Specification, done properly.

The practice that issues the spec, exactly as today. When a graduate drafts a clause wrongly and a senior misses it, the liability sits with the practice, not the graduate or the software they used. Avoice is the drafter. A qualified person still reviews, edits and issues.

Avoice never issues anything on its own. Like every spec software provider, it doesn't take on professional liability for the content you issue.

It helps to be clear about what technical authors at a spec publisher actually do. They watch for changes to British Standards, the Building Regulations and contract forms, and update a generic clause library to match: the standards each clause cites, the options it offers, the guidance notes and the generic workmanship clauses.

They never see your project. Turning that library into the specification for a particular building has always been the architect or engineer's job, and the architect or engineer has always been the author of record.

Avoice does each of those jobs too, line for line, and in several places with more information: the full regulations rather than a summary, the contract on the project rather than a standard form, and your own drawings. Then it does the part a library never did: it writes the specification for your actual building, and the architect or engineer reviews and signs it off as author of record.

That's a fair experience to have. AI can get things wrong, and anyone who tells you otherwise isn't being straight with you. Before deciding it can't be trusted, it's worth asking a few questions about how it was tested.

1. When did you last try it? If it was when ChatGPT first came out, or even a year ago, you were using a very different tool. AI models are improving in months, not years. We're used to software that changes a little every year or so. AI changes in weeks, and that takes some getting used to.

2. Which plan and which model? With most software, the free version is the paid version with fewer features. With AI, they can be completely different. Even within one provider such as Claude, ChatGPT or Gemini, the smaller, cheaper models and the most capable ones give very different answers to the same hard question.

3. How was it set up? A general chatbot answering from memory is not the same as AI given your drawings, the regulations and your office master, with tools built to check its own work. Most of what makes AI reliable for specifications is in that setup, not in the model alone.

4. Who ran the test? This one is worth asking gently, because it applies to all of us. Getting good results from AI is a skill, and like any skill it takes time to learn. The same tool can look unreliable to someone trying it for the first time and transform the work of someone who uses it every day. The difference is usually in how it was asked, what it was given to work from, and whether its answers were checked against its sources.

Architects and engineers know this feeling from the other side. Most have had a client, a neighbour or a relative look at a drawing and say the building will fall down, or that the design is wrong. The concern may be sincere, but it isn't the same as a structural engineer's judgement, and nobody would change a design on the strength of it. A judgement about AI works the same way. It carries most weight when it comes from someone who has spent real time with the tools.

So it's fair to ask: did the person who tested it use AI regularly? Did they choose the right model, give it the right information, and know how to push back when an answer looked wrong? If not, the test may say more about the setup than about the AI. None of this means the concerns were wrong. It means the test is worth running again with someone experienced, and we're happy to be that someone on one of your own projects.

The quickest way to settle it is to try it yourself. Even before you look at Avoice, take out a paid plan with one of the main providers, usually around £20 a month, choose its most capable model, and ask it something difficult from your own work.

Technical directors are paid to be cautious, and that's a good thing. Test thoroughly and look at every option. Just don't judge today's AI by what it could do a year ago, or by how quickly other technology has changed in the past. Nothing has moved this fast before.

No. Avoice proposes and you decide. Whether the trigger is a new consultant report, a drawing revision or a client change, every edit arrives as a proposed change showing exactly what would change and why. Nothing goes into the spec until someone accepts it.

No. Every empty field ends up in one of three states:

From the product. Filled from the manufacturer's published data, with the source and a note to double-check shown on hover.
Recommended. A value based on the Building Regulations, British Standards or trade practice, with the reasoning and other options you could pick.
TBC. It genuinely can't be determined from what you've given it, and the reason is stated.

That's how most people feel at first, and it's completely normal. Avoice changes how a spec gets written, not just how fast, so it takes a little time to click.

Just send us an email at hello@avoice.co. We'll book a session and start with one section on one of your live projects, ideally something you know well, like internal partitions or floor finishes. We'll go through it slowly with you: how the draft was written, where each clause came from, what's still open, and how you'd review and edit it.

Between sessions, a few things help it sink in:

Ask the spec questions. Why did it choose that value? Where did that clause come from? What's still to decide? Seeing how it answers is the quickest way to understand how it works.
Check something you've already issued. Run a spec you know inside out through Avoice and see what it finds.
Use AI for small things every day. The more you use it, even for simple questions, the more natural it becomes.

At the end of the day, though, none of it will fully make sense until you try it yourself.