this distributor's own ordering history. | Catches both a stocking problem and a diversion problem early. | Daily |
| Claim cycle time | Days from claim raised to settled. | The most reliable predictor of distributor goodwill, and the easiest to fix. | Weekly |
| Return on trade spend | Incremental margin ÷ scheme cost, per scheme per outlet. | Most trade spend is renewed on habit. This makes it a decision. | Monthly |
A chain needs comparison and transfer logic. An independent owner needs three answers and no training. Same data underneath, two completely different products on top — and getting that wrong is why most channel rollouts stall.
Use cases we build first
Performance normalised for footfall, catchment, size and format — so a small high-street outlet is not judged against a highway superstore.
Per-store, per-SKU reorder from that store's own velocity and shelf capacity, not a chain-wide rule applied uniformly and wrongly.
Move stock from where it is not selling to where it is, before it becomes a markdown. The cheapest inventory correction available.
What should be on the shelf against what is, with the revenue cost of the difference — measured, not photographed and forgotten.
What is selling, what to order, what is owed. No navigation, no configuration, no dashboard-building. If it needs a manual it has already failed.
Suggested order from their own sales history, adjustable, submitted from a phone in a stockroom with one bar of signal.
"These six lines earn most of your money. These four have not moved in a month." Not a chart they have to interpret.
What is outstanding and when it is due, visible before ordering rather than discovered at delivery. Fewer awkward conversations for everyone.
Two views, one model
Store-level reality, normalised
| Metric or principle | Definition | Why it earns its place | Cadence |
|---|---|---|---|
| Like-for-like growth | Same-store sales versus its own prior period. | The only growth number not flattered by new openings. | Weekly |
| On-shelf availability | Must-stock lines present ÷ lines listed. | The most common cause of a lost sale, and the least measured. | Daily |
| Sales per square foot | Revenue ÷ selling area, by category. | Space is the scarcest asset in retail. This is its yield. | Weekly |
| Transfer avoidance value | Markdown avoided by moving stock instead of discounting. | Makes the case for transfer logistics in currency. | Weekly |
| No report building | Independents have no analyst and no interest in becoming one. | Three fixed answers on one screen, chosen for them. | Daily |
| No online requirement | Stockrooms have bad signal. Every time. | Offline-first, syncing when it can. | Live |
Every manufacturer we meet has the same shortlist. These are the ones with a measurable baseline, a provable delta and a payback you can defend in a board paper — which is why we start with one of them rather than a platform.
Everything you need to know about the build in one picture: the systems you already pay for stay exactly where they are, we reach them through their own interfaces, and the intelligence is assembled above them. Nothing below the layer gets replaced.
Vendor names are set in type rather than reproduced as logos — we are not a reseller and we do not imply a partnership we have not been granted. The layer is vendor-neutral by design: it reads whatever you run, including nothing at all.
Download the diagram (SVG) ↓Five pillars, one engagement, one team. Most manufacturers buy these from five different suppliers and spend the difference on making them talk to each other. Open a pillar to see what actually ships.
