Hire AI developers when AI is the core of your product, the work will continue for years, and you have someone who can manage and judge their output. Work with an AI development company when you need a defined system built, tested and handed over or run, especially a first version, or when nobody in-house can yet tell good AI work from bad. Plenty of businesses do both in sequence: a firm builds and proves the first version while the in-house team grows to take it over. Whichever you choose, test the people before you commit, because AI systems fail in ways an ordinary interview never shows. Below is how the two options split the risk, how firms are engaged and paid, what drives the cost, and four tests we would give any candidate.
What buyers are searching for
When we looked at a year of Google search interest for buying terms, the fastest growth in this area was in hiring. Searches for "hire ai developers" more than doubled in India and rose by about four-fifths in the United States over the year, alongside "hire dedicated developers" and "hire remote developers". On the company side, people searched for "best software development company", "ai software development company" and "custom software development cost", the last of them rising.
Both groups are asking the same underlying question: who should build this, and how do I know they can? The answer depends less on the technology than on risk.
The real difference is who carries which risk
| Question | Hiring developers | Working with a development company |
|---|---|---|
| Who manages the day-to-day work? | You | The company |
| Who carries the risk that it does not work? | You | Shared, according to the contract |
| What happens when a key person leaves? | You recruit and the knowledge may leave with them | The company's problem, if the contract says so |
| Who decides the architecture? | Your team, ideally someone senior | The company, with your approval |
| How fast can work start? | After recruiting | After agreeing scope |
| Who owns the code? | You, as employer, usually by default | Whatever the contract says, so write it down |
| Who maintains it after launch? | Your team | A retainer with the company, or your team after handover |
The row that catches people out is the one about judging the work. A first AI system needs someone who can tell whether it is reliable: whether it invents answers, whether it breaks on odd inputs, whether it has more access than it should. If nobody on your side can judge that yet, hiring developers puts you in charge of work you cannot assess. A firm with a track record of testing can fill that gap while you build the skill in-house.
The ownership row deserves its own reading if you work with a firm in India or elsewhere abroad. We set out how copyright in code works and what the contract should say in who owns the code when you hire a software company in India.
The four ways a development company is engaged
A development company can be paid in four ways, and we use all four with our own clients:
- Fixed price, with the scope agreed up front. It suits work that can be specified precisely.
- Time and materials, billed against the days worked. It suits work whose shape will change as you learn.
- A monthly retainer. It suits a system that needs steady improvement and support after launch.
- A subscription to a product the company builds and runs. It suits needs that a ready-made product already covers.
The first two are the choice most buyers face, and the trade-offs are bigger than they look. Fixed price moves the risk of overruns to the supplier, who prices that risk in and then resists changes. Time and materials keeps the price honest but needs you to watch progress closely. We compared them in detail in fixed price vs time and materials. For AI work, where you often learn what the system needs only after testing it on real data, a fixed-price discovery phase followed by time and materials is a pattern worth considering.
What drives the cost either way
We do not publish figures, because the cost of an AI system depends far more on what surrounds the model than on the model. The four things that move the price most in our experience are the same whether you hire or contract:
- Integrations with the systems you already run: payment gateways, government portals, older internal software and anything without a modern interface.
- Compliance: data protection obligations, where data must be stored, and what has to be auditable.
- Data migration: moving years of existing records out of an old system, which usually need cleaning first.
- The number of people who must approve each decision. Every extra approver adds time to every decision.
Notice that none of these is "which AI model". When you compare quotes or salaries, compare how each option plans to handle these four, because that is where estimates go wrong.
Four tests to give any AI developer or firm
These come from experiments we ran on our own systems and published. Each takes under an hour to set up and tells you more than a portfolio.
- The unanswerable question. Give them a short document and two questions, one it answers and one it does not, and ask how their system will respond to both. In our hallucination test, current models answered six of nine unanswerable questions with facts from outside the text, until a single instruction line fixed it. A good candidate knows this happens and has a method to measure it.
- The messy extraction. Hand them ten real messages from your customers and the fields you need, including a few traps: a budget in another currency, a date where you expect a duration. In our structured output test, broken JSON never happened and wrong values did. Ask how they will check that each value is right as well as well-formed.
- The permission map. Ask them to list every action the system's credentials will allow, and which of those it actually needs. OWASP names excessive agency among the top risks for AI applications. A good answer separates reading from acting and keeps a person approving anything that sends, pays or deletes.
- The evaluation set. Ask how they will know the system still works after the next change. The answer you want is a fixed set of real cases, including past failures, run after every change. The US standards body NIST treats this kind of ongoing measurement as a core part of managing AI risk in its AI Risk Management Framework.
A developer or firm who answers all four with specifics has built AI systems that had to work. One who answers with the name of a model or a framework has not been tested by real users yet.
Questions to ask before you sign
- Who will own the code, the data and the prompts, and is that in the contract?
- How will you test the system before launch, and what will you show me as evidence?
- What can the system do without a person approving it?
- What happens after launch: who fixes problems, how fast, and on what terms?
- What would you advise us not to build? A good partner says no to something.
- How will you price changes we discover along the way?
If you are weighing whether to hire or to work with a firm for an AI project, we are happy to talk it through, including when the honest answer is to hire. Tell us about the project.
