What an in-house AI team really costs
The twelve-month cost model, including the six lines that never appear in the business case. Put your own salary figure in and the arithmetic is yours.
A single AI engineer costs their salary plus roughly 25 to 30 per cent in employment costs, plus recruitment, plus the six to twelve weeks before they ship anything useful. Add a manager's time, tooling, and the roadmap your team is not shipping meanwhile. Over twelve months an in-house hire is almost always cheaper than an outside firm for continuing work. Over one project it is almost always dearer, because the ramp-up and the recruitment are paid once and amortised across nothing.
Most business cases for an AI hire contain one number. The salary.
That number is roughly half the real one. Here is the rest of it, and a model you can put your own figures into.
When does this question come up?
Three moments, and they produce different answers.
The board asked what the AI plan is. Headcount gets approved before anyone has defined the job. This is the most expensive version, because the definition arrives after the offer.
A project needs building now. One feature, a deadline, and no spare engineer. Hiring is the wrong tool for this and it is the tool most often reached for.
AI work has become continuous. Several things shipped, more queued. This is the case where hiring is clearly right, and the only case where the twelve-month arithmetic favours it comfortably.
The twelve-month cost model
Fill in your own salary figure. Everything below is a multiplier or an addition on top of it.
Take an AI engineer at salary S.
| Line | Amount | Notes |
|---|---|---|
| Salary | S | The only line most business cases contain |
| Employment costs | 25 to 30% of S | Pension, insurance, payroll taxes, equipment |
| Recruitment | 15 to 25% of S | Agency fee, or your own time if you do it yourself |
| Ramp-up | 12 to 25% of S | Six to twelve weeks before useful output |
| Management | 5 to 10% of S | A senior person's attention, weekly |
| Tooling and inference | Real, and separate | Depends entirely on what they build |
| The roadmap not shipped | Not on any invoice | The largest hidden line, and unmeasured |
So the loaded first-year cost sits at roughly 1.6 to 1.9 times salary, before tooling.
That multiple is the number to take into a budget conversation. Year two drops to about 1.3 times salary, because recruitment and ramp-up are gone. This is why the twelve-month comparison and the three-year comparison give opposite answers.
Comparing it against an outside firm
Same arithmetic, different shape.
An outside engagement has no recruitment, no ramp-up on employment terms, and no year two unless you want one. It costs more per week and it stops.
Work out which you are buying with one question: how many weeks of this work exist in the next twelve months?
- Fewer than twelve weeks. An outside firm is cheaper, and it is not close. You would be paying recruitment and ramp-up to amortise across nothing.
- Twelve to thirty weeks. Genuinely close. Decide on the other factors below rather than on cost.
- More than thirty weeks. Hiring wins on cost, and the knowledge stays in the building, which is the stronger argument anyway.
Most companies asking this question for the first time are in the first bracket and believe they are in the third.
The scorecard
| Compared on | In-house hire | Outside firm | Your existing engineers |
|---|---|---|---|
| First-year cost | 1.6 to 1.9 × salary | Higher per week, and it stops | Already paid |
| Time to useful output | 3 to 6 months | 2 to 4 weeks | Immediate, and slower thereafter |
| Knowledge afterwards | Stays, in one person | Only if handover was contracted | Stays, spread across the team |
| If the work runs out | You carry a salary | It ends | Nothing stranded |
| Systems knowledge | Learned over months | Learned partly, at your cost | Already there |
| Can say nothing is worth building | Almost never | Only if paid for the assessment, not the build | Yes, and they will |
Read the last row twice. A new hire under pressure to justify the role is in the worst possible position to conclude that nothing here is worth doing.
The arrangement that usually works better
Neither column on its own.
Run a short assessment first, from outside, to produce a ranked list with numbers against it. Then hire against a job that is now defined, or find out you do not need to. The assessment costs a fraction of a year of salary and it removes the expensive failure mode, which is hiring for work that turns out not to exist.
If a build is already committed, put two of your own engineers on it alongside whoever you bring in. They will own it in month four regardless, so the only question is whether they learned it while it was being built or after.
How to sanity-check any AI hiring business case
Four questions. Ask them of your own case before somebody else does.
1. What does this person build in their first year? In specifics. If the answer is a category rather than a list, the role is not defined and the ramp-up estimate is fiction.
2. What is the loaded multiple? If the case contains only salary, it is out by something like eighty per cent. Send it back.
3. How many weeks of work exist? Count them. Then halve the number, because everybody overestimates this.
4. What happens if we are wrong? For a hire the answer is a redundancy conversation. For an engagement it is that it ends. That asymmetry has a value and it belongs in the case.
Three mistakes that cost the most
Hiring before the work is defined. The single most expensive one. Three months in, the bottleneck turns out to be data quality or a rules engine, and you are carrying a salary against a job that did not exist.
Counting salary only. Makes the hire look roughly half its real price, so the comparison against any outside quote is not a comparison.
Ignoring the roadmap cost. Your existing engineers doing AI work are not doing something else. That something else had a value, and it is the only line here that appears on no invoice anywhere.
One limit worth stating
This model has no salary figure in it, which is deliberate and is also its main weakness. Engineering salaries vary by several times between markets, so a model with a number baked in would be wrong for most readers. The cost is that you have to do the arithmetic, and a reader who wants a single answer will not get one here. The multiples are the transferable part; the absolute numbers are yours.
What to do this week
Count the weeks. Open the roadmap and count the weeks of AI work that genuinely exist in the next twelve months, then halve it.
If the answer is under twelve weeks, you have your answer and it is not a hire. If it is over thirty, write the job description this week, because the market is slow and you will need the time.
If nobody can count the weeks because nobody has decided what is worth building, that decision comes first and it is cheap. How to tell a good outside firm from a confident one is here, and our own bands are on the pricing page.
If you are sitting on a process that costs more hours than anyone wants to admit, that is the conversation to have.
Book the callWritten by
Radwan Altaf
Radwan runs AISynq. Before that he delivered software inside enterprise programmes at DHL, AT&T, DirecTV and Accenture, which is where the habit of measuring a result against its baseline came from. More about the firm.
Read next
- Best AI consultancies for startups, 2026Seven AI partners a startup can actually hire, what each one is genuinely for, and the two questions none of them answer on their websites.
- AI freelancer or a firm?A good freelancer is the best value here and the highest variance. How to spot one, and the three jobs where a firm is worth the premium.
- How to choose an AI consultancyA four-stage way to evaluate an AI firm, the questions that separate them in twenty minutes, and the five mistakes buyers make most.
Get the next one
New writing in Buying AI help as it goes up, roughly twice a month. One article per email and nothing else in it.