How to choose an AI implementation partner
By Kole OgundipeWhat you are actually choosing when you pick an AI implementation partner
You are choosing who answers the phone in eight months, when the agent has stopped working and nobody has changed anything. That is the decision underneath all the others, and almost nothing in a pitch tells you about it.
An AI implementation partner is the firm that takes a job your business does by hand and leaves it being done by software that your staff rely on. The work splits in two, and the halves are sold as if they were one thing. AI agent development is the build: the thing that works in a demonstration. AI implementation services is everything after it: access to your systems, a pilot on real cases, somebody's day changing to suit it, and the repair the morning your accounting software renames a field. The second half is where these projects come apart, and the first is what a sales call is built to show you.
So the questions below are not about who writes the better code. They are about who will still be holding the thing when it breaks.
Most advice on choosing an AI partner is written for a company you are not
The guides on this question read as though they were written for a buyer who has a procurement process to lean on and people to delegate the choosing to. If that is you, the standard advice is good advice.
If you are a UK business of five to two hundred people, where the owner is the buyer, the IT department and the person who will have to live with the agent, it is advice that sends you into a call asking for things nobody your size will be given. Enterprise security certification, a named delivery team, references from your own trade: some firms will have one of those and few will have all three, because the firms doing this work well at your end of the market are small and their clients are often unwilling to go on the record.
Asking for those anyway has a cost. The firm that says yes to all three is usually the one telling you what you want to hear, and you have just filtered for it. What follows is written for the smaller buyer: the same rigour, aimed at the answers a firm your size can actually get.
The nine questions worth putting to an AI implementation partner
Ask these on the first call, in roughly this order, and write down the answers. The pattern matters more than any single reply: a firm that is straight with you about the limits of its own work will be straight with you in month nine.
| Ask | What a straight answer sounds like | What to do with a vague one |
|---|---|---|
| What of yours is running in production right now, and who would notice if it stopped this afternoon? | A specific job in a specific business, the volume it handles, and the name of the role that would ring them | If every example is a demonstration or a pilot, you are their production experiment. That can be fine, at a price and with your eyes open |
| Can I watch it do the job on my own work before anything goes live? | Yes, as a stage of the project with a date on it, running on your real cases with nothing sent | A refusal here is the single clearest signal in the whole conversation. Everything else can be recovered from; this cannot |
| Whose accounts does this run on — your cloud, your repository, your model keys, or ours? | A plain answer either way, with the consequence of it said out loud | If it is all theirs and they will not say so, ask what you hold the day you leave. That is the real answer to this question |
| Who repairs it when my software changes underneath it, and is that in the fee? | A named arrangement: included in a monthly fee, or an hourly rate agreed now | "It just runs" means nobody. Your systems change on their own schedule and each change can stop an agent while it still looks healthy |
| What does the second agent cost, and the fourth? | A number, or a flat fee that covers them | If each one is a fresh project, price the next three now. The second job is usually one you spot after watching the first agent work, which is after you have signed |
| What sends a case to a person, and whose job is that queue? | The rule for escalating, and the admission that you have to name the person | A firm that has never raised this has not run one of these in a real business. Escalations that land nowhere are worse than mistakes |
| How will we both know it is working, in month three? | A measure taken from your business before anything is built, and a number to beat | Without a before, there is no after, and the project ends in an argument about impressions |
| What happens when the model underneath is retired? | An acknowledgement that models are withdrawn on the provider's schedule, and whose job the swap is | This one catches out firms new to the work. Models are withdrawn on the provider's schedule rather than yours, and somebody has to act on the notice when it arrives |
| What would make you tell me not to do this? | A real example of work they turned down, or a shape of job they do badly | A firm with no answer sells to everyone, which means the fit between your job and their work was never checked |
The two answers that settle it faster than the other seven
Question two and question four. A firm that will run the agent on your own cases before anything goes live has agreed to be judged on your work rather than its showreel, and a firm that names who repairs it has told you the project has an owner after go-live. Those two together are most of the risk in an AI project.
They are also the two that cost a firm something to say yes to, which is why they discriminate. Any firm can agree it will measure outcomes. Not every firm will put its unfinished agent in front of your real inbox.
What you are entitled to ask for at your size, and what you are not
Scale your demands to the market you are buying in, or you will mistake a sales process for a safeguard.
Worth insisting on, at any size. A written description of the job as it is actually done, including the awkward cases, before anything is built. A list of exactly what the agent can reach in your systems and what it may do there. A pilot on your real work. A named answer to who repairs it. The ability to stop.
