Embedded operators

An AI implementation partner
that sells judgement, not headcount.

Every embedded engineering vendor offers the same thing: senior engineers in your repository, starting next week. That is a staffing answer to a judgement problem. We embed alongside the people who do the work to decide which parts of it should become a model, which should become ordinary code, and which should stay with a person. Usually, most of the system should not be AI.

  • Five named artifacts, not a headcount
  • Built on top of your ERP. Never a migration.
  • Paid mapping first, so the build quote is a finding, not a guess
Book a qualification call →Read: how to buy AI delivery

Five-day Mapping Sprint · First workflow live in 6 weeks · Europe

What the category sells

A pod of engineers

Senior engineers embedded in your repository. Seven-day start. Four to eight weeks to ship. Priced per engineer per month. You still decide what to build, and you still carry the risk of building the wrong thing.

What we sell, you are here

The decision about where AI belongs

We work through the real process with the people who run it, find the exceptions nobody documented, and draw the boundary between what should be a model, what should be code, and what should stay human. Then we build the small part that needs building.

Why embedded delivery exists

The platforms already conceded this.

The forward deployed engineer model originated at Palantir, and the AI labs adopted it: OpenAI and Anthropic both run forward deployed engineering units. In June 2026 AWS announced a $1 billion investment to embed forward deployed AI engineers with customers, describing thousands of experts working inside customer environments with self-sufficiency as the exit condition.

When the companies whose entire business is the model spend a billion dollars embedding engineers inside customer teams, the bottleneck is not the model. It is your data, your exception paths, your approval chains, and the three systems that do not have an API. None of that is visible from a demo, and none of it is in your process documentation.

Sources: AWS, “AWS invests $1 billion to embed AI forward deployed engineers with customers” (2026); Databricks, “Forward Deployed Engineering”.

How to buy this without a billion dollars →
What you are actually buying

Five artifacts, each one nameable.

You cannot inspect a pod of engineers. You can inspect a document. Each of these has a definition, an owner, and a date it is delivered, which is what makes the engagement auditable instead of a matter of trust.

Exception LedgerEvery real deviation from the documented process, how often it happens, and who absorbs it today.
AI Boundary MapEach step marked as deterministic code, model judgement, or human approval. Most steps are not AI.
Evaluation suiteA golden dataset, per-step pass rates, and a monthly accuracy report. A named line item, never bundled.
Decision logEvery agent action, its inputs, its confidence, and who approved it. A surface a compliance officer can open.
ERP-additive covenantWe build on top of your existing systems. We do not propose replacing them. No migration.
Artifact one

The Exception Ledger

A written record of every real deviation from your documented process. For each one: how often it occurs, who currently absorbs it, and where the routing logic actually lives. That last column is the useful one, because the answer is almost never “the documentation”. It is one person's head, and it has been there for eleven years.

The only way to produce this is to watch the real process run, exception by exception, with the people who handle them, and to keep asking why until the rule comes out. A team working from your repository and a weekly status call does not do that. The information is not in the repository, and nobody volunteers it on a call, because nobody thinks to mention the thing they have always just known. It is not a documentation gap. It is what expertise looks like from the outside.

  • Every deviation, with observed frequency rather than an estimate
  • The person who absorbs it today, named
  • Where the decision rule lives, and whether it can be written down
  • Which exceptions must stay human, and why
Artifact two

The AI Boundary Map

Your workflow, annotated step by step: deterministic code, genuine model judgement, or human approval. The headline number is the determinism ratio, and the headline finding is almost always the same one. Most of the system should not be AI.

In one European manufacturer's order-handling workflow we mapped, three of eleven steps genuinely needed a model. The other eight were parsing, lookups, validation and routing — work that ordinary software does more cheaply, faster, and with an answer you can reproduce tomorrow. (SUPALABS engagement data.)

CostEvery step you move out of the model is a step you stop paying per token to run.
LatencyDeterministic steps return in milliseconds. A chain of model calls does not.
AuditabilityA rule can be shown to a regulator. A sampled generation has to be argued for.

This is why an AI-first vendor is the wrong shape of supplier for this problem. If the answer to “how much of this should be AI” determines the size of the invoice, you are not going to get an honest answer.

Artifact three

The evaluation suite and the monthly accuracy report

A golden dataset drawn from your own historical cases, per-step pass rates against it, and alerts when a step regresses. It is quoted as its own line item, never folded into the build, because a capability you cannot see on an invoice is a capability that quietly gets cut when the timeline tightens.

  • Golden dataset built from your real cases, including the awkward ones
  • Pass rates per step, not one aggregate score for the whole system
  • Regression alerts when model behaviour drifts after a provider update
  • A monthly report written for the person who has to sign off on the system

This is also the honest reason the relationship continues after the build. Model providers change behaviour without asking you. Somebody has to notice.

Artifact four

The decision log

Every agent action, its inputs, its confidence, and who approved it, rendered as a surface a compliance officer can open without asking an engineer for help.

