Don't choose an AI consultant on the strength of a demo, because a working demo can be built in a day these days. What matters is whether the AI works on your own material and who fixes it when something changes. Seven questions bring out the differences before you sign: proof, pricing, maintenance, data location, running cost, ownership and how success is measured. Then test the answers with a small, fixed-price first step.
Many now offer AI consulting: large consultancies, software houses, advertising and marketing agencies, freelancers and new AI specialists. Proposals and demos often look alike. The differences only show in day-to-day use, when the AI handles messy material every day and the systems around it change. If you don't yet know whether you need a consultant at all, first settle whether a custom AI solution or an off-the-shelf tool makes sense for you.
Choosing an AI consultant in brief
- Proof:
- Ask for implementations in production and a customer you can call. A demo is not proof.
- Pricing:
- A fixed price for a defined outcome shifts the risk to the vendor, hourly billing to the buyer.
- Maintenance:
- Find out before signing who fixes the implementation when a model or system changes.
- Data and ownership:
- Know where the data goes and who keeps the code, instructions and test set.
- Metric:
- Measure the baseline before implementation, or success cannot be shown.
- Getting started:
- A small, fixed-price first step tells you more about a vendor than a proposal does.
What should you get from an AI consultant?
AI consulting is either advisory work, which works out what is worth doing, or advisory work plus implementation, where the same vendor also delivers the plan. Either way, the outcome should be concrete: a defined target, its price and the expected benefit in euros, not just recommendations. First decide which of the two you are buying.
| Advisory | Advisory and implementation | |
|---|---|---|
| What you get | Use cases, a roadmap, training and guidelines for using AI | All of that, plus a working and maintained implementation |
| Suits you when | You have your own development team or already know who will implement | You have no development team of your own, like most SMEs |
| Biggest risk | The plan is never implemented | The vendor only recommends what they can deliver themselves |
You can recognise a good plan by whether it could be handed to an implementer as it stands: every target has a price, an expected benefit and an owner. If the consultant only plans, ask before signing who will implement and who is responsible for the result. Otherwise responsibility falls between two vendors.
When the consultant also implements, the gap between demo and production is what decides it. In practice it looks like this: the demo works on ten sample invoices, while production also has to handle the eleventh, where the reference number is missing and the total is handwritten. A production-grade implementation includes exception handling, a test set, quality monitoring and a plan for who fixes it when something changes. These are exactly the parts missing from the cheapest proposals, and the parts you never see in a demo.
Which 7 questions should you ask before signing?
The following seven questions cover the points where AI projects most often fail after go-live. Ask for written answers so you can compare vendors side by side. The questions assume the vendor also implements. If you are buying advisory work only, put questions 3, 5 and 6 to whoever does the implementation.
1. What have you put into production, and can I talk to the customer?
A good answer names the customer, says what the work delivered and points to a public description you can check. For example, in Metsä Board's transport condition monitoring, sensor data pinpointed where on the route packages were being damaged. A claim like that can be verified. Warning sign: only demos and concepts, or every reference is behind a non-disclosure agreement.
2. Am I buying an outcome or hours?
A good answer says what will be delivered, at what price and what is outside the scope. Warning sign: an hours estimate with no cap and no defined outcome. The differences between pricing models are in the table below.
3. Who maintains the implementation when a model or system changes?
AI automation does not stay in shape on its own. Model providers retire old models: Anthropic gives at least 60 days' notice, after which requests to the retired model fail. A good answer names who monitors quality and tests changes. We covered the causes in why AI automation degrades without ongoing development. Warning sign: "done is done" and a 30-day warranty with no plan for what comes after.
4. Where does the data go, and who can access it?
A good answer says which model provider the data passes through, in which region it is processed, whether it is used for model training and who has access. There is a checklist in AI security for business. Warning sign: the vendor cannot say which service your documents are sent to.
5. What will running the implementation cost per month?
AI model usage is billed by use, so the monthly cost can be calculated in advance from the number of runs and the model's price. Typically it is tens of euros a month per automation. A good answer is a calculation, not a guess. Warning sign: "it depends on usage" with no figures.
6. Who keeps the code, instructions and test set?
A good answer: you do, and the implementation runs in your own environment, for example your company's own Azure subscription, documented so that another vendor could take over. Warning sign: the implementation lives only on the vendor's own platform, and the instructions (prompts) and test set are not handed over.
7. How is success measured, and what is the baseline?
A good answer measures the current state before implementation, such as handling time, error count or working hours per week, and uses the same metric afterwards. Warning sign: "efficiency will improve significantly" with no starting figure.
Fixed price or hourly billing: which protects the buyer?
A fixed price for a defined outcome protects the buyer better, because the risk of overruns moves to the vendor. Hourly billing suits work where the goal cannot yet be defined, but then the risk sits with the buyer.
| Fixed price | Hourly billing | |
|---|---|---|
| Who carries the overrun risk | Vendor | Buyer |
| Suits work where | The outcome is defined and measurable | The goal is still open or changing |
| Buyer's biggest risk | The scope is too narrow and extra work is billed separately | The project drags on with no clear end |
| What to put in the contract | What is outside the scope, and what extra work costs | A price cap and checkpoints |
A fixed price requires the work to be scoped before the proposal. If a vendor gives a fixed price without scoping, the price either contains a large buffer or the scope gets clarified later as extra work. So always ask what assumptions the price is based on.
How should you make the choice in practice?
The safest way to choose an AI consultant is to start with a small, fixed-price step whose outcome can be measured. One or two weeks of work tells you more about a vendor than any proposal: how they scope, how they calculate and whether they tell you what is not worth doing.
- Send the seven questions to 2-3 vendors and compare the written answers.
- Call a reference customer and ask whether the implementation still works and who maintains it.
- Start with a defined first step that produces a target, a price and the expected benefit in euros.
- Agree on maintenance before implementation, not when the first error is spotted.
At Empirica, this first step is the Automation Assessment: a fixed-price analysis that gives each automation target its value in euros, implementation price and payback period. The cost of the assessment is credited if implementation starts within 60 days. Whoever you choose, ask the same of the first step: figures you can base the next decision on.
Frequently Asked Questions
What does AI consulting cost?
It depends on whether you buy hours or a defined outcome. Based on Empirica's implementation history, implementing one automation target typically costs €5,000-28,000, plus model usage, typically tens of euros a month. The running cost is worked through in what running AI costs per month.
Should you choose a large consultancy or a small vendor?
Size is not what decides it. What matters is what the vendor has put into production and who is responsible for the implementation after go-live. The same seven questions work for both.
How long should the first step take?
For a defined target, 1-2 weeks is usually enough. If the first step runs for months before anything is measured, the scope is too broad.
What if the vendor won't answer the questions?
If a vendor cannot answer on data location, maintenance or running cost before the contract, the answers are unlikely to become clearer after it. Ask for written answers before you commit.
Test a vendor with a small step before a bigger commitment
The Automation Assessment is a fixed-price, 1-2 week analysis that gives each automation target its value in euros, implementation price and payback period. The cost is credited if implementation starts within 60 days.
Empirica Finland is a Finnish provider of operational AI and automation that builds automations from the data produced by a company's systems as well as its devices and sensors, and is responsible for keeping those automations running. Empirica is a Claude Partner Network member and a Microsoft partner.
Sources
What are the claims in this article based on?
- Model deprecations
Anthropic, published 7 October 2026
Source for model retirement: at least 60 days' advance notice, and requests to a retired model fail.
These sources were last checked on 19 October 2026.
