AI didn't teach me how to renovate a bathroom. It taught me how to buy expertise.

I’ve spent the past week trying to organise a bathroom renovation.

This is the sort of thing I’m spectacularly unqualified to do. I don’t know how long tanking should take, what a reasonable tiling price looks like, or whether a quote should include adhesive, grout, plywood, pipe fittings and waste removal as a matter of course.

And that’s exactly why AI has been so useful. Not because it can replace the bathroom fitter, but because it can dramatically reduce the information disadvantage between me and the bathroom fitter.

Since I first wrote this, the job has moved from choosing a fitter to preparing for day one. That second phase has made the argument more interesting, because AI’s usefulness did not stop once I had bought the expertise. It shifted from helping me evaluate the expert to helping me coordinate everything around the expert so the work could actually happen.

From a wishlist to a spec

We started with a family wishlist rather than a specification: the bath out, a shower in, a new toilet and vanity, new tiles, a heated towel rail, a mirrored cabinet, and a Thai-style bidet spray.

AI helped turn that into something I could actually send to contractors, and that alone changed the buying process. Instead of asking several builders to “quote for a new bathroom” and getting back three different interpretations of what that meant, I could get them all pricing roughly the same job.

A good deal of that groundwork happened before I’d spoken to a single contractor, using the Checkatrade app now built into ChatGPT. It came back with a shortlist of ten strong candidates, which I then investigated myself — reading reviews, checking past work, deciding who actually looked right for this job. That was a noticeably better result than going to Checkatrade directly, where I kept being pointed towards builders who felt less relevant, some of whom seemed to be there because they’d paid for sponsored placement rather than because they were the best fit. Asking through ChatGPT got me past that, at least this once, and that’s where most of the preamble happened — while I was still thinking out loud, not once I was already on the phone to somebody who fits bathrooms for a living.

It also helped me work out what to ask. Whether the shower enclosure would actually fit. What waterproofing system they’d use. Whether waste removal was included, and who was supplying the building materials. What would happen to the extractor. How the bidet spray would be installed with proper backflow protection. And what happened if hidden work turned up after strip-out.

A number is only useful once you know what it buys

Then the first proper quote arrived: £6,950.

A few years ago I’d have looked at that and asked whether £7,000 sounds reasonable for a bathroom — which is close to a useless question. Instead I could look at what the number actually bought: roughly £700 for strip-out, £2,100 for plumbing alterations and shower preparation, £2,200 for waterproofing, tiling, grouting, silicone and painting, £1,600 for second-fix installation, and £350 for waste removal, with the underlying pipework, fittings, timber, plywood, waterproofing, adhesive, grout and silicone included in those figures rather than billed separately.

I also had an older quote from the year before: a more comprehensive option at £5,640 labour plus £1,400–£2,400 materials. Against that, £6,950 in 2026 including materials wasn’t just something that “felt okay”. It had a benchmark behind it.

Then more quotes arrived. One came in at £5,300 after a site visit. Another initial estimate was £4,700 before the contractor had seen the room.

That was useful in a different way. The question stopped being whether £6,950 was a reasonable bathroom price and became whether the cheaper offers were actually buying the same outcome. A low number isn’t evidence of value if waterproofing, waste, materials or finishing have quietly disappeared from the scope. Equally, a higher number isn’t evidence of quality just because it feels safer.

The work was in normalising what each number meant.

The terms mattered as much as the price

AI also helped with a part I’d probably have skimmed past: the payment terms. The original proposal was 40% up front, 40% when tiling begins and 20% on completion — which meant paying 80% of the job before the final stage. I asked whether that could change, and the contractor agreed to 20% at the start, 50% after first-fix plumbing and waterproofing, and 30% on completion.

I also queried the one-year workmanship warranty. He offered a second year for £450. Working through what that actually covered, my conclusion was to keep the one-year warranty — another year wasn’t attractive enough to pay for.

The best contractor was the one who challenged my brief

Then something happened that made the argument clearer.

The contractor who had initially estimated £4,700 came to inspect the bathroom. He didn’t just look around and confirm the job. He measured it.

