What is AI consulting?
AI consulting is a time-boxed, paid engagement in which an outside team helps you decide which problems are worth solving with AI, whether your data and systems can support them, and how a first project should be scoped and measured. Its output is a set of written decisions, not software. A good engagement ends with a short list of priorities, an honest view of the data, a pilot definition that someone can build against, and a list of risks to resolve before anyone signs a development contract.
The term covers several different kinds of work, and it helps to know which one you are buying:
| Type of engagement | Question it answers | Typical output |
|---|---|---|
| Use-case discovery and prioritisation | Which of our ideas are worth doing first? | A scored list of use cases, with the reasons for each decision |
| Data and systems assessment | Is our data usable, and where does it live? | A data inventory with quality, access and personal-data notes |
| Approach and deployment advice | Which approach fits, and should it run on a hosted model, our own servers or both? | A recommended approach per use case, with its trade-offs |
| Proposal evaluation and procurement support | How do we compare vendor offers or write a specification? | A vendor-neutral scope and evaluation criteria |
| Risk and governance | What can go wrong, and who is accountable? | A risk register, human-approval points, questions for legal counsel |
| Pilot rescue | The demo worked; why can't we put it into production? | A gap list covering data, integration, acceptance criteria and ownership |
In practice, one engagement often combines two or three of these. What matters is that the contract names which ones are included, so that at the end you can check each of them was delivered.
AI consulting is not the same as a vendor's discovery meeting, a legal opinion or a certification audit. A vendor's discovery scopes that vendor's own project; a lawyer gives the legal opinion; a certification body audits against a standard. A consultant may prepare the questions for each of them but should not claim to replace them.
When do you actually need AI consulting?
You need it when the decision in front of you is larger than the information you have, and nobody inside the organisation has the time or experience to close that gap. These are the situations where an engagement usually earns its fee:
- There is a list of ideas but no way to compare them. Each department has proposed something, and every idea sounds reasonable on its own. What is missing is one set of criteria applied to all of them.
- A pilot has stalled. The model works in a demonstration, but production questions such as integration, permissions, ownership and acceptance criteria were never answered.
- Nobody knows the state of the data. Documents and records sit in different departments and systems, and it is unclear what is usable. The UK government's guidelines for AI procurement put a data assessment before the procurement process itself for this reason.
- Proposals cannot be compared. Three vendors have quoted three different scopes based on three different assumptions, and you need your own requirements in writing before you can compare them.
- Privacy questions are holding up the decision. No one will sign until it is clear which data may be processed where. The Personal Data Protection Authority's guide on generative AI and personal data (November 2025) is a useful starting list of the questions, from the controller's role to transfers abroad.
- A public body must write a specification. The procurement route and the specification shape everything that follows, and the people who will write them may not have bought an AI system before.
When don't you need it?
You don't need a consultant when the problem, the data and the success measure are already clear; in that case, go straight to vendors' discovery meetings and a small pilot. Several other situations point the same way:
- An off-the-shelf product already does the job. Trial it on your own data under a usage policy instead of commissioning a study.
- The problem is not an AI problem. The UK government's AI Playbook (February 2025) asks teams to be open to the conclusion that sometimes AI is not the best solution, and that some problems are solved more easily with established technology. A slow approval chain or a missing integration is a process or software problem.
- You need a legal answer. Questions about lawful basis, contracts or liability belong with a lawyer.
- The decision is already made. If the engagement is meant only to confirm a choice, it will cost money without changing anything; say so openly and skip it.
- Your own team has done this before. A team that has already taken an AI system into production can usually run the checklist below itself and ask for an outside review only where it is unsure.
Need-or-not checklist
Answer each question with yes or no. The pattern of answers matters more than a total score.
| Question | If yes | If no |
|---|---|---|
| 1. Can you state the business problem in one sentence and name the person who owns it? | Continue | Start with an internal workshop; a consultant cannot own the problem for you |
| 2. Do you know where the relevant data lives and who can grant access? | Vendors can scope the work directly | A data assessment is the first deliverable you need |
| 3. Is there a measurable success criterion that everyone accepts? | Vendors can quote against it | Defining it is a core consulting deliverable |
| 4. Are several use cases competing for the same budget? | Prioritisation is worth paying for | A single use case rarely needs a strategy engagement |
| 5. Are personal data, data location or approval rules still open? | Bring legal counsel and a technical review in together | One less reason to engage |
| 6. Do you have proposals that cannot be compared? | A vendor-neutral scope document helps | Ask vendors for discovery first |
| 7. Has anyone in the organisation taken an AI system into production? | A short outside review may be enough | Choose an engagement that includes pilot design |
| 8. Must the decision be defended to a board, an auditor or a tender committee? | Written, reasoned deliverables matter | A lighter format may be enough |
How to read the result: if questions 1, 2 and 3 are all yes and question 4 is no, you most likely need a vendor's discovery and a narrow pilot rather than consulting. If question 2 or 3 is no and the decision is significant, an engagement deserves serious consideration, because those gaps are expensive to discover halfway through a build. If question 5 is yes, plan a legal review whichever route you take.
What deliverables should you expect?
Expect written documents that someone could act on without the consultant in the room. A useful set looks like this:
- Use-case inventory: The same criteria applied to every candidate (expected benefit, data readiness, risk and integration effort), including the reasons a use case was dropped.
- Data and systems assessment: Where each source lives, how current and consistent it is, who can access it and whether it contains personal data.
- Approach and deployment model: For each priority use case, the recommended approach and where the data will be processed.
- Pilot definition: Scope, data, acceptance criteria, how they will be measured and which result would stop the project.
