Custom AI pricing benchmarks, 2026

If you search what custom AI costs in 2026 you will be told, on page one, that it is $5,000 and that it is $2 million. Both pages look authoritative. Neither is lying. The problem is that nobody states which market they are pricing, so the buyer — usually a 10 to 500 person company — comes away with a range so wide it is useless for a board conversation.

So here is the tedious version. Published numbers only, each attributed, with the definitions attached. Use it as a reference, quote it, argue with it. Everything is checkable.

1. My published bands

These are not indicative. They are the ranges the estimator on this site returns, and the ranges a fixed-price quote from me lands inside once discovery is done. I publish them because a buyer who cannot get a number out of an agency’s website has already learnt something about that agency.

Source: the AI Makers cost estimator. Bands are before the multipliers in the next section.

2. The two multipliers that move the number most

The base band tells you what kind of thing you are buying. What decides where you land inside it — and whether you land outside it — is almost never the model. It is how many other systems the thing has to talk to, and where it has to live. Those are the two multipliers in my estimator, and I publish them because they are the part every buyer is trying to reverse-engineer.

Worked through: an automation project touching nine systems and deployed on your own servers is €9,950 to €49,950, times 1.6, times 1.15 — roughly €18,300 to €91,900. That arithmetic is the whole formula. There is no fifth secret variable, and if an agency will not show you theirs, the reason is usually that it ends in a rate card rather than a range.

The reason the systems count dominates is that each integration is a small project: agree the contract, handle the errors, handle the retries, handle the day the vendor changes the API. I wrote up how that plays out in practice in connecting an AI assistant to your business tools.

3. What everyone else publishes

Four sources, each linked, each quoted exactly. I have converted nothing between currencies and rounded nothing.

Sources: Kellton, EncodeDots, SFAI Labs, all read 18 August 2026, and the AI Makers estimator.

4. Why the ranges disagree by more than 10x

Set the tables next to each other and the gap looks absurd. Kellton calls a chatbot $50,000 to $150,000. EncodeDots calls a chatbot $5,000 to $20,000. That is a factor of ten on the same noun. Three things explain almost all of it.

The word describes different objects. Kellton’s chatbot tier sits alongside computer vision at $600,000 and recommendation engines at $1.5 million — it is an enterprise deployment with security review, procurement, integration to systems of record and a support organisation behind it. EncodeDots’ chatbot is a support bot on a website. The same word, two different products, both priced correctly.

The buyer decides the price more than the build does. Notice that Kellton also publishes a starter tier at $5,000 to $50,000 for smaller companies. The engineering is not ten times harder for the enterprise version; the process around it is. Requirements sign-off, security questionnaires, vendor onboarding, steering meetings, a project manager and a delivery lead who never write code. If you are a 40-person company you are not buying that, and you should not be paying for it.

Nobody says which currency of risk they are quoting. A range on a marketing page is a starting point for a conversation. A range in a fixed-price quote is a commitment. Mine are the second kind, which is why they are narrower and why they have multipliers attached instead of a rate card.

Once you filter for tier, the SMB numbers across independent sources converge remarkably tightly: $5,000–$20,000 (EncodeDots) and $5,000–$50,000 (Kellton) against my €4,950–€19,950. Three companies with no reason to agree, describing the same slice of the market, land in the same place. That is the number an SMB should plan against, and the rest of the spread is a different industry’s pricing accidentally sharing a keyword.

5. Rate and salary benchmarks

Two ways to sanity-check any quote: what the hours cost, and what the people cost.

Salary figures: Stack Overflow Developer Survey 2025. Rates: SFAI Labs and Kellton, as linked above.

The salary line is the one worth sitting with, because it is the honest floor under the whole market. A US AI/ML engineer’s median total cost to an employer is meaningfully above the $189,500 headline once you add employer taxes, equipment and benefits. Hire one and you have committed to that number for a year before a line of code is written. A €30,000 fixed-price build is a fraction of one quarter of that person — which is the actual comparison an SMB is making, and I made the long version of it in AI agency vs in-house.

6. What it costs after launch

The build number is the one everyone quotes and the smaller of the two over three years. Here the published sources agree closely, which is unusual enough to be worth trusting.

The gap between my €50–€500 and SFAI’s $200–$5,000 is a volume gap, not a disagreement: an SMB support assistant handling a few thousand messages a month and an enterprise system serving a million requests are not the same meter reading. Take the percentage figures as the planning number and the monthly figures as a range to place yourself inside. The mistake I see most often is a business budgeting the build and nothing else, then treating year-two maintenance as an unpleasant surprise rather than the 15 to 30% every published source told them to expect.

7. What actually moves your number

Ranked by how much I see each one swing a real quote, largest first.

1. Systems involved. Up to ×1.6 on my own bands, and the effect is the same everywhere. Integration is the work; the model is a line item.
2. Blast radius. Software that drafts for a human to approve is cheap. Software that acts on its own needs evaluation, guardrails, audit and a fallback path, and that is a different budget.
3. The state of your data. If it lives in one clean system, fine. If it lives in three spreadsheets, a legacy database and someone’s head, the first weeks go on making it usable before anything intelligent happens.
4. Regulation. Financial services, healthcare and public sector work carries an audit-and-evidence tail that is real engineering, not paperwork.
5. Interface. A chat window is the cheap answer. Something embedded in your own product is design and front-end work that can be a third of the project.
6. Where it runs. ×1.15 for on-premise on my bands. Usually worth it if it is what gets the deal through your own security review.

Two adjacent references if you are costing a specific shape: what a custom AI agent costs for anything that takes actions rather than answers questions, and what an AI employee really costs for the subscription-shaped end of the market, where the sticker price and the total are furthest apart.

How to use this page

1. Find your tier, not your keyword. Decide whether you are the 40-person company or the 4,000-person company first. Every published range is only meaningful inside that answer.
2. Take the base band, then apply the multipliers. Count the systems honestly — people forget the accounting package and the thing the warehouse uses.
3. Add 15 to 30% a year. Every independent source in the table says so. Put it in the business case at the start and the project survives its second year.
4. Then get a real number. The estimator applies all of the above in about thirty seconds, or design the system yourself in the builder and send it in for a written fixed-price quote. The spec is yours either way — take it to anyone.

Common questions

Related reading

• Custom AI development cost — the interactive estimator
• Custom AI development in 2026: real cost, timeline and how to pick a developer
• What an AI employee really costs in 2026
• Custom AI agent development cost 2026
• How to estimate cost and time for any custom AI project
• AI agency vs in-house: the real comparison