| # | Pillar | What it does | Deliverables |
|---|---|---|---|
| 01 | Enterprise FoundationOrganise. Connect. Standardise. | Everything begins here. The typical manufacturer has an ERP, a CRM, accounting, inventory, production, pricing, CAD, drawings, product PDFs, SharePoint, a decade of email and a great deal of tribal knowledge — and four different answers to how many active SKUs it sells. None of it disagrees on purpose; it disagrees because nothing ever reconciled it. We unify all of it into one trusted source of business knowledge. | 18open |
|
Core layer
Product and content
Surfaces
Goal One trusted source of business knowledge. | |||
| 02 | Market Growth EngineBe found. Be trusted. Be recommended. | Marketing is no longer only Google. A growing share of buyers ask an assistant before they open a search page, and an assistant returns one answer with two or three sources rather than ten links. Visibility now has to be built in both places at once — and the things that move an answer engine are not the things that moved a results page. | 22open |
|
Findability
Content
Channels
Goal When someone asks an AI who to buy from, your name is in the answer. | |||
| 03 | Revenue AccelerationTurn interest into customers. | Marketing generates attention; sales generates revenue. This pillar is the commercial engine between the two — and most of it is document work that a small group of expensive people currently do by hand against immovable deadlines. | 20open |
|
Pipeline
Documents
Retention
Goal Every lead has a next best action, and every document assembles itself. | |||
| 04 | Decision IntelligenceYour ERP should tell you what is next. | This is the part that makes the rest worth owning. We do not replace the ERP — we build intelligence above it. The ERP tells you what happened; the layer tells you what should happen next, how confident it is, and what the action is worth. Every role receives recommendations, predictions, alerts and next best actions in the language of their own job. | 28open |
|
Planning and demand
Commercial
Supply and logistics
Role intelligence
Goal Recommendations, predictions, alerts and next best actions — per role. | |||
| 05 | Trust, Governance and ScaleAI with human control. | This is where most AI projects fail, which is precisely why it is worth being good at. An agent that can act without a person is an agent that will eventually act at 3am against the wrong customer. Everything automated ends at a named human with a button — and everything that human approves is logged, explainable and reversible. | 26open |
|
Control
AI governance
Security and compliance
Adoption
Goal AI that executives trust, employees adopt, customers appreciate and auditors understand. | |||
| Full catalogue across all five pillars | 114deliverables | ||
Not every engagement includes every line — this is the full catalogue, not a mandatory scope. Most start with one pillar and add the next once the first is earning. Deliverable counts are computed from the list, not rounded up.
Eight screens from the system we build around your data — one for each person who has to act on it. Same data model underneath, a different surface and a different argument per role. Each one carries the AI rail: what the model noticed, what it suggests, what it connected, and the button that acts on it.
Planned · Actual · L3 behind at 74%
$216K total, ranked by size rather than by department
Distributor secondary sales is the constraint on every downstream model
Darker means more systems disagreeing. Three definitions converged this quarter.
| Sev | Exception | Owner | Age | Agent has |
|---|---|---|---|---|
| P1 | Sell-through gap · Dist 27 | Trade | 2h 10m | Allocation hold drafted |
| P1 | Credit exposure · 3 accounts | Finance | 4h 02m | Collections list ranked |
| P2 | Cover below reorder · Dallas DC | Supply | 6h 44m | Replenishment proposed |
| P2 | Batch certificate missing · Lot 88-114 | Quality | 11h | Chased supplier twice |
| P3 | Price list drift · 9 SKUs | Commercial | 1d 3h | Diff ready for review |
| P3 | Duplicate outlet records · 22 | Data | 2d | Merge proposed, needs sign-off |
Micro-stops, not breakdowns, account for most of the gap
Line 2 Thursday is the single unresolved constraint in the draft
Mix did the work. Freight ate a third of it back.
Current · 30 · 60 · 90+ — distribution carries most of the tail
Houston is 29% above network average — and it is not distance
Inside the band all quarter — the dip is normal variation, not a trend
Where the bars diverge you are filling a warehouse, not selling
0–30 · 31–60 · 61–90 · 90+ — Dist 27 has 690 units over 60 days old
Every point of availability is worth roughly 0.9pt of like-for-like here
Waco has been flat for eight weeks — a range issue, confirmed by the compliance scan
Layouts are from patterns we have shipped, populated with the same illustrative figures used throughout this page. Your build is designed to your process and your identity — this is the standard of finish, not a template you would be handed.
Coverage, growth and untapped opportunity for every state. Click any territory for its position, its constraint and the action the system recommends. Figures are representative — the structure is what your data would fill.
National position
Click any state. Twelve are modelled in detail; the rest sit below the coverage threshold where territory guidance becomes unreliable — which is itself a finding.
Tile cartogram: one equal square per state. Chosen deliberately over an outline map, because when the variable is sales opportunity rather than land area, Rhode Island should not be a pixel and Montana should not dominate.
Every one of these renders from live data in production. Shown here with representative figures so you can judge the thinking, not the numbers.
Monthly units, one SKU family. The band is the model's stated uncertainty — a single-line forecast hides the thing you most need to know.