I’d included a 1000 × 800mm shower enclosure and tray in the specification. After measuring the room, he told me it was too large for the existing layout. We could use it, but we’d have to move the toilet.

I’d spent hours using AI to research the renovation. I’d turned the wishlist into a detailed specification. I’d compared prices, waterproofing, materials, payment stages and guarantees. I’d specifically thought about whether the shower would fit.

And my specification was still wrong.

That’s not a failure of the process. It’s the boundary that makes the process useful.

AI hadn’t made me a bathroom expert. It had given me enough structure to have a much better conversation with one.

More importantly, I could now recognise something about the quality of the expertise I was buying. The contractor wasn’t simply agreeing to execute my specification. He was interrogating it. He measured, found a problem and then asked me to send him the complete product list so he could check everything against the room before I ordered anything.

That changed how I evaluated the quotes.

The cheapest contractor didn’t become attractive because he was cheapest. He became attractive because he was demonstrating judgement.

In fact, by that point I realised that if his price moved upwards after the inspection, I wouldn’t necessarily see that as a negative. If the extra money bought proper remediation, waterproofing or work needed to solve a real problem, I’d rather pay for the problem to be solved than protect the original estimate.

That’s a different kind of buying confidence. Not knowing enough to dictate the answer, but knowing enough to recognise when somebody is doing more than agreeing with you.

The brief kept getting better because the context kept getting better

The site visits also triggered things that weren’t on the product list at all.

There had previously been water leaking from the bathroom into the flat downstairs. That needed to become an explicit requirement: investigate anything visible during strip-out, pay particular attention to wastes and connections, waterproof properly, and test before everything gets concealed again.

There had also been rats getting into the bathroom through gaps around the pipework, floor or walls. Strip-out would expose those routes in a way I’d probably never get another chance to see, so that became another requirement: inspect the openings and close practical entry points properly before the new bathroom hides them again.

Neither detail came from asking AI to generate a better bathroom specification. They came from the conversation continuing long enough for relevant context to surface, then being captured instead of forgotten.

That’s increasingly how I think about useful AI work. The prompt isn’t the whole input. The quality of the decision depends on what the system can retain, retrieve and route back into the work at the moment it matters.

A product list becomes a specification. A contractor visit changes the product list. A remembered leak changes the contractor brief. A history of rats changes what needs to happen during strip-out. The working context gets progressively richer, and the decision improves with it.

It’s easier on the contractor too

None of this was really adversarial, and I don’t think it should read that way. A contractor doing honest work benefits from a client who arrives with an actual specification rather than a wishlist, who isn’t renegotiating scope after strip-out because nobody agreed what was in it, and who isn’t asking for 80% up front out of anxiety rather than trust. Knowing what you want and knowing what’s fair isn’t only protection for the buyer. It’s less friction and fewer disputes for the person doing the work.

The expert was hired. The coordination problem remained

Once I’d chosen the fitter, paid the deposit and fixed the start date, I assumed the difficult part of the decision was largely over. In one sense it was. I no longer needed to compare contractors or keep optimising the design. But a different kind of complexity appeared around the work: product availability, delivery timing, parking, neighbour communication and access to the building’s water supply.

This turned out to be a useful test of what AI was actually doing for me. It still wasn’t supplying the technical answer. When the original wall tiles were going to arrive too late, my wife chose an alternative range and the fitter calculated the quantities. When I asked about the bathroom light, he chose the size and colour temperature. He rechecked the shower dimensions, built the product basket and decided when the water needed to be isolated. Those were his calls because they depended on trade knowledge or the physical room.

AI’s role was to keep those decisions connected. The replacement tile order had to arrive before the job started. The sanitaryware order was better routed through the fitter’s trade account because he could check the fittings and arrange next-day delivery. The parking permit needed the van registration. The history of leaks and rodent entry still needed to survive all the way to strip-out rather than disappearing underneath a pile of newer logistics.

The clearest example was the stopcock. The water supply can only be isolated from a basement area downstairs. I arranged with the tenants for the fitter to arrive before they left for work, so he could get in and turn the water off. That looked like the problem solved until the fitter immediately pointed out the missing half: he would also need to get back in later to turn the shared supply on again.

