Asimov wrote fiction. We build reality: the AI laws a business needs.
Asimov's robot laws were a plot device. Ours run in production: four rules that keep AI useful and honest in business software, with costs and limits stated plainly.
Clients ask us one question more than any other at the moment: how can AI actually help our business? It deserves a better answer than either camp offers: the evangelists promise magic, the sceptics promise disaster. Our answer starts somewhere unexpected: with a science-fiction writer, eighty years ago.
In 1942 Isaac Asimov gave his fictional robots the Three Laws of Robotics: do not harm humans, obey orders, protect yourself, in strict order of priority. Here is the part people forget. He then spent the next forty years writing stories about how those laws fail: edge cases, conflicting orders, loopholes. The laws existed to show why declared rules, on their own, are not enough.
We build AI into business software every week: assistants that answer questions, systems that read documents and listen to recordings. So we wrote our own laws. The difference is that ours are not a policy document. They are built into the machinery the software is made on, not pinned to a noticeboard. And, learning from Asimov, we added a law that outranks all the others.
Law Zero: assume no AI is perfect
Everything else follows from this. In a recent client pilot we asked the system to read interview records — typed notes, a Word form, a scanned page of handwriting — and extract the answers to a fixed set of questions. It scored 41 out of 42. That is genuinely useful. It is also not 42. So the working rule we gave the client was simple: the AI does the reading, people check the record. A supplier who says their AI needs no checking has just told you a great deal about the supplier.
Law One: suggest, never commit
Nothing an AI produces becomes real in your business until a person makes it real. Our in-app assistant will happily draft three follow-up tasks from one conversation, but each lands as a pre-filled form that a person reviews and saves. A public booking assistant we run can create appointments only in a provisional state that a person must confirm. In our demonstration sales system, the assistant can propose a discount, citing the business rule that caps it; only the signed-in manager can apply it. The AI drafts. You decide.
Law Two: see only what your user sees
An assistant answers as the person asking. It goes through the same permission system as the buttons and pages, with no side door. Ask "how is the pipeline looking?" as a regional manager and you get your region; ask as a director and you get the whole company. Same assistant, same question; the difference is the authority of the person asking, because the authority was never the assistant's. Information marked above a sensitivity threshold is stripped out before the AI ever sees it. And when the AI is refused, it says so honestly instead of pretending the data does not exist.
Law Three: show the working, and the doubt
Every answer keeps its receipts: the record of what the AI was asked and what it did is kept, every call is logged, and sources stay checkable. When the answer is not there, this law requires the gap to be admitted. In that document pilot, the one miss out of 42 was the system reporting that it could not find an answer, where inventing one would have looked better. We scored that a pass. What you will not get from us is a decorative confidence percentage. We would rather the AI said "I could not find this" than "87% sure".
What it costs
Less than people expect, with an honest "it depends". The three-document pilot above cost about fifteen pence in AI charges, end to end. Assistant questions work out at roughly two pence each. Across a month of real use, our working estimate is £2 to £20 per user, driven almost entirely by how much people lean on it. The pattern to notice: the cost scales with use, so trying it is cheap and the bill only grows with the value.
The honest limits
AI errs confidently; that is why Law One exists. Poor handwriting and muffled recordings set real floors on accuracy. Some inputs we refuse outright rather than guess at. The models themselves change over time, so testing is a routine, never a one-off. And some work should not be automated at all; part of our job is saying which.
How we test deserves its own post: the same source material pushed through typed notes, documents, handwriting and audio, and rival AI models measured against each other on your actual tasks, with the results shared so you can choose the trade-off knowingly. That is the next post.
Asimov's laws made great stories because they broke. Ours are duller, and that is the point. If AI is going to work inside your business, it should work like a good employee in their first week — useful, supervised, and honest about what it does not know.
Curious what AI could do in your business, and what it should not? Book a discovery call. Half an hour, no charge, and we will bring real numbers, no magic.