Not worth insisting on below a couple of hundred staff. A formal RFP with scored responses, which mostly filters for whoever is best at writing them. Enterprise security certification, which is normal to ask after and not a reason on its own to rule a small firm out. Three case studies in your exact trade. A fixed total price for a job neither of you has scoped yet, which you can have, but you will pay for the uncertainty and it will be scoped narrowly enough to be safe for them.
Worth asking for and worth accepting a no on. One client who will take your call. A sight of the agent they are proudest of. The evaluation set they test with. A no to these is not a red flag on its own; the reason given for it tells you what you need.
Who owns the code, the prompts and the keys when a partner builds your agent
Ask, get the answer in writing before work starts, and understand that either answer can be right. What is never right is not knowing.
There are three honest shapes. You own everything and host it, which gives you the most and asks the most of you: somebody in your business has to be able to pick it up, and that person has to exist and stay. The partner owns and runs it and you pay monthly for it working, which asks almost nothing of you and means you hold nothing if you leave. Or the middle: they run it, and you hold a copy of the prompts, the evaluation set and the description of the job, so a different firm could rebuild it without starting from your own memory.
The thing to settle is not ownership in the abstract. It is this one sentence: on the day we part company, what do we still have? Write the answer down. AI agent development sets out what you should be handed at each stage of a build, which is the same list read from the other end.
What an AI implementation partner costs, and the three shapes the fee comes in
The fee shape tells you more about what you are buying than the number does.
A fixed-price project ends when the agent is delivered. It is the easiest to approve and the one that most readily leaves nobody owning the thing in month six, because the fee bought a delivery rather than a working agent. Time and materials, at a day rate, is honest about scope being unknown and gives you no ceiling. A monthly retainer prices the running of it rather than the building, which is the half that decides the outcome, and the thing to check is what it includes: whether repairs are inside the fee, and what the second agent costs. What an AI agent costs sets out what UK agencies charge to build one and what the off-the-shelf tools charge per month, so you have the going rates in front of you before any of these conversations.
Our own shape, for the avoidance of a sales conversation you did not ask for: One flat monthly fee for unlimited agents. Nothing extra per call, minute or message. Stop any month. We give you the price on the call.
When an AI implementation partner is the wrong answer for you
Three cases. In the first, the job is not written down anywhere and nobody agrees on how it is done; a partner will not fix that for you, and a build on top of it produces an agent that handles the demonstration and not the Tuesday. Settle the job first, with people, then come back.
In the second, the work you want handed over needs judgement that cannot be written as a rule — pricing a difficult quote, deciding whether to keep a customer. An agent can prepare that decision and should not make it, and a firm that agrees to automate it is one to walk away from.
In the third, you have an in-house developer with the time and the appetite. Then the honest answer is that you need a few days of advice and not a partner, and AI consultancy in the UK is the smaller thing to buy.
How to compare two proposals that are not the same shape
Put both on one page, in your words rather than theirs, with four lines: the job each one is taking over, what you have to do, what you hold if it ends, and the total cost over two years rather than the quoted figure.
The two-year line is the one that moves the decision, because the shapes diverge after the first agent. A project price plus a second project plus an unpriced repair is a different number from a monthly fee, and it is usually the larger one. If you are at the stage of drawing up a shortlist, the UK AI agencies we scored on their own published proof is a starting list and the method we used to score it.
What our answers to these nine questions are
Fair is fair: the page is published by a firm that does this work, so here is where we sit on our own questions. Agent Place is an AI automation agency and consultancy based in the UK which builds bespoke agents for operations, sales and marketing. Every agent goes through the same four stages, in this order: scope, build, test, live. Our fee is the monthly shape described in the fees section above, and the sections above are where it is set out rather than restated here.
Everything else on the list above is a question to put to us on the call, in the same words you would put it to anyone else, and the answer you get is ours to give rather than this page's. That includes the three cases earlier on: if your job is one of them, the answer is still not to hire anyone, and that applies to us too.
Frequently asked questions
- What is an AI implementation partner?
The firm that takes a job your business does by hand and leaves it being done by software your staff actually rely on. That covers the build and everything around it: writing down how the work is really done, getting the agent access to your systems, running it on real cases before anything is live, changing what a person does each day, and repairing it when your software changes. A firm that does only the build is a developer, not an implementation partner, whatever the proposal says.
