FactoryInstitutional memory for revenue
We build AI that moves renewal retention.
Agreed with your CFO before we start. Defined in writing. Measured against your own figures. Built inside your stack, and yours to own when it ships.
Four records. One customer. One history. That is where every build begins.
We are not an AI company. We are a company that solves one expensive problem at a time.
Nobody wakes up wanting artificial intelligence. They wake up because quotes are going quiet, because the best dealer stopped ordering, because a customer left and nobody saw it coming. We start at the problem, put a number on it with your CFO, and build the smallest thing that moves it.
A third of your open quote value has had no event for twelve days. Your installed base is buying spares from someone else. Dealers who used to order stopped, and nobody called. None of this needs a new customer to fix.
Finance and sales compute open pipeline differently. Two people give the board two figures and both are right. Every improvement gets argued about by whoever is not benefiting from the way it was measured.
Twenty years of judgment about who pays late, which discount shape means the deal is already lost, which plant manager actually decides. None of it was ever written down, because there was nowhere to write it.
We do not run programmes, workstreams or roadmaps. One problem, one number, one thing built, then the next one.
Your systems stay where they are. We read them in place. Nothing is migrated, replaced or rebuilt to suit us.
If the answer is a spreadsheet, a process change or nothing at all, we will say so, and you will not get an invoice for it.
If it cannot be traced to a line in your P&L, we will not build it.
Most AI spend in a company your size cannot be traced to a dollar, and that is not a technology failure. Nobody agreed which number would move before the work started, so the result gets defended with adjectives instead of arithmetic.
Top line means new accounts sourced, quote to order conversion, renewal retention, bind rate, share of wallet. Bottom line means days sales outstanding, discount outside policy band, commission leakage, claim denials, cost to serve, procurement spend, hours returned to people who are expensive.
Whichever it is, we agree it with your CFO first, in writing, with its definition and its exclusions. If we cannot find one worth several times the cost of the work, we say so and stop.
Nine industries where we already know where the money goes.
The method is industry agnostic. The knowledge is not. Every engagement in a sector deepens what we already know about how that sector quotes, serves, renews and leaks, which is why our second build in any industry is materially better than our first.
01Manufacturing and machine tools
Revenue arrives as a quote and keeps earning through an installed base. Both ends leak, and both are pure arithmetic on data you already keep.
Quote stall detectionInstalled base win-backAMC renewal engineDealer coverageOpen sector
02Auto and Tier-1 components
You answer every RFQ by reflex because refusing feels like losing the relationship, and nobody counts what a quote costs to produce or what escalation you never claimed.
RFQ triage and fitEscalation claimsPlatform whitespaceScrap pattern analysisOpen sector
03Real estate and project sales
You buy the enquiry from the portal and lose it to whoever called first. Site visits are counted, what happens after them is not.
Enquiry response SLAVisit to bookingChannel partner coverageSpend attributionOpen sector
04Insurance distribution
Three acquisitions in, nobody can say how many clients you have, so every cross-sell and retention question rests on a number no one trusts.
Book resolutionSubmission to bindRenewal retentionCross-sell across linesOpen sector
05Logistics and freight forwarding
You lose bookings to whoever quoted first rather than best, and some lanes lose money on every single shipment while looking like growth.
Rate turnaroundLane profitabilityDetention recoveryCarrier invoice auditOpen sector
06Retail and consumer brands
You see primary sales clearly and secondary sales through a fog. Schemes reward volume that was coming anyway.
Outlet coverageTrade spend returnDead stockRepeat and lapseOpen sector
07Professional services firms
You bill less than you agreed and discover it at year end, because scope creep accumulates where nobody is watching it.
Realisation by partnerWrite-off analysisUtilisationClient whitespaceOpen sector
08Industrial distribution and spares
You count what you sold, never what you could not supply, and the warehouse holds money that stopped moving three years ago.
Reorder predictionStock-out lost salesDead stockRebate captureOpen sector
09Medical equipment and diagnostics
The installed base is the business and it has the least visibility. Contract dates sit in one system, reagent volumes in another, and nothing joins them.
Installed base pull-throughAMC and CMC renewalTender qualificationReplacement cycleOpen sectorTwelve sectors are documented in full, each with its own plays, systems and the number we would agree first.
See all twelve industriesWe never ship the Factory. We build in it, and ship you what comes off it.
The Factory is ours and stays ours. What it produces is built for you alone, runs in your own environment, and is yours to keep. Nobody is asked to buy it, licence it, or log into it.
Forty-one plays across the revenue surface. Nine run from the public record today. The rest come online as your systems are connected and the case history builds.
Lookalike sourcing
Your next fifty customers resemble your last fifty and nobody can see the pattern.