Actual stops at August. September is projection only.
| Month | Actual | Forecast | Low | High |
|---|---|---|---|---|
| Jan | 820 | 800 | 760 | 845 |
| Feb | 910 | 890 | 840 | 940 |
| Mar | 870 | 900 | 850 | 950 |
| Apr | 1040 | 1010 | 950 | 1070 |
| May | 1120 | 1090 | 1020 | 1160 |
| Jun | 990 | 1030 | 960 | 1100 |
| Jul | 1180 | 1160 | 1080 | 1240 |
| Aug | 1260 | 1240 | 1150 | 1330 |
| Sep | — | 1330 | 1200 | 1460 |
Coverage against growth, sized by untapped opportunity. The top-left quadrant is where your next quarter is — high growth, low coverage, real money on the table.
Bubble area is proportional to opportunity value, not radius — a bubble twice as wide is four times the money.
| Territory | Coverage % | Growth % | Opportunity | Action |
|---|---|---|---|---|
| North TX | 82 | 14 | $1,250,000 | grow |
| Gulf Coast | 61 | 22 | $2,100,000 | grow |
| Midwest | 91 | 4 | $640,000 | hold |
| Southeast | 44 | 19 | $1,750,000 | grow |
| West | 73 | -3 | $880,000 | fix |
| Mountain | 38 | 8 | $520,000 | watch |
| Northeast | 86 | -8 | $1,100,000 | fix |
| Houston | 55 | 27 | $1,900,000 | grow |
| Oklahoma | 29 | 11 | $430,000 | watch |
The red line is the reorder point. Two locations are already below it — this chart exists so nobody has to notice that manually.
| Location | Days of cover | Below reorder point |
|---|---|---|
| Plant DC | 21 | No |
| Dallas | 14 | No |
| Houston | 11 | No |
| Dist 04 | 9 | No |
| Dist 12 | 5 | Yes |
| Dist 27 | 3 | Yes |
| Chain A | 12 | No |
| Chain B | 8 | No |
Where the two diverge, you are not selling — you are filling a warehouse. This is the single most useful chart a manufacturer can be shown.
Indexed to sell-in = 100. Dist 27 has taken stock it is not moving.
| Distributor | Sell-in | Sell-through | Gap |
|---|---|---|---|
| Dist 04 | 100 | 94 | 6 |
| Dist 09 | 100 | 88 | 12 |
| Dist 12 | 100 | 61 | 39 |
| Dist 18 | 100 | 79 | 21 |
| Dist 27 | 100 | 43 | 57 |
| Dist 31 | 100 | 91 | 9 |
Revenue share and margin share rarely match. The channel that looks biggest is often not the one paying for the factory.
| Channel | Revenue share % | Margin share % | Difference |
|---|---|---|---|
| Distributors | 46 | 31 | -15 |
| Retail chains | 28 | 24 | -4 |
| Single stores | 14 | 22 | +8 |
| Direct / online | 12 | 23 | +11 |
Faster is not always better — it is better only where a faster decision exists. Streaming a monthly planning input is spend without return.
| Cadence | Supports | Still cannot do | Index |
|---|---|---|---|
| Monthly | Closed-period accounting, board reporting, annual planning | Anything operational. By the time you see it, the month is spent. | 20 |
| Weekly | Trade schemes, territory reviews, replenishment cycles, S&OP | Stock-outs inside the week. Production resequencing. | 45 |
| Daily | Production sequencing, allocation, collections priority, route planning | Intraday breakdowns, shift-level quality drift. | 70 |
| Hourly | Shift performance, quality drift, warehouse throughput, order promising | True machine-level anomaly detection. | 88 |
| Live | Fleet position and HOS exposure, POS transactions, machine telemetry, order feasibility | Nothing operational — but it costs more, so only stream what earns it. | 100 |
Rendered as inline SVG — no charting library, no CDN call, and every chart carries a hidden data table so screen readers and AI crawlers read the numbers rather than see a picture. Power BI, Looker or your own front end can render the same models if you prefer.
