Logiciel Solutions Contact Us
Success Stories Tech News Contact Us
whitepaper

AI Governance: What Buyers Should Ask.

Every vendor will answer these questions, and answer them fluently. That is the problem: the question set is the easy part, and reading the answer is the skill. This page pairs each question with the plausible, confident, evasive reply most buyers accept and the specific reply that should reassure them, across six themes from accountability to exit. Put every question in writing, ask for the artefact rather than the assurance, and treat any answer that names a certification instead of a control as a finding.

In depth

Every Vendor Will Answer. Judge The Answer, Not The Badge.

01

Two themes decide most of your residual risk.

Accountability and data both get settled in the contract rather than in the product, so ask them first: who is a named owner of model behaviour when a decision goes wrong, what the liability clause actually carves out, whether your data trains anything by default, and what the IP indemnity covers. A vendor who cannot answer these will not improve on the harder ones. Ask them in the first meeting, while declining to proceed is still cheap.

In shortwhile declining to proceed is still cheap
02

A security page is not technical documentation.

A SOC 2 report, a trust centre and an annual audit answer a different question from the one a conformity assessment asks, which wants intended purpose, architecture, data governance, risk management, accuracy and logging, each with an owner and a revision date. Ask to see the documentation index in the meeting rather than after it. An artefact that is always one week away is itself the answer.

In shortAn artefact that is always one week away is itself t…
03

A leaderboard score is not an evaluation.

What you need is a task-level suite built from real cases in your domain, with the grading rubric, the sample size, per-version results and the regression gate that blocks a release, plus an error profile broken down by type and severity for the one decision you intend to automate. Ask to bring your own holdout set into a sandbox before signing. Results you keep whatever they show, inside a defined trial window, is the term that separates a trial from a demo.

In shortis the term that separates a trial from a demo
The detail

The Two Questions Buyers Ask Last And Regret Asking Last.

Model behaviour changes under you, and lock-in is assembled quietly out of your own prompts and evaluation work. Both are cheap to settle before signature and close to impossible afterwards, which is why they belong in the first conversation and not the renewal.

Zone · 01

Version notice

The weak answer is that they continuously improve the model, so you are always on the best available version at no extra cost. The strong answer is pinned versions you choose, a notice period written into the contract, a published changelog and behavioural diffs between versions on a standard suite.

Zone · 02

Deprecation window

A status page and a developer newsletter are not commitments. What you want is a minimum support window stated in months, a period where both versions run, migration support, and a documented right to stay on the old version until your revalidation finishes. Retirement of a model you depend on should not arrive as a newsletter item.

Zone · 03

Exit package

Exporting your data through the API is not an exit. Ask for an enumerated package: records, prompt and configuration history, evaluation sets and results, embeddings, logs, and either the fine-tuned weights or a clear statement that they cannot leave. Ask what cannot be exported and treat that list as the true cost of switching.

By the numbers

The figures that make it a board-level conversation.

7
artefacts to demand before signature rather than during onboarding
$6.07M
the cost of a model inversion incident, the most expensive AI incident type
$4.99M
global average cost of a breach, up 12% year on year
Inside the report

What you'll take away.

01

Step 1 - Get the data processing terms in writing

Read the consent language rather than the marketing. A training default that a setting can switch off is a commercial position, not a control, so the no-training term belongs in the agreement.

02

Step 2 - Read the ISO 42001 scope statement aloud

Scope is the whole question. A certificate covering a corporate management system tells you nothing about the product you are buying, so ask for surveillance audit dates and any open non-conformities too.

03

Step 3 - Run your own holdout set in their sandbox

Your cases, your graders, your rubric, and results you keep whatever they show. Make that right explicit in the trial agreement, since a guided proof of concept on a curated dataset proves nothing.

04

Step 4 - Have the IP indemnity clause read out

What it covers, whether output is included as well as the service, the cap, the conditions that void it, and who controls defence. Ambiguity here is always resolved against the buyer.

Questions

Frequently asked.

What is the single fastest way to sort strong vendors from weak ones?

Ask for one artefact in the meeting rather than after it. The documentation index, the certificate scope statement, the eval suite, the indemnity clause. A vendor who has done the engineering can show at least one of them on screen within minutes. One that is always a week away has given you the answer.

The vendor says the AI Act does not apply to them outside the EU. Is that right?

Rarely, and it is worth pressing. Placing a system on the EU market or using its output in the EU pulls a provider in regardless of establishment. Article 50 transparency duties have applied since 2 August 2026 and were not delayed, and the December 2027 date they will cite covers Annex III high-risk obligations only.

Is a SOC 2 Type II report enough evidence of AI governance?

It is evidence of information security controls, which is not the same thing. It says nothing about training data provenance, evaluation practice, model change notice or human oversight. Accept it for what it covers and ask separately for the documentation index, the eval results on your cases and the version policy.

How much of this can we realistically get into a contract?

More than most buyers attempt, and the asks that succeed are specific ones. No training by default, a named notice period for model changes, a minimum support window in months, an enumerated exit package with formats, and audit rights that survive termination. Vague commitments to best-efforts support are what get conceded instead.

What does a bad answer on liability actually look like?

Output is provided as-is, the customer is responsible for reviewing it, and total liability is capped at twelve months of fees. That is the market default and it is negotiable. Look for carve-outs on the failure modes that matter to you, a separate sub-cap or super-cap, and a review obligation scoped to defined decision types.

Where does this fit with your other governance papers?

This one points outward at vendors. AI Governance: An Engineering Reference is the same subject turned inward, with the controls and artefacts you build yourself, AI Governance Under Regulation covers which duties bind you and by when, and Benchmarking AI Governance scores where you currently stand.

Who should be in the room when these questions are asked?

Whoever owns the decision the system will influence, plus someone who can read a contract and someone who can read an eval. Procurement alone gets fluent answers accepted. The combination catches the gap between what the product does and what the agreement obliges the vendor to keep doing.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send AI Governance: What Buyers Should Ask straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Ask us the same questions you ask everyone else.

Put this question set to our engineering leads in a working session, and judge the answers by the artefacts rather than the badges. A vendor due diligence review, on us. SECTION 7 - FAQ - 5 to 8 questions

Book a due diligence review