AI writes 10,000 lines of code in an afternoon. We still tell most firms to wait.

By Dave Sellick and Jonathan Gaunt 

We build AI tools every single day. That is exactly why we tell most accountants to wait before building their own.

Somewhere out there, a clock is running. Accountants have around 449 days left. After that, so the story goes, generative AI automates the work, replaces the software you pay for and quietly retires the profession.

We build with this technology every day. Dave runs Siggrove, a portfolio finance function for around fifty clients in London. Jonathan built and sold Xavier Analytics to Dext, and now builds Socket. Between us we have built plenty of these tools, and plenty of them are in daily use. We laid out the full case at the AI in Practice Summit, and you can watch the session here.

And we think the clock is a sales pitch. The people setting the countdown are rarely the people doing the building. We are. So here is the view from the inside. This technology will write you ten thousand lines of code in an afternoon, then hand you a mess your firm cannot safely rely on. Both are true at once. That paradox is the whole story. It is why we build every day and still tell most firms to wait.

The AI takeover is a mirage sold by people who don't build

From a distance, generative AI looks like high definition. The tool that finally replaces the software and the staff. Up close, hands on the keyboard, it pixelates. It breaks up. The promise falls apart.

We call it the GenAI mirage, and it has a tell. Here is the test we put to every room we speak in. Name one app or website that has actually improved since generative AI arrived. We are still waiting.

What we see instead is software getting buggier while the marketing gets louder. There is a reason the message keeps escalating. Adoption is real, but nowhere near enough for these companies to be profitable. The wider that gap grows, the harder the countdown gets pushed. Fear sells faster than proof.

It is advanced autocomplete, so the code needs rewriting

So what is doing the building? Underneath the language, these tools are transformer-based and generative. In plain terms, a very advanced autocomplete. It breaks your request into tokens and guesses what comes next, one token at a time, with no view of the whole.

That is why the same technology can feel brilliant and brainless in the same minute, and why its code looks the way it does. You can generate ten thousand lines in one sitting, hand them to a real engineer and hear the same verdict. That is repeated. That is repeated. Throw it away and start again. It builds fast. It rarely builds well. This is the distance between the sales pitch and the keyboard, and the first reason we tell firms to wait.

Keep client data out unless you have real guardrails

This is where the countdown gets dangerous. It pushes people to rush straight past the one thing that matters most.

A calculator on your website is harmless. Everything changes the moment client data enters the picture. One wrong setting can let one client see another client's data. In our world that is a breach. A breach is a professional indemnity claim. That is a massive risk, all to save twenty pounds a month on a tool that already worked.

So we hold one rule without exception. Client data only goes into something we have built when three things are in place. A secure production database. An engineer involved in the build. An enterprise API agreement with a data-processing agreement behind it. Short of all three, client data stays out.

There is a quieter risk too. The technology is costly to run and hard to sustain, and its price is far more likely to climb than fall. Build a core workflow on that and the ground could shift under you.

The value is narrow: deterministic tools that fill real gaps

Everything so far has been a warning. Here is the turn. We build with these tools constantly, because used narrowly they earn their place. The value is real. It is simply sharper and narrower than the countdown wants you to believe.

The sweet spot is replicating a process you already understand into a simple, repeatable, deterministic tool. Take a job you do more than once and can test against an existing spreadsheet. Turn it into something the whole team runs the same way every time. Half an hour saved on a routine task, across dozens of clients, adds up fast.

The next win is filling gaps. An API and a well-scoped tool can do what your core software will not. Jonathan chose PayCircle for payroll for one reason. It had an open API, so any missing feature could be built instead of waited for. The core software could not post a payroll journal split across departments for management accounts. That gap became a tool.

The third is plumbing. Wiring up APIs, connecting systems and pushing notifications used to sit out of reach for most accountants. Now a tool like Claude Code walks you through it. Aim for deterministic automation you can rely on, rather than generative AI running live over your data.

Some of it is small. A marginal-rate explainer that holds no client data and doubles as a lead magnet. A salary and dividend calculator that gives dividends the weight most tools miss. Some of it is serious. At Siggrove, one cash-flow tool grew into a full suite. It covers management accounts, year-end papers, payroll reallocations and month-end schedules, almost all converted from existing spreadsheets. And some of it, honestly, is a confetti cannon fired from a hidden menu in Google Sheets, because learning this means playing with it.

Prove it with an MVP, then hand it to an engineer

Every one of those wins rests on discipline the countdown never mentions.

Prove an idea by vibe coding an MVP. Then hand it to a real engineer to rebuild, before it goes anywhere commercial or near scale. Treat the first version as a sketch. The finished product comes later.

Ask the honest question before you build at all. Do you need generative AI here? Most of what we run sits on older, deterministic technology that is proven and reliable. AI just helps us build within it faster. The core of how both practices run is unchanged.

One last thing, for anyone tempted to write off engineers. It is the same sentence aimed at you. You do not need an accountant anymore. You know how wrong that is. The answer to "can you" is almost always yes. The real question is "should you".

Build what does not exist yet. Leave what works alone

So back to that clock.

We cannot tell you where the pricing lands, or whether the technology gets more sustainable or less. It may end up feeding on its own output and its own mistakes. What we can tell you is simpler. The countdown is a sales tactic, nothing more.

Watch the full session here.