Custom AI development company
A custom AI development company is the more expensive answer and usually the wrong one. Buy the tool if a tool fits. What follows is the honest test for when it does not fit — and what a build has to include so it does not become the project everyone regrets in month four.
Buy the tool if any of this is true
- One system, one workflow, and the tool covers it. You are paying for someone else's maintenance and that is a good deal.
- Your process is genuinely standard. If it looks like everybody else's, someone has already built it better than a first version of yours.
- You need it next week. A subscription starts on Monday; a build does not.
Build when the tool cannot see the whole picture
- The work crosses systems that do not talk. Every off-the-shelf tool is excellent inside its own boundary and blind outside it. When the job needs your warehouse, your accounting and your sales channel in the same decision, no single vendor's tool can see enough to make it.
- The rule that makes you money is your own. Pricing logic, how you judge a return, which customer gets which answer. That is the part a generic tool flattens, and it is usually the part worth the most.
- The volume makes per-seat pricing absurd. Some jobs are cheap per unit and enormous in aggregate. At that scale the licence stops being the small number.
- It has to keep working without you. A tool a person has to drive is not automation. If nobody should need to open it on a Monday morning, that is a build.
Where these projects actually overrun
Almost never in the model, and almost always in one of these four. They are worth settling in writing before anyone starts.
- The data was assumed to be fine. It rarely is. The honest first step is often joining systems that disagree, and it takes longer than the clever part. Budget it as its own piece of work rather than discovering it halfway.
- Nobody defined what finished means. Without a number to hit — this report built by 7am, this share of questions answered without a person — the project ends when the money does.
- The running cost was never counted. Model calls, storage and the machine it runs on are a monthly number, not a one-off. It should be on the table before the build, not after.
- There is no owner on your side. One person has to be able to say the answer is wrong. If nobody owns it, nobody will notice when it drifts.
How to test whoever builds it, including me
Five questions. Any builder worth signing can answer all five in one call, and the answers tell you more than a portfolio.
- Show me something you built that is still running a year later, and tell me what broke in it.
- What will this refuse to do on its own, and how will I hear about it when it refuses?
- What is the monthly running cost at my volume, and what makes it go up?
- If we stop working together, what exactly do I hold — and can somebody else pick it up without you?
- Tell me about a project where you were wrong about the estimate. What did you miss?
What you own at the end
The code and the accounts, from day one. No lock-in, no seat licence on your own process, and nothing that stops working if you stop paying me. That is not generosity: a system you cannot take over is a system you cannot trust to run your business.
Things that were built rather than bought
- Data platform — the case for building rather than buying, in one system: several sources, one true record per product, and an AI layer on top of it.
- Photo factory — a full listing pipeline — prompt-writing, image generation, video and SEO text — built as one machine instead of five subscriptions.
- Personal AI assistant — voice notes transcribed, research done, and every client's history filed into one searchable place.
- Recommendation shelves — cross-sell that maintains itself: every product tagged by its real attributes, 968 of them, and the shelves rebuilt from those tags.
Honest answers
When is custom AI development worth it?
When the job crosses systems no single tool can see, when the rule that makes you money is your own, or when per-seat pricing stops making sense at your volume. If none of those is true, buy the tool — it is cheaper and it starts on Monday.
What does a custom AI build cost?
It is quoted per project, fixed up front, after a free first conversation. The honest driver of the number is the state of your data, not the sophistication of the model: the same job costs a fraction with clean inputs and multiplies when three systems disagree about the same record.
How long does a custom AI project take?
The first working piece in about two weeks, deliberately small enough to finish and real enough to judge. Anything longer than that before you can see something running is a risk being moved onto you.
Who owns the code and the data?
You do — the code and the accounts, from the start. Nothing here stops working if you stop paying me, and there is no per-seat licence on your own process.
What should be in the contract before anyone builds?
Four things, and they are the four places these projects overrun: what finished means as a number, who owns the data work, what the monthly running cost is, and what the system is forbidden to do on its own.
Tell me what no tool covers
Describe the job in a couple of lines and which systems it has to touch. I will tell you plainly whether a tool already does it — in which case buy the tool — and if not, what a build would involve. The first conversation is an hour and it is free.
Book that hour →