Should you build your own AI agents?

Getting an agent to work once is not the achievement it feels like. The work starts at the tenth awkward case.

A quoting agent handles a clean enquiry beautifully. Then someone sends a photograph of a handwritten spec, with a part number discontinued two years ago, for a customer who is on different terms to everyone else, and the delivery date they want is a bank holiday. Your prototype either guesses or stops dead. Both are the wrong answer, and deciding which it should do for every shape of awkward is where the fortnight goes.

None of that is clever engineering. It is the patient business of writing down what your best estimator does when the information is incomplete, which nobody has ever written down because nobody has ever needed to.

The prototype that impressed everyone in the meeting sent its output straight out. That is fine when the audience is four people who know what they are looking at. It stops being fine the first time it reaches a customer.

An agent doing real work needs a point where it stops and waits for a person. Which actions need that, at what value, with who allowed to overrule it, and what happens to the queue when that person is on holiday. This is design work and it has almost nothing to do with the model. It is also the first thing dropped when a working prototype needs to be live by Friday.

We wrote about where those checkpoints belong in how we use AI in our own work. The short version is that the approval step is the product, and the model is a component inside it.

This is the one that catches people, and it is the argument I would lead with if I only had one.

Model providers retire models. Not occasionally: regularly, with a few months of notice, on their schedule and not yours. The replacement is better on average and different in the details, which means an agent tuned against the old one has to be tested behaviour by behaviour against the new one and usually adjusted. A prompt that reliably produced the right format now produces something slightly different. A judgement it used to get right nine times in ten it now gets right eight.

That work arrives several times a year whether or not anyone is free. If you built it yourself, the person who does it is the person who built it, on top of their actual job, to a deadline set by a company in California. What happens in practice is that it gets postponed, the agent gets quietly less reliable, someone stops trusting it, and within a few months it is switched off. Nobody writes a post-mortem for that. It just stops being mentioned.

Home-built agents are almost always side projects. Side projects do not get documentation, because the person writing the code already knows how it works and is doing this on top of everything else.

So the agent runs on one person's understanding. That is fine while they are at their desk. It is less fine when they are on a customer site, on leave, or working out their notice. An agent nobody else can explain is a system your business depends on and cannot maintain, which is the same problem as an ageing system with no support, arriving years earlier than usual.

A customer rings up about a quotation they were sent in March. They say the price was wrong. You need to know what the agent saw, what it decided, which rule it applied and who approved it.

A prototype will not have that. Logging every action permanently, in a form a person can read six months later, is unglamorous work that nobody puts in the first version, and it is the difference between an answer and an apology. If you are in a regulated sector it is not optional at all.

Three cases, and I would say so on a sales call.

You have an engineering team with real spare capacity and you want to own this capability because it is close to what you sell. Then build it. Paying someone else to develop a skill you want in-house is a poor trade.

The process is internal and a wrong answer costs nothing. An agent that drafts an internal summary, tidies a spreadsheet or gives someone a starting point does not need approval queues and audit trails, because the person reading it is the check. Build that yourself this afternoon.

You are building something deliberately disposable to find out what is possible. This is the best reason of the three. A fortnight spent building a throwaway agent will teach you more about what to commission than any amount of talking to people like me, and you will be a much better client afterwards.

The bad reason is assuming that because the first version came together in an afternoon, the rest will follow at the same rate. It will not, and the gap between those two rates is where most in-house AI projects quietly end.

Not the build. The build is a fortnight and we charge a fee for it that reflects that.

What the monthly price buys is the part that has no end: the model migrations, the tuning when your rules change, the monitoring, the person whose job it is to notice on a Tuesday in March that something has drifted. We price agents by the processes they run rather than per agent, and you can see what that comes to with the AI agent calculator. The reasoning behind the numbers is in what AI agents cost UK businesses.

An agent is closer to a member of staff than a piece of software. You would not hire someone and then expect them to need nothing from you for three years. The reason to use a company is not that we can build something you cannot. It is that when the model is deprecated in February, keeping your agent working is our whole job and not the eleventh thing on someone's list.