How Much Does a Custom AI Project Cost?

How much does a custom AI project cost?

There is no honest single number for AI project cost, because most of the money sits outside the model call: scoping, data work, evaluation, the production build, and a monthly bill that never stops. The pilot is the cheap part. Gartner expects at least half of generative AI projects to overrun their budgets through 2028, and it blames architecture and operations, not the demo.

So the better question for a founder or COO is not "what does AI cost?" It is "where does the money go, in what order, and which parts do people forget to budget?" That is what this post answers. We build these systems for a living, and the same five line items show up every time.

Why does the pilot look cheap and production does not?

In June 2026, Gartner published a report called "10 Best Practices for Optimizing Generative and Agentic AI Costs." THE Journal summarized it on June 22, 2026, and two quotes from that coverage are worth taping to the wall:

• "Creating a production-ready GenAI system can be orders of magnitude more expensive than running a pilot."
• "Through 2028, at least 50% of GenAI projects will overrun their budgeted costs due to poor architectural choices and lack of operational know-how."

We have watched this happen. A pilot runs on a clean sample of data, with one friendly user, no permissions model, and nobody on call when it breaks. It looks great in the meeting. Then someone asks to put it in front of the whole team, and the real list arrives:

• Real data. The pilot used fifty tidy examples. Production gets every malformed PDF, half-filled form and legacy export you have.
• Edge cases. The pilot handled the common path. Production has to do something sensible with the weird ten percent.
• Permissions and data boundaries. Who is allowed to see what, and does the AI respect that every time?
• Logging. When an answer is wrong, can you find out why?
• Evaluation. How do you know it is still right next month?
• A human review step. Where does a person check the work before it matters?
• An owner. Who gets paged when it fails at 7 AM?

None of that is in the demo. All of it is in the bill.

Where does the money go in an AI build?

Here is the order money actually gets spent in a custom AI project, and what goes wrong when a stage gets skipped.

Figuring out the thin slice is the cheapest money you will spend. This is where you pick one workflow, name who owns it, and agree on what success looks like. Most budget overruns we see trace back to a project that never did this, so it grew into "AI for the ops team" instead of "draft the weekly vendor report."

Data work is usually the biggest surprise. The model is rarely the hard part. Getting clean, permissioned access to the data it needs, in a shape it can use, often is. If the data lives in a tool nobody can query, that is a finding, not a delay. An AI readiness check surfaces it early.

The eval harness is the part teams want to skip and should not. Before the build, write down twenty to fifty real examples with the right answer. That set becomes the acceptance test. Every prompt change, model change or new tool runs against it. Without it, every change is a guess.

Build and integration is what most people picture when they think of AI project cost. It matters, but if the first three stages are done well, it is the most predictable part.

The running bill is the one that never ends, which is why it gets its own section.

Why does the running bill matter more than the build?

A build is a one-time cost. Model calls are a monthly cost that grows with use. THE Journal's summary of the same Gartner report says Gartner expects inference, the cost of running a model on each request, to account for at least 70 percent of a model's lifetime costs.

Agents make this sharper. A chatbot makes one model call per question. An agent might call the model a dozen times, pull data from three systems, and retry when a step fails. If you have not budgeted per task, the bill tracks usage you cannot see. We wrote a full guide on how to budget a metered AI bill, including how to set a cost per task and alert on it.

There is a second recurring cost people miss: model retirements. Labs retire model versions on their own schedule. When that happens, someone has to move the system to the new version and rerun the evals to confirm nothing broke. If you have an eval harness, that is an afternoon. If you do not, it is a small project every time a lab ships an update.

What do AI studios and agencies actually charge?

Most AI studios and agencies do not publish a price, and that is not unreasonable: the range of possible projects is huge. AE Studio is a good example of a larger, enterprise-leaning shop. Its homepage, as we read it on October 16, 2026, shows no price or project minimum. It describes its strategy work as "Priced as fixed phases with evals as acceptance criteria, inside a program retainer," and says "Engagements often begin with a reliability diagnostic on the AI a client already runs."

That structure is worth noticing even if you never hire them. Fixed phases with evals as the acceptance test is a sensible way to buy AI work from anyone, because it ties payment to "it works," not "it was delivered."

We publish one number. A 60-minute consultation with us is $250. It is one hour, screen-sharing or audio, where you bring the problem and we work it together. We do not publish engagement prices, because a thin slice for a five-person team and a production system for a regulated product are different jobs, and a range would mislead more than it would help.

How do we keep an AI project inside its budget?

Our agent orchestration methodology is "Figure it out, Build it, Ship it," and most of the budget discipline happens in the first step. Here is what that looks like in practice:

1. Scope one thin slice. One workflow, one owner, one number that will tell us it worked. If we cannot name those three things, we are not ready to build.
2. Write the evals before the code. The test set is the acceptance criteria. When it passes, we ship. When a change breaks it, we know before a user does.
3. Budget the run cost up front. We estimate cost per task before the build, then set an alert so the bill never surprises anyone.
4. Pick the lab for the constraint. We work across Anthropic Claude, Google Gemini on Vertex AI and OpenAI, and the decisive constraint picks the lab. For Smile PreVue, our dental product, the constraint was patient data, so it runs on Google Vertex AI under a BAA. That choice was made on day one, not after the pilot.
5. Keep a human in the loop where it matters. A person reviews the output at the point where a mistake would cost money or trust.

None of that is glamorous. It is what keeps the production cost close to the number everyone agreed on.

If you are still deciding whether to build at all, start with build vs buy for back-office AI. An off-the-shelf tool is often the right answer, and a good studio should tell you that.

FAQ

Is building a custom AI system cheaper than buying an AI tool?

It depends on whether the workflow is yours alone. If a tool already does the job, buying is usually cheaper. Custom builds make sense when the workflow, the data or the constraints are specific to your business.

What costs the most after an AI project launches?

The running bill and keeping the system accurate. Model calls grow with use, and model retirements mean rerunning your evals on a schedule you do not control.

Why do AI projects go over budget?

Gartner's June 2026 report points to poor architectural choices and a lack of operational know-how. In our experience that usually means no thin slice, no eval harness and no run-cost budget.

Can we start small?

Yes. One thin slice is the right first project. Our 60-minute consultation is $250, and it is a good place to find out what that slice should be.

Start with the slice, not the budget

If you want an honest read on what your project would take, book a $250 consultation, or tell us about the project and we will scope the first slice with you. You can also see how we approach AI integration for business on our services page.