ما الذي يجب أن يسلّمه شريك تنفيذ الذكاء الاصطناعي: الوثائق الخمس التي يجب طلبها

What an AI Implementation Partner Should Hand Over: Five Documents, Not a Deck

You cannot inspect a vendor's client list, because the good ones will not name their clients and the ones that do are showing you logos, not work. You cannot inspect a pod of engineers. You can inspect a document. This article lists the five documents an AI implementation partner should hand over on every engagement, in the order they are produced, with who in your organisation reads each one and what it lets them decide. If a partner cannot name these, or produces a status deck instead, you have learned something useful before signing.

The list is not arbitrary. It is the set of artifacts that answers the five questions a buyer actually has: does the partner know how our work really runs, how much of this is actually AI, how will we know it is still right, can we defend a decision it took, and does this mean replacing our software. A status deck answers none of them. The five documents answer one each.

1. The Exception Ledger: Does the Partner Know How the Work Really Runs?

The Exception Ledger is a written record of every real deviation from the documented process: what triggers it, how often it happens, and who absorbs it today. It is the first document because everything else depends on it, and it is the one that cannot be produced from a workshop or a questionnaire. It comes from sitting beside the person who runs the workflow for long enough to see the forty email formats, the customer who phones, the supplier invoice referencing a purchase order that does not exist, and the approval that goes to a different manager on Fridays.

Who reads it: the COO or operations director. What it lets them decide: whether the process is worth automating at all, and which of its exceptions the automation has to survive. An exception nobody recorded becomes a step the model is quietly expected to handle, which is how a pilot that passed the demo dies in month two. We wrote about the underlying problem in why the documented process is never the real process.

The test for a partner: ask how long they spend with the people who do the work before writing a scope. A day is too little. A quarter is billing, not discovery. Five days on one workflow is the range in which the ledger gets written without the invoice becoming the point.

2. The AI Boundary Map: How Much of This Is Actually AI?

The Boundary Map takes the workflow step by step and classifies each step as deterministic code, model judgement, or a decision that stays with a person. The headline number it produces is the determinism ratio: the share of steps that are ordinary software. It is usually higher than the vendor deck assumed. In one European manufacturer's order handling that SUPALABS mapped, three of eleven steps genuinely needed a model; the other eight were parsing, lookups, validation and routing (SUPALABS engagement data, 2024–2026).

Who reads it: the CIO, and whoever owns the budget. What it lets them decide: what the system will cost to run, how fast it will be, and what can be shown to an auditor as a rule rather than argued for as a sample. Every step kept out of the model is cheaper, faster and reproducible. A partner who cannot produce this map is either selling a platform priced as if every step were the hard kind, or has not looked closely enough to know.

The test for a partner: ask what share of the proposed system is deterministic. A partner who answers "all of it is AI" is describing their demo. A partner who answers with a number and a list is describing your process.

3. The Evaluation Suite: How Will We Know It Is Still Right?

The evaluation suite is a golden dataset built from your own past cases, the orders your team processed correctly, the invoices that matched, the drafts a senior professional would have accepted, scored step by step, with per-step pass rates and a monthly accuracy report that runs against it after launch. It should be quoted as its own line item, never bundled into the build, because a bundled evaluation is the first thing cut when the build runs late.

Who reads it: the head of transformation, and the board pack. What it lets them decide: whether the system is still right six months after the partner left, and whether a task has earned more authority. The suite is also what makes the authority tiers real: a task moves from drafting to acting on its own only when the pass rate you set is met on real cases, and it moves back down when the monthly report shows a regression. Without the suite, "human in the loop" is a slogan; with it, it is a threshold.

The reason this matters more every year is that trust in model output is falling as usage rises. The 2025 Stack Overflow Developer Survey found 84% of developers using or planning to use AI tools, while 46% actively distrust their accuracy and only 33% trust it. Those are the people building the systems. A COO who is asked to trust one in production is entitled to a number, not an assurance.

The test for a partner: ask to see the evaluation line on the quote. If it is not there, ask why. If it is bundled, ask what happens to it when the build slips.

4. The Decision Log: Can We Defend a Decision It Took?

The decision log records every automated action, its inputs, its confidence, and who approved it, in a surface that a compliance officer, a finance controller or an acquirer's diligence team can open without asking an engineer. It is table stakes to build. The differentiation is entirely in whether it is shown to you, in a form a non-engineer can read, before you sign.

Who reads it: compliance, finance, and anyone who will one day have to explain a decision to a regulator, an auditor or a buyer. What it lets them decide: whether any single decision the system took can be reconstructed and defended. For a company that files with a tax authority, a works inspector or a financial regulator, that is the difference between an automated process and an unexplainable one.

The test for a partner: ask to see a decision log from a past engagement, anonymised. A partner who cannot show one either does not build one or does not think you will ask.

5. The ERP-Additive Covenant: Does This Mean Replacing Our Software?

The fifth document is the shortest and the most often missing: a written commitment, made at the start, that the partner builds on top of the systems you already run and will not propose replacing them. Not a preference. A clause. A replatforming programme is the largest career risk an operations or IT leader can take on, and it is almost never what the problem requires; the workflow is slow because of the exceptions and the handoffs, not because of the database underneath them.

Who reads it: the CIO, and whoever would carry the migration. What it lets them decide: that this engagement is not a replatforming in disguise. The covenant has a second half that is worth asking for in the same clause: when the engagement ends, the system stays. It runs in your accounts, on your systems, documented for your team from the first week, and stopping the service does not stop the workflow.

The test for a partner: ask whether they will put "no replacement, no migration" in the scope. A partner who hedges is keeping the migration in reserve.

Why the Order Matters

Partners who produce some of these documents often produce them in the wrong order, and the order is where the value is. The Boundary Map cannot be drawn until the exceptions are written down, because an unrecorded exception becomes a step the model is silently expected to handle. The evaluation suite cannot be built until the Boundary Map says which steps have a judgement to evaluate; there is no point scoring a lookup. The decision log only means something once the suite defines what a correct decision looks like. And the covenant has to come first of all, because everything else is designed around systems you keep. Skip a step and the next one is a guess. This is the sequence described in full on the SUPALABS method page, and it is why the same five documents come out of every engagement regardless of sector, which is also what makes them comparable if you are buying for more than one business unit or, as we cover for sponsors on the private equity page, more than one portfolio company.

What This Replaces

The five documents replace three things buyers are usually offered instead. They replace the logo wall, because a document about your process is more informative than a list of companies whose processes you cannot see. They replace the testimonial, because a monthly accuracy report is a testimonial the system writes about itself. And they replace the methodology deck, because a deck is the same for every client and these are not. If a partner offers you the first three and cannot produce the five, the gap is the answer.

Sources & References