Quote velocity
Your quotes take nine days and nobody can say which ones died or why.
Installed base win-back
You sold the machine and then stopped earning.
Trigger monitoring
You call when someone remembers to, not when something changed.
Dealer coverage
Your dealers are your pipeline and you cannot see inside them.
Receivables
You chase the big invoices, not the risky ones.
Every play execution, with the spans, tokens and cost behind it. A run that returned nothing because it could not cite a source is recorded as a failure, not hidden.
| Run | Play | Trigger | Spans | Duration | Cost | Result |
|---|---|---|---|---|---|---|
| r-8f24c1 | Lookalike sourcing | manual | 31 | 2m 58s | $0.51 | ok |
| r-8f24b7 | Trigger monitoring | cron, mon 06:00 | 4 | 58s | $0.09 | ok |
| r-8f2496 | Meeting brief | manual | 4 | 1m 22s | $0.09 | ok |
| r-8f2480 | Territory whitespace | manual | 4 | 55s | $0.08 | ok |
| r-8f2471 | Tender qualification | cron, daily | 6 | 1m 44s | $0.12 | no matches |
| r-8f2455 | Competitor watch | manual | 4 | 41s | $0.10 | ok |
Run r-8f2471 found no live tenders matching the stated capability and returned an empty list rather than loosening the match.
Every suite has a threshold. A play scoring below its gate cannot be promoted, and a play with no suite attached cannot be deployed at all.
Source quality is at 84.2 against a gate of 90, and has fallen every run for eight runs. Two plays are held. The cause is measurable: search increasingly returns aggregator pages where a company's own site was reached. Two plays that do not depend on that suite promoted normally in the same run.
| Suite | What it measures | Score | Gate | Labelled by | State |
|---|---|---|---|---|---|
| Evidence grounding | Every cited URL resolves to a source the search actually returned | 100 | 100 | Radexus | pass |
| Write-back safety | No write reaches a system of record without a named approver | 100 | 100 | Radexus | pass |
| Account resolution | Records from different systems resolve to the correct single account | 97.1 | 90 | Your sales ops team | pass |
| Pattern precision | Sourced accounts a rep marks as a genuine fit, top twenty only | 78.0 | 70 | Your sales team | pass |
| Confidence calibration | Claims marked high confidence that survive contradicting evidence | 92.4 | 85 | Radexus | pass |
| Source quality | Citations come from an authoritative page, not a directory listing | 84.2 | 90 | Radexus | below gate |
One definition of every entity, resolved across systems and versioned. Built once, then read by every play and every case in memory.
| Source | As it was stored | Resolved to | Confidence | State |
|---|---|---|---|---|
| SAP S/4HANA | MERIDIAN FASTENERS PVT LTD | Meridian Fasteners | 1.00 | merged |
| CRM | Meridian Fastners | Meridian Fasteners | 0.94 | merged |
| Dealer portal | M/s Meridian | Meridian Fasteners | 0.81 | merged |
| FieldEasy | (no name) Rajkot plant | Meridian Fasteners | 0.72 | merged, low confidence |
| CRM | Meridian Auto Components | held for review | 0.58 | held |
Every field carries lineage: which system, when read, which transform. Traceable in under a minute, which is what makes the number defensible in a board meeting.
Nothing is written into SAP or the CRM without one of your people approving it. Every decision is retained with the actor and the time, and every approval is itself a case the memory learns from.
| Action | Origin | Approver | Waiting | State |
|---|---|---|---|---|
| Create 42 target accounts in Salesforce | Lookalike sourcing, r-8f24c1 | Sales head | 4 h | waiting |
| Update AMC renewal dates on 18 installed units | Installed base check, r-8f2480 | Service head | 1 d | waiting |
| Create follow-up tasks on 6 stalled quotes | Quote velocity, r-8f24b7 | Sales head | 2 h | waiting |
| Write 38 target accounts | Lookalike sourcing, r-8f23f0 | Hari M | approved | |
| Correct 12 dealer territory codes | Dealer coverage, r-8f23c1 | Hari M | approved |
Approvers can be set per play, per field and per value threshold. Prior values are retained for thirty days and every approved write can be rolled back within that window.
