At a glance

AI is useful, tireless, and not to be trusted unsupervised. That is not a criticism. It is an accurate description of a talented developer three weeks into their first job, and it is more or less exactly how we treat it.

People ask whether we use AI. We use it constantly. The question worth asking is how, so here is the curtain pulled back.

What a very fast junior gets you

A good junior developer is genuinely valuable from day one. They write code quickly. They know the syntax of every language you throw at them. They will happily take on the tedious job a senior developer has been quietly avoiding for a fortnight. They do not get tired, bored, or defensive about feedback.

AI is all of that, dialled up considerably, and available at eleven o'clock at night.

Give it a clearly specified job with obvious boundaries and it will finish faster than the person who asked for it could have. That is real value, and we would be daft to ignore it out of principle.

What a very fast junior does not get you

Judgement. Context. The instinct that something is wrong before you can articulate why.

A junior does not know that this particular client takes ninety per cent of their orders in a four-hour window. They do not know that the previous developer left something unpleasant in the payments code. They do not know which shortcut is genuinely fine and which one becomes a two-year problem.

And here is the difference that actually matters: a good junior knows when to stop and ask. AI does not. Faced with an underspecified question it will produce something confident and plausible rather than admit the question was underspecified. That single behaviour is why supervision is not optional.

The rules: what AI never touches unsupervised

Three things, and they are not up for negotiation.

Architecture. How the system is divided, what talks to what, where the boundaries sit. Get this wrong and every later decision inherits the mistake.

Data design. The shape of the database, the relationships, the constraints, the indexes. This is the layer that is hardest to change once real data is sitting in it, and AI has no idea how your data is going to grow.

Security and access control. Who can see what, and how that gets enforced. AI reliably writes code that fetches the right record. It routinely forgets to check whether the person asking is entitled to it.

A human makes those three calls, every time, before AI writes a line.

What it does all day instead

Plenty, as it turns out.

First drafts of functions once the design is settled. Boilerplate and scaffolding. Test cases, including the awkward edge cases people forget to write. Converting data from one shape into another. Documentation, which it is genuinely good at and everybody hates doing. The tedious mechanical refactor across forty files. Explaining an unfamiliar library so we do not lose an hour to someone else's documentation.

None of that is trivial work, and all of it used to eat real hours.

Every line gets read

This is the part that makes the difference, and it is the part that gets skipped.

Every line AI produces is reviewed by someone who could have written it themselves. Not skimmed. Read. If we cannot explain why a piece of code is there, it does not ship.

Reviewing code you did not write is a real skill, and this entire way of working rests on it. It is also why "we use AI" and "we let AI build it" are completely different statements that happen to sound similar.

Does it actually save time?

Yes. Not as much as the marketing implies.

The typing gets dramatically faster. The thinking takes exactly as long as it always did. The reviewing is new work that did not exist before. Net, on the sort of projects we take on, it is a solid gain rather than a miracle. Anyone claiming a tenfold speedup is either building something trivial or not reading the output.

The honest bit

We are not precious about this. AI has made us faster, and it has made some genuinely miserable work disappear. We would not go back.

But every project still has a person who understands the whole thing and is answerable for it, exactly as it did five years ago. The tool changed. The responsibility did not. If you want that spelled out properly, how we actually use AI to build software covers the detail.

It comes back to where this series started: AI is a brilliant typist and a terrible architect. Hire the typist. Do not promote it.

Inherited an app and unsure how carefully anyone reviewed it? Our AI Code Audit is a fixed-price £495 review that tells you plainly what you are dealing with, or get in touch for a no-pressure conversation.