Hiring an AI agency vs an independent AI engineer

You have a project — an assistant that answers from your own documents, a workflow that eats a day a week, a process that should not still involve retyping numbers from a PDF. You start looking for someone to build it, and everyone calls themselves an AI agency. Some are twelve people. Some are one person with a good website. The prices range from four figures to six for what sounds like the same sentence.

I am an independent AI engineer, so I am one of the two options being compared here. Read this accordingly. What I can offer that a neutral article cannot is the actual shape of the trade — including, further down, the cases where you should hire the agency and not me.

What you are actually buying from each

The difference is not skill. There are excellent engineers inside agencies and mediocre ones freelancing. The difference is what your money is structurally paying for.

An agency sells you capacity and continuity. Several people can work at once, someone covers when someone is sick, and there is a project manager whose job is that you are told things. You are buying an organization, and part of the price is the organization existing — sales, management, bench time between projects.

An independent engineer sells you accountability and depth. One person holds the whole system in their head, so there is no context lost between the person who scoped it, the person who built it, and the person who tested it. You are buying attention, and the constraint is that there is exactly one of them.

Where an agency is genuinely the right call

This is the part these articles usually skip, which is why they are not worth much. There are projects I turn down, and you should know what they look like:

• You need several workstreams running at once — a mobile app, a backend, and a data pipeline, all landing in the same quarter. One person is a queue, and a queue is the wrong shape for that.
• Your procurement or risk process requires a vendor whose bus factor is greater than one. That is a legitimate requirement and no individual satisfies it, however good.
• The work needs specialisms that genuinely do not co-occur — a penetration test, a design system, ongoing 24/7 on-call. Anyone claiming all of those in one head is selling.
• You want a team you can scale up mid-project. Independents can recommend people, but adding capacity is not something we do quickly.

If two or more of those describe you, hire an agency and stop reading comparison articles. You will save yourself a bad engagement.

Where an independent engineer wins

The inverse cases are more common than most buyers expect, especially for a first AI project in the $5,000 to $50,000 range.

The person you talk to is the person building it. In an agency the senior people are usually in the room to win the work; the build happens elsewhere. That is not dishonesty, it is how the model works. But it means the nuance you explained in the kickoff has to survive a translation, and nuance is exactly what an AI project is made of — which edge cases matter, which mistakes are embarrassing versus expensive, what your team will actually tolerate.

Decisions are cheaper. When the person who wrote the retrieval layer is also the person who tests it, a wrong assumption gets caught in an afternoon instead of a sprint. There is no ticket, no handover doc, no meeting to schedule.

And nobody is learning on your budget. A first AI project is a bad place to train a junior, because the failure mode is not a visible bug — it is a system that looks like it works and is quietly wrong 4% of the time.

The question almost nobody asks either one: who tests it?

This matters more than the agency-versus-individual question, and it is the one buyers most reliably forget to ask.

AI projects fail differently from normal software. Normal software breaks loudly — a page errors, a job crashes, someone notices. An AI system fails quietly and plausibly. It returns a confident answer that is subtly wrong, on inputs nobody tried, and it keeps doing it until a customer or an auditor finds it. A demo cannot detect this, because a demo is a curated input in front of a friendly audience.

What detects it is evaluation: a suite of cases with known-correct answers, run automatically on every change, including all the cases that once went wrong. It is testing, applied to a component that is probabilistic instead of deterministic. It is unglamorous and most vendors do not do it, because it is invisible in a sales meeting.

It is a fair question to ask me too, and the answer is on this site: the IFTA filing agent carries 429 automated tests, including a backtest against a real Kentucky carrier's official state filing that matches to the penny. That is not a claim about my diligence — it is a number you can check.

What it costs, honestly

Published pricing is rare in this market, which should itself tell you something. Here is mine, so at least one side of the comparison is concrete: AI automation projects start at $8,000 as a fixed quote, advisory work is $150/hour, and independent testing on someone else's system starts at $3,000. Every project is quoted in writing before any work starts.

Agency pricing for comparable work usually lands higher for the same scope, because the price includes the organization. That is not a scam; you are getting something for it. The mistake is paying the organizational premium when what you needed was one person paying close attention.

Be more suspicious of the bottom of the range than the top. A very cheap AI build is almost always a thin wrapper around a model API with no evaluation layer, which is the thing that looks like it works right up until it does not.

A checklist that works on either

Whichever way you go, these six questions separate people who ship working systems from people who ship demos:

• Who specifically will write this code, and can I talk to them before I sign?
• Show me tests or evaluations from a previous project. What broke, and how did you find out?
• What happens when the model gets something wrong in production — who notices, and how?
• Which parts are handled by deterministic code rather than the model, and why those parts?
• Is the quote fixed, and what exactly changes it?
• At the end, do I own the repository and the accounts — and can I hire someone else to maintain this?

A good agency answers all six comfortably. So does a good independent. Anyone who gets vague on the second or the sixth is telling you something important.

And if it does not, say so on the call. Telling you an agency is a better fit costs me one project and saves you a bad quarter, which is a trade I will take every time.