Any provider, chosen per task. Nothing is locked to one vendor: extraction, reasoning, speech and embedding route separately, and any of them can be swapped without touching the play or the memory beneath it.
| Model | Context | Capability | In $/M | Out $/M |
|---|---|---|---|---|
| Anthropic: Claude Sonnet 5 anthropic/claude-sonnet-5 | 1M | tools reasoning structured | $2 | $10 |
| OpenAI: GPT-5.5 openai/gpt-5.5 | 1.1M | tools reasoning structured | $5 | $30 |
| Google: Gemini 3.5 Flash google/gemini-3.5-flash | 1.0M | tools reasoning structured | $1.5 | $9 |
| DeepSeek: DeepSeek V4 Pro 0813 deepseek/deepseek-v4-pro-0813 | 1.0M | tools reasoning structured | $1.115 | $3.346 |
| Mistral: Mistral Medium 3.5 mistralai/mistral-medium-3-5 | 262K | tools reasoning structured | $1.5 | $7.5 |
| Meta: Llama 4 Maverick meta-llama/llama-4-maverick | 1.0M | tools structured | $0.2 | $0.696 |
| NVIDIA: Nemotron 3 Super nvidia/nemotron-3-super-120b-a12b | 1M | tools reasoning structured | $0.085 | $0.4 |
Live catalogue, read 2026-09-01. Browse every model.
Because it is built in our Factory, this is what you actually get.
Everyone else sells you a seat and keeps the system. We keep the Factory and give you what comes off it, on a perpetual licence, running in your own environment, with the memory it has formed inside it.
Your ownership, in plain terms
Perpetual. Irrevocable. In your tenancy.Model your own numbers. Switch on the levers that apply to you.
Ten levers, each with its own arithmetic. Pick a business model to load sensible defaults, switch off what does not apply, and change any assumption. Nothing here is a benchmark. In discovery every figure is replaced with what your data supports.
Value model
5 of 10 levers on, year one.
This is the model we show in the room. Bring it to a discovery and we replace every assumption with your data.
Take these numbers to discoveryNine engagements. Nine numbers. Each reported as it landed.
Schemes were paying for volume that was coming anyway.
Dealer incentives had never been tested against counterfactual volume, and claims were reconciled by hand at quarter end by two people who had done it that way for years.18% lowerResult at 2 quarters, at flat volumeThey had sold the instruments and stopped earning from them.
The installed base existed complete only in a service engineer's spreadsheet, and contract renewal dates lived somewhere sales never looked.+14 ptsResult at 12 months, against prior year baseThree acquisitions in, they could not say how many clients they had.
Two agency management systems and a spreadsheet held overlapping books, so every retention and cross-sell question rested on a client count nobody trusted.4,100+Duplicates resolved, client records mergedSome lanes lost money on every shipment, and looked like growth.
Rate requests were answered in the order they arrived and lane margin was a blended average that hid the worst routes entirely.6.5 to 2.1 hRate response, median, business hoursSingapore ran the region. It could not see two of the countries it ran.
A regional headquarters with distributors in six markets reported one number to the group and had confidence in three of them. The two largest gaps were the two fastest-growing markets.3 to 6 marketsResult at 4 months, reported on one definitionThe escalation clause was in the contract. Nobody was claiming it.
Raw material increases were absorbed quarter after quarter because assembling the evidence took longer than the claim window allowed, and nobody had ever computed what that cost.31% to 78%Result at 3 quarters, of entitlement claimedTwo hundred customers stopped ordering and nobody noticed.
Reorder rhythms were entirely visible in the order history. Nobody was reading them, because the sales team called on their own rhythm instead.+118 accountsResult at 2 quarters, reactivated from dormantThe claims came back for eleven reasons. Everyone thought there were hundreds.
First-pass denial was treated as an unavoidable cost of doing business with payers. It turned out to be a short, countable list of preventable causes.+16 ptsResult at 3 quarters, on first submissionClient names are withheld under NDA. Sector, scale and figures are as recorded in the engagement.
Read all nine success storiesSixty minutes to a live result. Two weeks to a number. Ten weeks to production.
Seven stages, each with a gate. You can stop at any of them and keep everything produced so far, including the map of how revenue actually moves through your systems.
And then something else starts happening.
Fix one number and you have a result. Keep the cases, and the company starts holding knowledge it used to rent. That is the part nobody buys on day one, and the part every client ends up valuing most.
It remembers what happened
Every quote, site visit, call and decision becomes a case with its context intact: what was asked, what was offered, who decided, what happened next. The case does not live in a rep's head or a notebook, so it does not leave when they do.
It forms judgment, not just answers
Retrieval gives you the nearest document. Judgment is the pattern across hundreds of cases: that this discount shape precedes a loss, that this customer type stalls at week three, that this territory needs two visits before a quote. No individual works long enough to see it.
It expires
Every piece of knowledge carries a confidence that decays with age and with contradicting evidence. Prices move, competitors change, a plant closes. A system that never forgets becomes confidently wrong, and the moment your CFO catches it being wrong, the whole thing is abandoned.
Institutional memory. Compounding judgment. Knowledge that expires. Governed learning.
How the Factory System works