Share of revenue against share of contribution, once trade spend and cost to serve are attributed back to the channel that caused them. The order changes — which is the point. Your largest channel is your third most profitable.
Contribution is gross margin less trade and promotional spend less cost to serve (freight, warehousing, returns, service and, for direct-to-consumer, acquisition). Figures are illustrative and internally consistent with the rest of this page. Cost-to-serve studies on distribution and manufacturing books routinely find 20–30% of customers unprofitable once those costs are attributed; the point of the chart is the reordering, not the numbers.
| Channel | Revenue $M | Share of revenue | Gross margin % | Trade and promotion $M | Cost to serve $M | Contribution $M | Share of contribution | Contribution margin % | Difference, points |
|---|---|---|---|---|---|---|---|---|---|
| National retail chains | 42.0 | 34.3% | 26.0 | 5.5 | 2.1 | 3.4 | 19.2% | 8.0% | −15.1 |
| Distributors | 31.5 | 25.7% | 22.5 | 1.3 | 0.8 | 5.0 | 28.4% | 15.8% | +2.7 |
| Direct and key accounts | 18.4 | 15.0% | 34.0 | 0.6 | 1.3 | 4.4 | 25.2% | 24.0% | +10.2 |
| Foodservice | 12.6 | 10.3% | 28.5 | 0.4 | 0.8 | 2.4 | 13.7% | 19.0% | +3.4 |
| eCommerce and DTC | 9.2 | 7.5% | 41.0 | 1.4 | 0.9 | 1.5 | 8.4% | 16.0% | +0.9 |
| Private label | 8.9 | 7.3% | 14.5 | 0.1 | 0.3 | 0.9 | 5.1% | 10.0% | −2.2 |
Six stops on one run, scored while the truck is still moving. Fill, cost per mile, on-time and detention are the four numbers that decide whether today was efficient — and the layer proposes the change rather than reporting the miss.
A schematic network rather than a tiled map: on a marketing page a real map is a network call and a licence. In your build this is a live map against your own telematics, with the same panel beside it. Figures illustrative.
| Stop | Location | Status | Timing |
|---|---|---|---|
| 1 | Plano DC | Delivered | 06:40 · loaded 89% |
| 2 | Chain A · Frisco | Delivered | 08:05 · delivered |
| 3 | Chain A · Allen | On site | On site now · 12 min |
| 4 | Dist 04 · Garland | Scheduled | ETA 11:20 · window 11:00–13:00 |
| 5 | Chain B · Mesquite | Window at risk | ETA 14:35 · window closes 14:00 |
| 6 | Return · Plano DC | Scheduled | ETA 16:10 |
| Measure | Now | Movement |
|---|---|---|
| Trailer fill | 89% | ▲ 27 pts since consolidation |
| Cost per mile | $2.41 | ▼ $0.09 vs last month |
| On-time in full | 92.6% | ▼ 1.4 — two lanes dragging |
| Detention today | 41 min | ▲ one site, one customer |
This is the architecture behind every engagement. Click a layer to see what goes in it, what we use, and where your existing ERP and CRM sit — which is usually not where vendors tell you.
No layer is a black box and no layer is locked. Each one can be replaced independently — which is the practical definition of owning your stack rather than renting it.
Most lost revenue in manufacturing is not lost in the pitch. It is lost weeks earlier, in a stock position nobody was watching, at an outlet nobody visited.
Preventing the deficit
Securing the sale
Online ordering — revenue without a visit
A B2B ordering portal for distributors, chains and single stores that removes the physical visit from the transaction — and removes the phone call, the WhatsApp photo of a handwritten list, and the re-keying error that follows it.
Every new store, rep, territory, brand, category and SKU is a workflow with approvals, dependencies, documents and a measurable time-to-productive. Modelling it that way is the difference between scaling and coping.
Channel and people
Portfolio
What it looks like in practice — brand, category and product onboarding is the same engine, run at a different grain.
An agent that can do anything is an agent nobody trusts with anything. Each of these has a defined scope, a data context, an explicit approval boundary and a measurable outcome. Click any agent to see exactly what it does and what it is forbidden from doing.
