Seven layers. Everyone buys the same bottom. Nobody can buy the top.
Software gets installed. An operating model gets run. We deploy sovereign compute, an open base model, your industry's regulation and your company's records, then put an agent workforce inside your workflow — and instrument every decision it makes and every outcome that follows. The bottom four layers are procurement. The top three are why the operation gets cheaper and better every quarter it runs.
Compute and base model. A supplier relationship. Swapped when something better ships.
Industry corpus and your company data. Necessary, and available to anyone with the same budget.
What was decided, what happened, and what that teaches the next deployment.
Seven layers, three bands.
Competitors sell into the bottom two. We deliver all seven, and the top three cannot be delivered by anyone who does not run the operation.
What was decided, what was done, what happened, and what that teaches the next deployment. This layer cannot be procured. It only accumulates, and only if someone is instrumenting for it from day one.
The bottom four layers are a purchase order. The top three are a history.
The deposit under the operation.
Every operation sits on a deposit — interactions, transcripts, dispositions, exceptions, escalations, overrides, resolutions. In most enterprises it is retained for compliance, stored at cost, and read by nobody. It is ore, and ore is not gold.
The deposit is yours. It stays yours. We build the refinery on your side of the wall, and what leaves is a decision — never a record.
It is action-conditioned
Not only what the customer said, but what was done in response and what happened next. State, action, outcome. Most enterprise data records the state and skips the other two.
Outcomes resolve fast
In collections, servicing, renewals, claims and disputes there are a bounded number of things you can do, and you learn within days or weeks whether it worked. That combination is what makes learning possible at all.
The counterfactual is built in
Running an operation at scale means the same situation gets handled different ways across thousands of instances. That is an experiment which has been running for eighteen years without anyone calling it one.
It cannot be acquired
Compute is available to anyone with capital. Base models are available to anyone with a download. A record of what an operation decided and what followed exists only where someone was running it and instrumenting at the time.
Frontier labs read about the enterprise. We stand inside it.
A decision record has three parts: the state something was in, the action that was chosen, and the outcome that followed. The public internet is full of the first part, described after the fact by someone who was not in the room. The other two are almost never written down — and a model cannot learn a policy from outcomes it has never seen attached to the actions that caused them.
The same situation gets handled different ways across thousands of instances, so the counterfactual is built in. Pricing an outcome requires knowing the counterfactual — which is an actuarial asset, not a software asset.
Why the gap does not close with scale
A larger model reads more of what was written down. It does not create records of decisions nobody recorded. Scaling the top row does not produce the bottom row.
Why it cannot be bought
Regulated data does not leave the building — which is precisely why an architecture that requires the data to travel is structurally excluded from producing this record.
Why eighteen years matters
The record accumulates at the speed the operation runs. There is no way to buy the back years, and the acquisition of an operator gets you the seats rather than the history.
Capital can buy compute. It cannot buy eighteen years of outcomes.
Instrument before you automate.
A learning loop needs a stable thing to attach an outcome to. If the same customer is three different records across your CRM, your billing system and your collections platform, the outcome attaches to nothing and the loop never closes. Which is why phase one of every deployment is instrumentation — and why we do it whether or not you proceed.
A record with no loop is a data warehouse
Beautifully modelled, queryable, and it never gets better. Fifteen years of enterprise data programmes ended here.
A loop with no record is A/B testing
You learn that B beat A, but not for which customer, under which policy, in which state. So the finding does not generalise, and it does not survive a regulator's question.
Remember. Learn. Reuse.
Every account, decision, policy and outcome lives as one connected record instead of five disconnected systems. One customer, one history.
Because outcomes attach to specific decisions in specific states, the system can tell which approach worked for whom, and under what conditions. Not on average — specifically.
A pattern proven in one place raises the starting point in the next. The second deployment does not begin at zero.
Cost per resolved task, indexed. [Illustrative — replace with measured figures before publishing.]
Structure carries. Records don't.
A pattern learned in one deployment makes the next one cheaper. That is the economic argument, and it is also the sentence that makes a data protection officer sit forward. So here is the mechanism, precisely.
What carries is shape
That accounts have arrears states. That promises have cure rates. That certain sequences of contact perform differently under certain conditions. The industry corpus is the shared schema.
What never carries is instance
Your customers, your balances, your records, your anything. Company data is the instances, and instances stay behind the wall.
Department to department
Inside one firm, where a single controller decides. Your collections function and your servicing function share a learning curve.
Firm to firm, federated
Model updates move, raw records do not, and each firm consents explicitly to participate.
Market-wide
Only where participants have agreed to pool at signup. A ladder we are climbing, not a moat we already hold.
This is also why the moat is vertical and the platform is horizontal. A collections pattern learned at one bank transfers to another bank because both have accounts, arrears, promises and cures with the same shape. It does not transfer to an airline.
Certify, intercept, correct.
An autonomous agent in a regulated process needs three things before a compliance officer signs: proof it was certified against policy before deployment, the ability to intercept a decision in flight, and a record of every correction made after. We run that as a control plane over every deployed agent.
The control plane is also the capture plane. Every interception is a labelled example of the boundary. Every correction is a domain expert saying exactly where the system was wrong, and why.
Compliance is normally a tax on AI deployment. Instrumented correctly, it is the input — the governance layer regulated clients require is the same layer that produces the highest-quality training signal in the stack.
Frontier models get cheaper. The advantage gets deeper.
The most common question from technical buyers: what happens to your position when open models are good enough that anyone can do this? We are the buyer of that trend, not the seller.