- Risk and compliance notes: Human-approval points, logging, permissions and open legal questions. The voluntary NIST AI Risk Management Framework, built around its Govern, Map, Measure and Manage functions, is a reasonable checklist for this part, and its Generative AI Profile, released in July 2024, adds risks specific to generative systems.
- Roadmap and cost drivers: Phases, dependencies and the items that drive cost, written so that any vendor can quote against them.
Be wary of deliverables that cannot be acted on: a slide deck of market trends, a maturity score without a plan, or a tool recommendation without reasons and alternatives. Agree in the contract that the documents are yours and that you may share them with any vendor.
How is an AI consulting engagement scoped and priced?
A sound engagement is priced against a written scope: which processes and units will be examined, which systems and data sources will be reviewed, what will be delivered and what is explicitly out of scope. The main cost drivers are the number of processes and units involved, the number of systems and how easy they are to access, whether small validation tests on sample data are needed, the amount of on-site work, and the depth of security, personal-data and documentation requirements. Open-ended hourly arrangements are harder to control, because the scope can grow without anyone deciding that it should.
Public bodies have an additional constraint. Article 48 of Public Procurement Law No. 4734 lists services such as software development and preparing technical specifications among the services to be procured from consulting service providers, and states that consulting services are tendered among pre-qualified bidders only; consulting services below a set threshold may be procured through a regular service tender. Agree the route with your procurement unit before you write the scope, because it affects the timetable and the documents you need.
How do you judge an AI consultant?
Judge a consultant by the quality of their questions and the specificity of their past deliverables, not by how fluently they use the latest terms. The scorecard below turns that into questions you can ask in a first meeting.
| Criterion | What to ask | Good sign | Warning sign |
|---|---|---|---|
| Hands-on experience | What have you built and run in production? | Describes a system they built, what broke and how they fixed it | Only strategy documents |
| Willingness to say no | Have you ever advised against using AI? | Gives a concrete example and the reason | Every idea is a good fit |
| Independence | Do you resell a platform or expect to build the result? | Discloses it; deliverables are vendor-neutral and yours | Recommends its own product without alternatives |
| Data first | When will you look at our data? | Early, on a sample, in an environment you choose | Only after a development contract |
| Measurability | How will the pilot be judged? | Proposes acceptance criteria and a test set | "We will see once it is built" |
| Data handling | How will you handle our confidential data? | Non-disclosure agreement, masking, review in your environment | Asks for full exports to its own machines |
| Who does the work | Who will we actually work with? | Names the people and their roles | Senior people at the sale, unnamed people on the work |
| Sample deliverable | Can we see an anonymised report? | Specific, reasoned, estimates labelled as estimates | Generic templates |
Conflict of interest deserves a direct conversation. A consultant who may also build the system has a reason to recommend building. That is not disqualifying, since teams that build tend to know what is feasible, but the deliverables should be written so that any vendor can quote against them, and you should remain free to do exactly that. The UK procurement guidelines' advice to focus on the challenge rather than a specific solution, and to avoid vendor lock-in, applies to consulting output as much as to software. Once the scope is written, our guide to choosing an AI development partner covers the next step.
How should you prepare for an engagement?
Prepare the inputs only you can provide; the work will move faster and the quote will be more accurate. Before the first working session, collect the following:
- The idea list with owners: One line per idea and the name of the person who would use the result.
- A decision date: When, and by whom, the decision to continue or stop will be made.
- Representative samples: A small, approved sample of the documents or records involved, with personal data masked where possible.
- Constraints: Where data may be processed, which systems are off limits and which security reviews the project must pass.
- Earlier work: Previous pilots, vendor proposals and internal analyses, including those that did not work.
Also name one internal contact who can open doors to process owners and the IT department. Waiting for access to people and systems is a common source of delay, so settle it at the start.
Hypothetical example: nine ideas, one budget
This is a hypothetical example. A distribution company's management collects nine AI ideas from its departments. The consultant interviews the process owners, reviews samples of the data and scores each idea on the same criteria. The result might look like this:
| Idea | Finding | Recommendation |
|---|---|---|
| Answering staff questions from internal procedures | Documents are current, owners are known | First pilot, as an assistant that cites its sources |
| Reading delivery notes and matching them to orders | Scan quality varies; rules are clear | Second pilot, after a scan-quality test |
| Demand forecasting | History is short and product codes changed during it | Defer; clean and collect data first |
| Customer chatbot | Most questions are about order status, already shown in the portal | Not an AI problem: improve the portal |
| Automatic supplier selection | Contractual impact; data scattered | Drop for now; consider decision support later |
| Customer-service email triage | Volume and categories unknown | Measure for a month before deciding |
| Contract review support | Confidential and personal data; legal questions open | Legal review first, then reassess |
| Route optimisation | Established non-AI software exists | Buy, don't build |
| Marketing copy drafts | Low risk; existing tools are enough | Use an existing tool with a usage policy |
Only two of the nine ideas go forward to a pilot, and that is a good result. The budget is concentrated where data and ownership exist, three ideas are handled without building anything, and the deferred ones have a clear condition for coming back. The value of the engagement lies in that reasoning, written down so that the decision survives the next budget meeting.
How Hatson Tech approaches this
Our AI consulting engagement is a separate, paid piece of work with a written scope and written deliverables. It is distinct from the first meeting and from the discovery and feasibility phase of a software project, both of which are free. The report is prepared by the same engineering team that builds our own products, including Caseon, our LegalTech product for due diligence, contract review and drafting, which is still in development; that keeps the recommendations tied to what it takes to build them. After the report, you decide whether to continue with us, with another team or not at all. Our article on scoping an AI project is useful reading before any of these conversations. If you are not sure whether you need consulting at all, the free first meeting is a reasonable place to find out.