This is ordinary engineering. We are not going to claim otherwise. The differentiation is not that it is clever, it is that most suppliers skip it, and a system without it does not get permission to go near production in a regulated or audited environment. That is the whole argument for building it: it is the reason the system is allowed to run at all.

Artifact five

The ERP-additive covenant

We build on top of your existing systems. We do not propose replacing them. We do not require a migration. That is a commitment we make in writing at the start of the engagement, not a preference we hold until it becomes inconvenient.

A replatforming programme is the largest career risk an operations or IT leader can take on, and it is almost never what the problem actually requires. The workflow is slow because of the exceptions, the handoffs and the approvals — not because of the database underneath them. Changing the database does not fix any of that, and it takes two years to find out.

For organisation-wide programmes, see the AI Efficiency Programme →
How an engagement runs

Four rungs. You can stop after any of them.

0
Qualification call — free, 30 minutesA straight answer on whether there is anything here worth mapping. No deck, no proposal, no follow-up sequence.
1
Mapping Sprint — five days, embedded with the teamPaid discovery. Exception Ledger, AI Boundary Map, evaluation plan, and a fixed-price build quote based on findings rather than assumptions. If it shows you nothing you did not already know, you do not pay.
2
Build — six weeks to productionA fixed-price build quote, scoped from what the sprint found rather than from a guess. Majority deterministic software, built on top of your existing systems, with human approval at every irreversible step.
3
Run — ongoingMonthly accuracy report against the golden dataset, regression alerts, cost-per-run tracking, and quarterly re-scoping of the next bottleneck.
Book the qualification call →See the full ladder
Read this before you book

When you should not buy this.

There are three situations where an embedded operator engagement is the wrong purchase, and it is cheaper for both of us to establish that on a thirty-minute call than in month two.

You can staff a standing practiceIf you are large enough to run your own internal delivery team, hire it. Embedded delivery is how you learn what to hire for, not a permanent substitute for it.
The problem is arithmetic, not judgementIf the rules are knowable and stable, a rules engine or a well-built spreadsheet beats anything we would put in front of you, and costs less to maintain. We will tell you on the call.
No executive owns the outcomeWithout a named owner who can clear an approval boundary, the work stalls at the first one. That is true regardless of who builds it.

If one of these three is your situation, we will say so on the qualification call rather than after the invoice. It costs us a deal and saves you a programme.

FAQ

Frequently asked questions

How is this different from AI staff augmentation or an embedded engineering pod?A pod sells you capacity. You decide what to build, they build it, and the risk of building the wrong thing stays with you. We sell the decision itself: which steps of your workflow should become a model, which should become ordinary deterministic code, and which should stay with a person. In practice most of the system should not be AI, and a vendor paid per engineer per month has no reason to tell you that.
What is an Exception Ledger, and why does reading our documentation not produce one?It is a written record of every real deviation from your documented process: how often each exception occurs, who currently absorbs it, and where the routing logic actually lives. That logic is almost never in the documentation. It is in one person’s head, and it only comes out when you watch the exception happen and ask the person handling it why they did what they did. A team working from your repository and a weekly status call does not do that, so it cannot produce the ledger: the information is not in the repository, and nobody volunteers the thing they have always just known.
How much of the system will actually be AI?Less than you expect, and that is the point. The AI Boundary Map annotates each step of the workflow as deterministic code, genuine model judgement, or human approval. A typical result is that three of eleven steps need a model. Every step you move out of the model is cheaper to run, faster, and easier to audit, so the determinism ratio is a commercial number, not an engineering preference.
Do you replace our ERP?No. We build on top of your existing systems, we do not propose replacing them, and we do not require a migration. This is a covenant, not a preference. A replatforming programme is the single largest career risk an operations or IT leader can take on, and it is almost never what the problem actually requires.
How does an engagement start, and how is it priced?It starts with a free thirty-minute qualification call. From there the shape is always the same: a paid Mapping Sprint of five days, run alongside the people who do the work, which produces the Exception Ledger, the AI Boundary Map, an evaluation plan and a fixed-price build quote; then a fixed-price build scoped from what the sprint found rather than from a guess, with the first workflow in production in six weeks; then an ongoing run engagement once the system is live. Everything after the call is priced per engagement, and we quote it after the call rather than before, because the scope is a finding. If the sprint shows you nothing you did not already know, you do not pay for it.
When are you the wrong choice?When you are large enough to staff a standing internal practice, hire it instead — embedded delivery is how you learn what to hire for, not a permanent substitute. When the problem is arithmetic rather than judgement, a rules engine or a spreadsheet will beat anything we would build, and it will cost less to maintain. And when no executive owns the outcome, the work stalls at the first approval boundary regardless of who builds it.

Thirty minutes to find out whether there is anything worth mapping.

The call is free, and we will tell you if the answer is no. Bring one workflow that annoys you.

Book a qualification call →See all services