Every manufacturer has a small group of expensive people whose week disappears into documents. The work is real, the deadlines are immovable, and almost none of it is creative — it is retrieval, assembly, checking and formatting. That is the exact shape of work a language model does well, provided a human still signs.
These are hard-coded refusals, not settings someone can quietly switch off in year two.
Your ERP holds the money, the recipes and the customer list. A layer that reads all of it has to be governed harder than the systems underneath, not more loosely. This is the control set we build in from the first sprint, not the one we retrofit before a security review.
The boring controls, done properly and evidenced.
Because "the model said so" is not a defence.
Including PHI, where the rules are strictest.
Regulatory positions verified July 2026. We are a software studio, not your counsel or your auditor — we build the system that makes an audit survivable and keep your advisers in the loop, but interpretation stays with them.
In controlled and certified categories, documentation is not overhead — it is your licence to sell, your access to government tenders, and your defence in an audit. We build it as data with owners and expiry dates, not as a folder somebody maintains.
Four sectors. Open the one you are in — each is a summary of the actual obligation and what we build against it.
The FDA Food Traceability Rule turns recall response from a phone-call exercise into a data obligation with a 24-hour clock on it. Most food manufacturers can technically comply today — by hand, slowly, and only if the right person is not on holiday.
February 2026 changed the shape of device quality compliance in the United States, and the systems most manufacturers run were designed against the old shape.
Composite reinforcement sold into public infrastructure lives or dies on approved-product-list status, and that status is a records problem as much as a manufacturing one.
State track-and-trace is not a reporting nicety — it is the condition of your licence, and it reports to the regulator whether your data is right or not.
Two regulators, two clocks. Federal excise on a semimonthly cycle, monthly operational reporting, and a separate product-authorisation regime.
Every certificate, permit, test result and licence carries an owner, an expiry and an escalation path. Nothing depends on someone remembering.
From raw material lot to the shipment that left the gate, in both directions, fast enough to be useful during a recall rather than after it.
The evidence bundle a government buyer or auditor asks for, assembled on demand from live records.
Every filing, renewal, audit window and test interval on one calendar with lead-time alerts sized to how long the work actually takes.
Where a regulator holds a parallel copy of your data, continuous reconciliation catches the mismatch while it is still fixable.
Government concessions, duty drawback, export incentives and local-content claims need evidence assembled from production data. We generate the pack rather than reconstruct it.
Scope note. We are a software and data studio, not a law firm or a certifying body. We build the systems that keep records complete, current and auditable. Regulatory interpretation stays with your counsel and your certifier — and rules differ by jurisdiction and change often.
Seven tiers of people, four sources of data, one canonical model. Signal travels down as instruction and back up as evidence — which is what makes the forecast believable.
CEO / CXO
Business strategy
Five exceptions instead of two hundred KPIs. Each one linked to its operating cause, its financial exposure, its owner and the recommended action.
Analyst executive
From report-builder to decision partner
Stops rebuilding the same spreadsheet. Curates the semantic model, sets thresholds, investigates the anomalies the system already surfaced.
Operations manager
Operational decisions and monitoring
Suggestive manufacturing cycle: what to run next, on which line, with what material, at what margin, against which customer commitment.
Sales executive / rep
Execution support
The day arrives pre-planned — ranked accounts, an efficient route, order feasibility checked before the promise, next-best-offer in hand.
Distributor
Executives, coordinators, purchase heads
AI-suggested reorder, live allocation visibility, scheme and claim clarity, and their own sell-through instead of guesswork.
Retail chain
Multi-store owners and managers
Store-versus-store performance, planogram compliance, replenishment by outlet, and demand transfer between locations.
Single store
Small and independent owners
One screen: what sells, what to order, what is owed. Simple enough to be used without training.
POS data
Transaction-level truth
Basket, refund and discount data ingested at transaction grain — the earliest honest demand signal a manufacturer can get.
RIS data
Retail Information System
Sales tracking, instant multi-store inventory counts, employee and financial analytics, and the bridge from checkout hardware to back-end systems.
Third-party data
External context
Syndicated market data, weather, commodity and freight indices, competitor pricing, tariff and holiday calendars.