- What is the difference between an AI consultant and an AI implementation partner?
A consultant leaves you with a decision and a document; a partner leaves you with something running. Both are legitimate purchases and they answer different questions. If you do not yet know which job to hand over, a few days of advice is the cheaper start, which is what AI consultancy in the UK covers. If you know the job, paying for a strategy phase first is usually paying to be told what you already said on the first call.
- What should I ask an AI implementation partner on the first call?
Start with what of theirs is running in production today and who would notice if it stopped, then ask whether you can watch the agent do your job on your own work before anything goes live. The second one settles more than the other eight: a firm that agrees has accepted being judged on your work rather than its showreel. The full nine, with the answers worth accepting, are in the table on this page.
- Is it a problem if a partner will not name a client?
Not on its own, and expecting otherwise at the small end of this market will mislead you. The firms doing this work well at the small end are small themselves, and a client who has handed a job to software is often unwilling to say so publicly. What matters is the reason given. "We cannot name them, here is the job it does and the volume it handles, and this client will take your call" is a good answer. A change of subject is not.
- Who should own the code when a partner builds our AI agent?
Either of you, as long as it is written down before work starts, and the question that matters is narrower than ownership: on the day you part company, what do you still have? There are three workable shapes — you own and host it, they own and run it and you pay for it working, or they run it and you hold a copy of the prompts, the evaluation set and the written description of the job. The third is the workable middle, and it is the one a proposal is least likely to spell out unless you ask.
- Does a model provider's official partner badge mean anything?
Less than the weight it is given in a pitch. Whatever a badge requires, it is awarded by the company selling the models, so it is not a judgement on whether anything that firm built is still working in somebody's business a year later, which is the thing you are trying to find out. Ask the nine questions of a badged firm exactly as you would of an unbadged one, and if you want to know what a particular badge involved, ask the firm to tell you what it had to do to get it.
- How much should we expect to pay an AI implementation partner?
It arrives in three shapes — a fixed project price, a day rate, or a monthly fee — and which shape you are quoted matters more than the figure, because a project price covers the building and usually leaves the running of it unowned. What an AI agent costs sets out the going rates UK agencies charge and what the off-the-shelf tools cost per month. Ours is one flat monthly fee for unlimited agents, and Agent Place gives you the price on the call.
- Should we build AI in-house or hire a partner?
In-house wins when you already employ a developer with time and appetite, and when the job is close enough to your own product that nobody else will understand it faster than your team. A partner wins when the first agent has to work before anyone is convinced, because the slow part is not the code — it is agreeing what the work actually is and getting access to the systems it runs on, and a firm that has been through that before moves faster than anyone doing it for the first time.
- What are the red flags in an AI implementation proposal?
A refusal to run the agent on your real work before go-live. No named answer to who repairs it. Every example being a demonstration. A price for the first agent and silence on the second. No mention of what happens to cases the agent cannot settle. And a firm that cannot name a job it would turn down, because a partner who sells to everyone has never checked the fit between your work and theirs.
- What should an AI implementation contract settle before work starts?
Five things, and they fit on one page. What job the agent is taking over, in a description you have corrected. What it may reach in your systems and what it may not. Who repairs it when your software changes, and whether that is in the fee. What you hold on the day the engagement ends. And how to stop. Scope limits and small print beyond that are normal, but if those five are not settled, nothing else in the document protects you.
- How long should an AI implementation take?
The parts that take time are not the building. They are agreeing what the work really is and getting the agent access to the systems the work lives in, and both run at your business's pace rather than the partner's. A firm that quotes you weeks without having asked who holds the admin login for your accounting software has not costed the slow half. AI implementation services sets out the stages and which of them is yours.
- What do we do if the agent breaks and the partner has gone quiet?
Switch the job back to being done by a person on the day it happens, rather than waiting, because the cases keep arriving either way. Then it is a question of what you hold: the written description of the job and the prompts let another firm rebuild it in a fraction of the original time, and without them you are starting from somebody's memory of how it worked. Which is why the question of what you keep is worth settling on the first call and not the last.
Bring us the job
Tell us the work, or the part of somebody's week you would like back. On the call we will say which of it an agent can hold, which stays with a person, what access it would need from you, which part of it worries us, and the price.
Unlimited agents. One flat monthly fee. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are.
Written by Kole Ogundipe for Agent Place, an AI automation agency and consultancy based in the UK which builds bespoke agents for operations, sales and marketing.