How much does AI implementation cost, and what drives the number?

Search for what AI development costs and you get wide ranges, mostly published by firms that sell AI development. I would rather give you the lines the number is made of and the public figures you can price them with, so the estimate is yours and you can defend it to a finance committee. Every external figure on this page comes from a primary source listed at the end, read on the date above, and none of them is 1AYM’s price.

The build is the line everyone asks about, but a budget is only as sound as its least-scoped line: an integration nobody scoped, a data clean-up nobody expected, or a system that works but that nobody changed their day to use.

Each driver below can be estimated before a build starts. If a supplier cannot tell you which of them sets their quote, the quote is a guess.

Firms that sell AI work rarely publish their prices. The public exception is the federal government’s Multiple Award Schedule, where GSA publishes the hourly ceiling rates it has awarded to each vendor’s labor categories, in its CALC+ tool. GSA’s user guide describes them as not-to-exceed prices it has determined fair and reasonable, and says they should not be viewed as exact estimates for a particular location, order size or complexity. An order cannot exceed them, and commercial prices can differ, so read them as reference points rather than quotes.

On 29 September 2026 we queried every labor category CALC+ returns for “artificial intelligence” and “machine learning”: 204 categories from 68 vendors, at their current-year hourly ceilings. The spread matters more than any one figure.

The full range ran from $55.60 to $485.47 an hour. At the median, a 40-hour person-week is $7,629.60 at ceiling.

Compare that with employing the skills yourself. The Bureau of Labor Statistics puts the May 2025 median pay for its combined group of software developers, quality assurance analysts and testers at $134,040 a year, or $64.44 an hour, and for data scientists at $120,230, or $57.80 an hour. Wages and salaries were 70.0% of private employers’ compensation costs in June 2026. Applying that all-industry share to $64.44, a rough loaded cost for that median is about $92 an hour, before recruiting, management, equipment and the months a hire takes to arrive.

Neither number is the answer on its own. An in-house team costs less per hour and more per idle month, because you pay for it whether or not the work is ready. An outside firm costs more per hour and, on a fixed scope, carries the risk of an overrun. What decides it is how long you need the skills and whether the work is defined well enough to price as an output.

A software company adding AI to the product it sells makes a narrower version of this choice, and our guide to AI features in a software product compares partnering, hiring and doing both.

Model usage sounds alarming in the abstract and is often modest for a single workflow. Anthropic’s own pricing page works an example of 10,000 support conversations of about 3,700 tokens each on Claude Haiku 4.5, at $1 per million input tokens and $5 per million output tokens, for roughly $37 in total. Set beside a loaded engineering hour of about $92, that is small, so for one workflow at moderate volume the model bill is unlikely to be the line that decides the budget.

At enterprise scale the arithmetic changes, and the design decides it. At a large international marketing agency, one change, replacing a per-user API call with a single ten-minute sync, carries a forecast saving of £2–4M a year, modelled on token consumption and reviewed by the client’s finance team. The detail is in engagement file D-01. The saving came from the design, which is where run cost at scale is decided.

The run costs that surprise people sit around the model: monitoring, retesting when a vendor retires or changes a model, seats for the people using AI tools, and a named owner’s time every week. Budget them per year, and put a ceiling per run and per month into the design so a runaway loop cannot turn into an invoice.

The same work can be bought three ways, and the shape changes the risk more than the headline figure does. Buy the shape that matches how well the work is defined today.

Whichever shape you choose, ask three questions before signing: what exactly each fee pays for, what the system costs to run for a year after launch, and what you own if you stop paying. A small build fee followed by a monthly charge for workflows held in the supplier’s account can cost more over two years than a fixed-scope build you own outright.

For a US company, how the build is accounted for can matter as much as what it costs. In September 2025 the Financial Accounting Standards Board issued ASU 2025-06, which amends the guidance on internal-use software in Subtopic 350-40 and removes its old project stages. Under the new guidance, capitalisation starts when management has authorised and committed to funding the project and it is probable that the project will be completed and the software used to perform the function intended.

That threshold is not met while there is significant development uncertainty, which the standard ties to two questions: whether novel or unproven functions have been resolved through coding and testing, and whether the company has settled what it needs the software to do. An AI use case that is still an experiment may not clear either. My reading is that discovery gains a second job here, because the decisions it records are the kind of evidence an auditor looks for when that uncertainty is resolved.

FASB expects capitalisation to change little for most types of software, and says it could decrease for software developed to be provided through a cloud computing arrangement. The amendments are effective for all entities for annual reporting periods beginning after December 15, 2027, and interim periods within them, and early adoption is permitted as of the beginning of an annual reporting period. Whether an AI system counts as internal-use software, and how your costs are treated, is for your auditor. This is not legal, tax or accounting advice.

Fill this in before you ask anyone for a quote. A line you cannot fill is discovery work you have not done yet. The reference rates at the bottom are the public figures above, with their date; replace them with your own quotes and payroll data as they arrive.

If lines 3, 4 and 6 are the hardest to fill, the project is not ready to be priced as a build, and the honest next step is a short, fixed-scope discovery that fills them.

If the discovery lines of that worksheet are the ones you cannot fill, because nobody has yet measured the work, checked the data or settled build against buy, that is what our AI Opportunity & Feasibility Sprint is for: a fixed-scope engagement of typically two to four weeks that scores candidate use cases on value, feasibility and risk, and ends in a plan and an investment case your finance team can fund. After it, the build is a fixed-scope statement of work that can start within a day of the scope being signed, or, where a funded programme needs its architecture owned for months, a retained implementation team for typically two to three days a week.

We take small fixed-scope statements of work as well as larger builds, and US clients can contract through 1AYM’s US entity. We do not publish prices, because a price without a scope is the problem this page is about. A 30-minute call is enough to tell you which line of your worksheet we would start on.