Your own database
Where orchestration starts
ERP, MES, WMS, spreadsheets, email, calls, legacy SQL. Profiled, cleaned, mapped, mastered and given one canonical model.
The same team that builds your data layer runs your demand side, on the same numbers. Brand, content, design, motion, campaign operations, lifecycle email and outbound — measured against orders, not impressions.
Channel names are set in type rather than as logos — we do not reproduce other companies' marks, and a wordmark says the same thing without implying an endorsement.
Buyers increasingly ask an assistant before they open a search page, and an assistant returns one answer with two or three sources rather than ten blue links. The uncomfortable part for anyone who has spent a decade on SEO: ranking and being cited are now different games. Roughly 83% of AI Overview citations come from pages outside the organic top ten, and a meaningful share of the most-cited pages in ChatGPT have no organic visibility at all.
Pick a rhythm to see the volume behind it. These are the numbers we run for an industrial manufacturer with one marketing lead and us behind them — not agency maximums designed to justify a retainer.
Anyone can be pleased at go-live. The test is whether the thing still gets opened, still gets trusted, and still gets funded twelve months later.
We stopped arguing about whose number was right and started arguing about what to do. That took about six weeks and it changed how the Monday meeting runs.
The forecast is not perfect. What changed is that it now tells us how wrong it might be, so we plan the range instead of pretending the number is the truth.
We kept the ERP. That was the part I did not believe until it happened — nobody on the floor had to learn a new screen.
Two of our distributors were sitting on stock they were never going to move. We could see it in week three. That one report paid for the build.
Placeholder quotes and blank client marks. We do not publish a client name, logo or quote without written approval — swap the strings and the sample chips disappear.
Filter by what you came here for. Every answer stays on the page whichever filter is on, so you can search the lot with ctrl-F — and so can the assistants your buyers ask.
A separate application that reads your ERP, CRM and operational systems continuously, holds its own governed data model, and produces forecasts, exceptions and recommended actions. It writes back only through supported interfaces, and only when a person approves. Your ERP keeps doing what it is good at — being the system of record. The layer does the part the ERP was never designed for: deciding what happens next.
Three ways. It sees data your ERP cannot — EPOS feeds, distributor sell-through, telematics, spreadsheets that never made it into a system. It is priced as a build rather than per seat, so it does not get more expensive as you grow. And you own it, which means the model logic, the thresholds and the definitions are yours to change without raising a ticket.
The meeting stops being an argument about whose number is right. Exceptions are already ranked by money, each one carries a drafted action with its expected value, and the person who owns it is named. The measure of success is boring: fewer meetings about data, more decisions per meeting.
No, and usually you should not. Your ERP is the system of record and it is good at that. We read from it, never modify its core, and build the prediction, AI and reporting layer alongside — so your upgrade path stays clean. Replacing the ERP only makes sense when its licence cost has outgrown its usefulness, and even then we would move you to an open-source foundation rather than another seat contract.
Yes. If your processes live in spreadsheets, email and a legacy database, we can build the canonical data model first and put purpose-built applications on top of it. An open-source foundation such as ERPNext or Odoo Community covers the transactional ledger with zero licence fee, and everything above it is yours. That path costs more up front and nothing per seat afterwards.
SAP, Oracle, Dynamics 365, NetSuite, Infor, Epicor, Acumatica, Sage, Odoo and ERPNext are the ones we see most, plus the industry-specific systems that never appear on a vendor list. The integration route matters more than the badge: if it has an API, a database we can read, a scheduled export or an EDI feed, we can work with it. We have yet to meet a system we could not read.
Nothing breaks, because the layer talks to it through documented interfaces rather than reaching into its tables. That isolation is one of the strongest arguments for building above rather than inside: a vendor upgrade becomes a compatibility check instead of a re-implementation.
As a source and a destination. We read pipeline, accounts, activity and quotes; we write back scores, next best actions and suggested orders so the rep sees them where they already work. Salesforce, HubSpot, Zoho, Pipedrive and Dynamics all support this properly. Nobody is asked to learn a second system to get the benefit.
