Choosing an AI implementation partner: the questions to ask before you sign
Choose an AI implementation partner by how they would run your first pilot, not by their list of tools. Ask five things: which one process they would start with, how the pilot will be measured, who owns the code, prompts and keys, what happens to your data, and how the work is handed over. A partner with clear answers to all five is worth a call.
This guide is for an owner, director or manager in a company of any size, from a small team to an enterprise, who has decided to try AI on real work and is choosing who to do it with. It is written by an AI and automation consultancy, so read it with that in mind: the questions are the ones we would want a client to ask us.
Start with one process, not a platform
The first answer to listen for is which process a partner would start with. A good one asks about your work before naming a product. A weak one leads with a platform and looks for places to put it.
Automating a broken process makes it break faster. In our reporting automation work, the audit removed several steps entirely before anything was automated. The same is true of AI: the cheapest improvement is often a step that can simply stop.
| Ask | A clear answer sounds like | A warning sign |
|---|---|---|
| Where would you start? | “With the process that takes the most time, after we have mapped it.” | A list of tools before any question about your work |
| What would you not automate? | Steps that need judgement, named | “Everything can be automated” |
| What do you need from us? | A process owner and a few working sessions | Nothing, or “access to all your data” |
Ask how the pilot will be measured
A pilot is a test with an answer. Agree the answer’s shape before it starts: one measure, its value today, and the value that would justify going further.
- One measure. Hours a week on the process, errors per month, time to answer a customer. Not “productivity”.
- A baseline. Measured before the pilot, by the people who do the work.
- A threshold. The value at which you continue, and the value at which you stop.
- A cost on both sides. The build, and what it costs to run each month. Time released is capacity, not cash, unless a cost actually changes.
Write the four down together. If a partner will not, the pilot cannot fail, and a pilot that cannot fail cannot prove anything either.
Check who owns the code, the prompts and the keys
Everything the partner builds for you should end up in your hands, in a form your team can run.
| Item | Where it should live | Why it matters |
|---|---|---|
| Code and configuration | Your repository, as files | A rebuilt server comes back as the same system |
| Prompts and assistant settings | Your repository, versioned | You can see what changed when an answer changes |
| Model and API keys | Your accounts, behind a gateway | You control spend and can switch provider |
| Directory groups and permissions | Your directory | Joining and leaving the company are handled in one place |
In the private AI workspace we built for a mid-sized company, groups, permissions, the model list and every setting are files applied by scripts, and a gateway holds the provider keys and enforces budgets per team. That is the standard to ask for, whoever builds it.
Ask what happens to your data
AI touches documents, mail and customer data sooner or later. Ask about it before the first line of work, not after.
- Where it runs. On infrastructure you control, or a hosted service, and in which region.
- What the model provider sees. Which data leaves your network, and on what terms.
- Whose permissions apply. An assistant that reads mail or files should read them as the signed-in person, with the permissions that person already has.
- What is logged. Who can see questions and answers, and for how long.
- How it is switched off. Removing access, and deleting what was stored.
Ask how it is handed over
The last question decides whether you are buying a system or a dependency. Ask what the handover contains, and ask to see one from a past project.
- Documentation your own IT team can deploy, update and support from.
- An operational guide: what to watch, what an alert means, who to call.
- Training for the people who will use it, built around their own tasks.
- A check that tells the next person, or the next AI agent, whether a fresh environment is ready to run it.
Every Empes engagement ends with documentation and a handover, so the client’s team runs it without us. Whoever you choose, put the handover in the contract.
Compare the answers side by side
Put the partners’ answers in one table. The differences are usually clear.
| Question | Partner A | Partner B | Partner C |
|---|---|---|---|
| Which process first, and why? | |||
| How will the pilot be measured? | |||
| Who owns the code, prompts and keys? | |||
| What happens to our data? | |||
| What does the handover contain? |
A partner with a clear answer in every row is worth a call. A partner with a clear answer in none is selling something else.
Frequently asked questions
How much does AI implementation cost for a small company?
It depends on the process, not on the AI. Price the first pilot the way you would price any automation: what the manual process costs each month, minus what the new one costs to run, against the one-off build. The formula is in the guide on reporting automation. Be wary of a quote given before anyone has looked at the process.
Do we need an AI consultancy, or just an AI tool?
A tool is enough when the job fits what the tool already does and your team can set it up. You need a partner when the job crosses several systems, when data or permissions need care, or when nobody in the team has time to own the rollout. A good partner will tell you when a tool is enough.
How long does an AI pilot take?
Long enough to measure the one thing it was meant to change, and no longer. Agree the length together with the measure before it starts. A pilot without an end date tends to turn into a slow rollout that nobody decided on.
Should the first project be an AI agent or a chat assistant?
Start with whichever touches one process your team already runs every week. A private AI chat helps many people a little; an agent inside one process helps one team a lot. Either way, the first project should be small enough to measure and safe enough to switch off.
What should we prepare before the first call?
One process, written down step by step, with roughly how long each step takes and where errors come in. Add which systems it touches and who owns them. That list lets a partner say something useful in the first conversation instead of a general pitch.
Can we take the work in-house later?
You should be able to. Ask for it in the contract: the code in your repository, the configuration as files, the keys in your accounts, and documentation your own IT team can follow. If a partner cannot hand it over, you are buying a dependency, not a system.
Where to start
Pick the one process that costs your team the most time each week and write down every step, with the minutes each takes, for two cycles. Then ask two or three partners the five questions in this guide about that process. The differences between their answers will tell you more than their websites. If you want Empes to be one of them, bring the list to a scoping call.
Related reading
- Private AI chat for a company: what it is, when it pays, and the decisions that matter
- Reporting automation: cost, architecture and a worked example
- Keeping AI agents under control: a method for predictable agent work
- Empes AI: AI agents and private AI chat for companies in Poland
- Empes Synthesis: business process automation