AI couldn’t answer that question because it wasn’t really a plumbing question. The fitter owned the technical decision about when the water could be restored. The tenants controlled the initial access. Their landlord could arrange access later in the day. My job was to connect those people early enough that the fitter did not arrive on day one with a building full of neighbours and no way to restore their water.

That is much closer to what I mean when I say prompting is becoming routing. The useful system does not need to pretend it owns every answer. It needs to retain enough context to recognise what kind of problem has appeared, who has the authority or expertise to resolve it, what depends on that decision and what should happen next.

By a week before the start date, most of the renovation uncertainty had stopped being a vague source of anxiety and become a short set of named dependencies with owners: the fitter for technical decisions and the final product order, the council system for parking, the landlord and tenants for basement access, the tile supplier for delivery, and me for keeping the hand-offs moving. That is not the same thing as AI managing the renovation. It is AI helping me manage the edges between people who each know or control something I don’t.

I started this process thinking the interesting use case was AI helping me buy expertise. I now think buying the expertise was only the first half. The second was making sure that expertise arrived in a situation where it could actually do its job.

What it hasn’t done

None of this has made me a plumber. I still wouldn’t know whether someone had tanked a shower correctly by looking at it, and I wouldn’t use AI instead of hiring someone who has spent years learning the trade.

It also hasn’t made me right by default. Asking a sharper question only helps if I can tell a good answer from a confident-sounding wrong one, and AI can’t do that part for me — it can tell me what to ask about backflow protection on a bidet spray, but not whether the answer I get back is actually correct. What changed isn’t my judgement. It’s my vocabulary, and the gap between the two is exactly where I’m still as exposed as I ever was.

The shower measurement made that limitation concrete. The AI-supported specification got me far enough to ask the right question. The fitter’s tape measure supplied the answer.

I made a version of this argument with a different kind of purchase when choosing a family computer. There the risk was an outdated or incomplete price rather than lopsided expertise, but the same discipline carries over: know which facts deserve the expensive check, and how close to the actual commitment that check needs to happen.

Where this stops being straightforwardly good

I don’t think this is unambiguously a good thing, and I’d rather say so than leave it implied. The same week I was doing this, the BBC reported that AI-assisted complaints are straining councils and schools: documents that used to run to a page of A4 are arriving at twenty pages, sometimes misquoting the legislation they cite, and one body estimated it needed double its current workforce to cope.

That’s the same mechanism I used on my bathroom quote — closing the gap between someone who understands a subject and someone who doesn’t. But there it’s pointed at an adversarial relationship instead of a collaborative one, and the effect isn’t fairness, it’s strain: more confident, more voluminous, and not obviously more accurate. My contractor and I both wanted the same bathroom built properly. A council fielding an AI-drafted complaint and the resident who sent it don’t share that alignment, so the same tool that got me fairer payment terms is, somewhere else, getting institutions buried in longer arguments they still have to check line by line.

The difference wasn’t the technology. It was whether the gap it closed sat inside a relationship where both sides gained, or one where only one side was trying to win. The same underlying pattern — AI reducing the disadvantage of being the less-prepared party — shows up outside transactions too, including in personal conflict rather than purchasing, and the same question applies there: is it being used to find the fair position, or to build the strongest possible case regardless of what’s fair.

That’s one renovation, not a study. I’d want to see the same pattern in somebody else’s solicitor’s letter or garage bill before calling it more than a strong personal case.

Buyers and coordinators of expertise

We talk constantly about whether AI can replace experts. I’m increasingly interested in a different use case: AI helping non-experts become better buyers and coordinators of expertise — with builders, mechanics, accountants, solicitors, mortgage brokers, insurers, IT suppliers and consultants, wherever one side of a transaction understands the subject and the other basically doesn’t.

The bathroom experience has made me slightly more precise about what that means. The goal isn’t to use AI until I know as much as the expert. It’s to know enough to structure the decision, preserve the important context, ask better questions, compare the answers, recognise the moment when the expert knows something I don’t, and then make sure the surrounding dependencies do not prevent that expertise from being useful.

AI doesn’t have to close the expertise gap to be useful. Sometimes its real value is making the gap visible enough to use properly — and coordinating the work around it without pretending the gap has disappeared.