It limits what the commercial models can say, and we will tell you which ones. In practice the fix is not a data-entry campaign — it is removing the reasons the data was never entered. Auto-capture from email and calendar, suggested orders that pre-fill the record, and enrichment against external sources do more for CRM hygiene than any amount of asking.
It is the normal starting condition, and it is why the first two weeks are a data reality check rather than a design sprint. We profile every source, reconcile the SKU master against what actually ships, quantify the gaps, and tell you which decisions are supportable today and which need instrumentation first. AI applied to bad data produces confident wrong answers, so we fix the foundation before building on it.
Only as fresh as the fastest decision it supports. Streaming a monthly planning input is spend without return; running a stock-out alert off last month’s extract is negligence. We map each decision to a cadence — monthly, weekly, daily, hourly or live — and instrument to that. Most manufacturers need three of the five, not all of them.
One place where “active SKU”, “active outlet”, “on-time”, “margin” and “lead time” are defined, and one grain — SKU by location by day — that every dashboard, report and AI answer resolves against. Without it, two departments quote two numbers and the meeting becomes an argument about data lineage. We have seen a single definition gap — 60-day versus 90-day active outlet — account for a 214-outlet disagreement about coverage.
Your cloud account, your region, under your policies — AWS, Azure or GCP. We build in your tenancy rather than ours so there is no vendor to leave and no data to extract if you ever part company with us. Single-tenant deployment where regulation or procurement requires it.
Yes, and most clients do. Power BI is a good presentation layer — the problem is usually what sits behind it. We build the governed model underneath, so Power BI, Tableau, Looker and the applications we build all read the same certified numbers. What changes is that the report stops disagreeing with the other report.
A dashboard hands you a figure and leaves. Every view we build carries four things a dashboard does not: what the number will be, how confident that is, what is driving it, and the action that changes it — with a button. If a chart cannot change a decision, we do not ship it.
No, and that is deliberate. Finance argues in margin bridges and ageing buckets; operations argues in constraint grids and OEE; logistics argues in cost per mile and lane performance; retail argues in on-shelf availability by SKU and like-for-like. Same model underneath, a different surface and a different argument per role.
Each is a narrow specialist that watches one thing continuously and drafts one kind of action. The finance agent generates and sends invoices on delivery confirmation, chases overdue payment, matches remittances and forecasts 13-week cash. The route agent consolidates loads. The compliance agent tracks certificate expiry. Twelve of them, each with a stated job, a data context, a measured metric and a hard guardrail.
Release a payment, write off a balance, accept or waive a contract clause, submit a bid, or send a price below your margin floor. Those are hard-coded refusals rather than settings someone can quietly switch off in year two. Agents draft and recommend; people approve. Every automated action is logged with actor, policy, timestamp and a rollback path.
Choosing the right tool for each job. Forecasting and scheduling use statistical models and constraint solvers, not language models — a solver cannot invent a machine that does not exist. Language models are used where language is the task: drafting a document, summarising a thread, answering a question over governed data. And the assistant on this page refuses to guess rather than improvising an answer.
Seven layers: your existing sources, orchestration, the data platform and semantic model, the intelligence layer split between ML and Gen-AI, an API and action layer, the consumer layer for every front end, and design and research orchestration around all of it. There is a full diagram on this page, and you can download it as an SVG for your own architecture review.
REST and GraphQL APIs, native connectors, EDI and flat files, change data capture, webhooks and events, over a private link where required. We choose the least invasive route that meets the freshness the decision needs — and never the one that requires direct writes into a vendor’s tables.
Because everything is API-first, the front end is a choice rather than a constraint: web for the office, tablet for the plant floor and warehouse, mobile for reps and drivers with offline-first sync, a permission-scoped portal for distributors and customers, your existing BI tool, a chat assistant, and embedded views inside the ERP or CRM screens people already use.
The cost curve is the point. Licences grow with headcount forever; a build grows with data volume, which is far cheaper and far more predictable. Adding a plant, a territory or forty users changes your infrastructure bill by a small amount and your licence bill by nothing, because there is no licence.