Better models make the operation cheaper. The base model is a line item that falls every quarter and swaps out when something better ships. Our unit of measurement is cost per resolved task, not benchmark rank.
Cheaper operations win more contracts. When capability improves, margin on a contracted outcome rises. When multilingual coverage improves, [120+] languages get cheaper to serve.
More contracts deepen the layer nobody can buy. None of it substitutes for a record of what an operation decided and what followed.
Two layers change. Five stay constant.
That is what makes this a platform rather than five bespoke projects — and why a deployment in your sector starts from a running position rather than a blank page.
| Sector | Industry corpus | Action and outcome data |
|---|---|---|
| Banking | Prudential and AML codes, dispute rules | Collections outcomes, recovery rates |
| Telecom | Consumer codes, tariff structures | Churn saves, credit decisions |
| Healthcare | ICD-10, HL7/FHIR, care guidelines | Prior-auth results, pathway outcomes |
| Fintech | E-money licensing, KYC tiers | Fraud calls and their outcomes |
| Travel | Fare rules, disruption policy | Rebooking and waiver outcomes |
Sovereign by construction.
Sovereignty here means in-perimeter, not government-only. Every regulated enterprise needs it; not every regulated enterprise is a ministry.
On-premise or in-region
Your choice, written into the contract rather than set as a configuration option. The record is captured and stays inside your perimeter.
Model-agnostic
No dependency on a single frontier lab, and none on a lab being permitted to operate in your jurisdiction.
Auditable by construction
Every decision reconstructable with the policy version in force at the time, the state the account was in, and the reasoning trace.
Compute funded as a real asset
Where an in-country facility is needed it is capitalised deal by deal by a real-asset partner, with an infrastructure operator running it. We anchor as the operating tenant. It sits on neither your balance sheet nor ours.
One number we contract against
Cost per resolved task, published monthly against your baseline. Not benchmark scores, not model size, not tokens, not seats.
You keep the record
The instrumented record is built inside your perimeter and it is yours. If we part, it stays.
Additive, not a re-platforming programme
We deploy alongside your existing core, CRM and telephony rather than replacing them. Scope can start narrow, prove value, and expand function by function.
Compliance is architectural
Audit trail, data residency and regulatory guardrails are built into the stack rather than bolted on, so the evidence your examiners ask for is an output of the system.
Full-population monitoring, not a sample
Fraud, mis-selling, evasion and agent absence monitored across every interaction, with inference running at the edge inside your environment.
Where we win, and where we don't.
On-premise deployment used to be the moat. It is becoming table stakes, and we are glad — because it moves the contest to the thing nobody can buy.
| Category | Where they are strong | Where the line falls |
|---|---|---|
| Hyperscalers | Bundled, already inside the bank, unmatched infrastructure | We compete on depth of operation, not infrastructure — and increasingly we run on top of them |
| Well-funded AI-natives | Greenfield, digital-first accounts. Better capitalised than us | They do not win eighteen-year incumbency inside a Tier 1 bank's core operations |
| Regulated-vertical specialists | Closest to us on deployment posture | Narrower on scope, and software vendors rather than operators |
| Global CX incumbents | Installed telephony and workforce platforms | Structurally slow, because their revenue depends on the seat model agents replace |
| Traditional BPOs | Scale, relationships, delivery footprint | Buyers of AI rather than builders. Our channel, not our rivals |
What we lose, said plainly: greenfield digital-native accounts, and accounts already under a global vendor or hyperscaler mandate. We win where operations are complex, regulated, multilingual and already running.
Installed position
[176] regulated enterprises, [7] countries, 120+ languages, [62] local channel partners. Years to replicate, not dollars.
Switching cost
The software runs the operation. Leaving means abandoning institutional memory and re-certifying from zero.
The decision record
Eighteen years of situation, action chosen and outcome, captured because we operate. The only one that cannot be bought at any price.
How it lands.
Instrument
Unify the record — one account, one history. Baseline every metric before anything changes. You keep this whether or not you proceed.
Automate
Agents deployed on the highest-volume, lowest-judgment queue, running in parallel with your existing operation until the SLA threshold is met.
Compound
Loops close, patterns accumulate, and the next function starts from a running position rather than zero.
Phase 1 is the one most vendors skip, because it does not demo well. It is the only phase that determines whether Phase 3 ever happens.
What technical buyers ask.
Isn't the base model doing all the work?
The base model does the reasoning. It has no memory of your operation, no record of what worked, and it resets every session. Layers five through seven are the difference between a capable model and an operation that improves.
Why can't our incumbent vendor do this?
They can buy the same compute and download the same model. What they do not have is a record of decisions and outcomes, because they were not instrumenting for it. That takes years to accumulate, not quarters.
Couldn't a competitor just acquire a BPO?
They can buy the seats. The instrumented history is not on the balance sheet, and the acquisition itself takes years.
What happens to our data if we leave?
You keep the instrumented record. It was built inside your perimeter and it is yours. We keep learned structure, which contains none of your instances.
How do we explain an AI decision to our regulator?
Every decision is reconstructable with the policy version in force at the time, the state the account was in, and the reasoning trace behind it. That is the reason the control plane exists.
What if the model you use gets deprecated?
It will. Base models are a supplier relationship and we swap them. Layers three through seven survive the swap intact.
Start with the instrumentation.
Phase 1 stands alone. We baseline your operation, unify the record, and show you what your current process actually costs per resolved task. You keep the analysis and the instrumentation whether or not you proceed.