Victor LaybatsVictor Laybats
The sprintLive systemsEvidenceGuidesAvailableBook the €490 sprint

ai consultant

AI consultant

What an AI consultant should deliver before any build: one process, data boundaries, human review points and a measurement plan, plus the limits to expect.

Victor Laybats · · 1065 words

  • €490 fixed pilot
  • 72h after access
  • Human review before external action
Victor Laybats, freelance AI and automation engineer in Paris
AI Agent & Automation EngineerVictor · Paris
Editorial scope: Victor Laybats publishes practical guidance for scoping, securing and measuring AI and automation projects.

What an AI consultant is for, and what they are not

An AI consultant should help you decide whether a specific piece of work can be automated safely and usefully. Writing code or picking a fashionable model comes later. The most valuable early output is often a clearer description of the problem than the one you started with. Sometimes that description shows the problem is not ready for AI yet.

This guide is published by Victor Laybats, who works on AI and automation engineering from Paris. The public approach covers the path from initial scoping to deployment and later follow-up. Treat the points below as questions to ask, not as promises about results. What any engagement achieves depends on your systems, your data and the conditions you are starting from.

Start with one accountable process, not a tool

Requests that begin with "we need AI" or with a favourite platform tend to drift. A sounder starting point is one business process with a clear trigger, an intended outcome, a named owner and a known set of systems. A competent AI consultant should be able to reduce your request to a workflow that can be observed, tested and stopped. The decisions that remain human should stay visible throughout.

The choice of tool should follow this map. If a proposal names a vendor before anyone has described the process, the order is reversed. Ask the consultant to explain what the workflow does today, who is accountable when it fails, and what success would look like in business terms. A vague answer to the last question means the business objective is not yet explicit enough to build against.

The discovery checklist to request before implementation

A bounded discovery phase makes competing proposals comparable before anything turns into code. It should separate what was actually observed in your environment from assumptions and from options for later. Without that separation, a delivery estimate can quietly absorb unverified integrations and unresolved permissions.

Use the list below as a decision aid. If several items are missing from a consultant's deliverable, the scope is not yet defined, however polished the slides look.

  • A workflow map with a named owner for each step
  • An inventory of integrations, marked as verified or assumed
  • A classification of the data involved and where it may and may not go
  • Action permissions: what the automation may do on its own and what it may only suggest
  • Human review gates and who staffs them
  • Known failure cases, expected behaviour on failure, and a named recovery owner
  • Explicit non-goals
  • A measurement plan stating the baseline, the metric and who reads it
  • Dependencies needing personal access, provider approval or business sign-off, listed openly

Controlled data, human review, and when automation should wait

Two principles decide whether an AI project can be operated safely. The first is controlled data: you know which information enters the system, how it is classified and where outputs are stored. The second is human review at the points where an error would be costly or hard to reverse. A consultant who treats either as a later detail is passing risk back to you.

A good first engagement may conclude that part of the process should stay manual. That can be the right answer when volumes are low, the work depends heavily on judgement, ownership is unclear, or the systems involved lack dependable interfaces. Such a conclusion is not a failed engagement. It protects you from an impressive demonstration that nobody can run responsibly afterwards.

Measure in production, not in the demo

A prototype or reference architecture can show how a method works. It does not prove return on investment, time saved or a result for any client. Treat claims built only on demonstrations with care, including claims in a consultant's own material. The real test starts when the automation handles live inputs with real exceptions.

Production measurement means agreeing the baseline before launch and choosing a metric tied to the business objective. It also means deciding in advance what would trigger a rollback. Accuracy alone is a weak metric. Also track how often reviewers override outputs, how long exceptions take to resolve, and whether the process owner still trusts the system after several weeks.

Worked example: reading an AI consultant proposal (hypothetical)

Example, invented for illustration only. A mid-sized distributor asks two consultants to automate supplier invoice handling. Proposal A opens with a named platform, a two-week build and a projected percentage of time saved. Proposal B opens with a discovery phase. It maps intake, matching and approval, and flags that the accounting system's interface has not been verified. It also proposes that invoices above a set threshold always go to a human approver.

Proposal B is easier to evaluate. Its unknowns are visible, its review gate is explicit, and the finance lead is named as owner. Proposal A's projected saving is a forecast rather than evidence, and its timeline assumes an integration nobody has checked. The rule of thumb: prefer the proposal that shows you what could go wrong and who handles it.

Frequently asked questions

What should an AI consultant deliver before building anything?

Before any build, an AI consultant should deliver a bounded discovery output. It should cover one mapped business process with named owners, verified and assumed integrations, data classification and boundaries, action permissions, human review points, failure handling with a recovery owner, explicit non-goals and a measurement plan. Facts observed in your environment should be kept separate from assumptions.

Is a working AI prototype proof that a project will pay off?

No. A prototype shows that a method can work under demonstration conditions. It does not establish return on investment, time saved or results in your environment. Those can only be judged in production, against a baseline agreed before launch, using a metric tied to the business objective.

Can an AI consultant recommend not automating a process?

Yes, and sometimes that is the most useful recommendation. Automation may be premature when volumes are low, the work depends heavily on human judgement, ownership is unclear or the systems involved lack reliable interfaces. Keeping part of a process manual can be safer than deploying something nobody can operate or stop with confidence.

Sources and further reading

These resources provide the wider reference frame. Product statements on this page are limited to the public information provided by Victor Laybats.

Who, how and why

Editorial responsibility: Victor Laybats

An automated assistant prepared a first draft. It then passed the published structure, similarity and unsupported-claim checks. Please report any useful correction through the main site.

Method, checks and corrections

Victor LaybatsStart a project