Because the two arguments are the same argument. What the factory can make and what the market is being told are usually managed by different people with different numbers, which is how you end up promoting a line you cannot supply. Same team, same model, one calendar tied to your production plan.
Being cited inside an assistant’s answer rather than only ranking on a results page — and the two are now genuinely different games. Around 83% of AI Overview citations come from pages outside the organic top ten. What moves it, in order: freshness (recently updated pages appear roughly 4.3× more often, and about 85% of AI Overview citations are under two years old), unambiguous entity data in schema.org that matches your third-party listings, corroboration off your own domain (community and third-party platforms carry over half of all citations), statistics and named sources in the body text, and crawler access you have actually checked rather than assumed.
We ship one because it costs nothing, and we tell you not to count on it — measured traffic from llms.txt is around 0.1% of AI crawler activity. The decision that actually matters is which crawlers you allow. Crawl-to-referral ratios differ by orders of magnitude: Google historically sat near 5 pages crawled per visitor sent, Perplexity around 111:1, GPTBot over 1,000:1 and ClaudeBot into the tens of thousands. Allow the agents that cite and attribute; decide about the training-only crawlers deliberately rather than by leaving a default in place.
Less than agencies quote and more than most manufacturers do. Roughly 3–5 LinkedIn posts and one email a week, 4–8 substantial pages a month, a mini-refresh of anything important every 60–90 days, and one master video shoot per month cut into six to ten derivatives. There is a full cadence table on this page with the benchmarks it is drawn from.
Role-based access at the data layer, not as a UI filter — every query is shaped by the user’s permission context, so a rep cannot reach another territory even through an API. SSO with MFA, encryption in transit and at rest, private networking to your ERP, and an audit trail on every automated action. Built to SOC 2 and ISO 27001 control families with evidence collected continuously.
NIST AI Risk Management Framework as the operating model — Govern, Map, Measure, Manage — with ISO/IEC 42001 as the certifiable management system on top, so one implementation produces evidence for both. Every model ships with its purpose, data lineage, known limitations, measured accuracy, drift monitoring and a retirement condition. A model with no stated failure mode is not finished.
It is a large part of what we build. FSMA 204 traceability with the 24-hour sortable extract, FDA QMSR and ISO 13485 records, 21 CFR Part 11 audit trails, HIPAA controls and a BAA where PHI is in scope, TTB and Metrc filings, ASTM and AASHTO conformance packs. We build the system that makes an audit survivable — interpretation stays with your counsel and your certifier.
Only if you sell into Europe. Transparency obligations apply from 2 August 2026; the Digital Omnibus defers most standalone high-risk obligations to 2 December 2027 and embedded high-risk to 2 August 2028, but only once published in the Official Journal. We design to the earlier date and let you enjoy the later one.
Four to six weeks for the first working decision, with a measured baseline and a measured delta. Not a slide deck — a thing someone opens on a Monday that changes what they do that week. Full first release in 8–12 weeks. Anything quoted faster than that is either trivial or untrue.
You do. Schema, pipelines, models, prompts, dashboards and source code — plus the Figma files, the design system and the research. It is delivered in your repositories and your cloud account, and it keeps working if you never speak to us again. That is the practical difference between owning software and renting it.
The licence figures are published 2026 list prices with sources, applied to headcounts you set yourself — so change the inputs and watch the answer change. They exclude implementation services, and enterprise vendors discount heavily off list. At low seat counts the calculator will tell you licensing is genuinely cheaper, because that is sometimes true and you would find out anyway.
Most clients do, and we would push you toward it. One pillar, one measurable decision, one number that moves. The catalogue on this page is the full menu, not a mandatory scope — the sequence that works is: prove it in one area, then let that fund the next.
No questions in that group yet.
Every price and statistic on this page, in one place. Checked 2026-08-04. Software pricing moves — if a figure looks stale, it probably is, and we would rather you tell us.
I answer questions about how this would work for a manufacturer — cost, timelines, your ERP, your data, regulated categories. Pick one, or type your own.
Cost and commercials
Getting started
Your systems
Specifics