Most of the AI conversations I have with business owners start in the same place: someone on the team has tried a chatbot, it produced something impressive in thirty seconds, and now the question is whether the company should “do something with AI”. The honest answer is that it depends entirely on which part of your business you point it at. AI is very good at some kinds of work and quietly unreliable at others, and the difference is not obvious from a demo.
Where it earns its keep
The safe ground is work where a mistake is cheap to spot and cheap to undo. Summarising a long document so you know whether to read it properly. Drafting a first version of a letter, a job description or a procedure that a person will edit anyway. Turning a pile of similar emails into a table. Extracting names, dates and amounts from invoices into a system where someone still checks the total. Classifying support tickets so the right person sees them first.
What these have in common is that a human is still the last step, and the human can tell at a glance whether the output is right. The AI removes the tedious part, not the responsibility. In my experience this is where the real, measurable time saving is — not in replacing people, but in taking the first draft and the sorting off their plate.
Where I keep a person in the loop
I would not let an AI system, on its own, do anything that is hard to reverse or hard to check. Sending money. Answering a customer with a promise about price, delivery or a legal position. Changing a record in your CRM or accounting system without review. Deciding who gets hired, who gets credit, or who gets a warning. Anything medical, legal or financial where being confidently wrong has consequences.
The reason is not that the models are bad. It is that they are fluent. A wrong answer looks exactly as polished as a right one, and the person reading it has no signal that something is off. Traditional software fails loudly; language models fail politely. So the rule I apply is simple: the harder an error is to notice, the closer a person has to stand to the output.
The rule of thumb
Before automating any step with AI, ask three questions:
- If this output is wrong, who notices, and how quickly?
- What does it cost to undo?
- Can I sample-check ten outputs a week and would that be enough?
If the answers are “nobody”, “a lot” and “no”, the step is not ready for AI — regardless of how good the demo looked.
Before you hand over your data
The second half of AI safety is not about the model at all; it is about where your information goes. Most tools that promise to “read your documents” send those documents to someone else’s servers, and the contract decides what happens next. Before signing, I ask a vendor these questions and I expect written answers:
- Where is my data processed and stored — which country, which provider?
- Is my data used to train or improve your models? Can I opt out, and is the opt-out the default?
- How long do you keep prompts, uploads and outputs, and can I have them deleted?
- Who at your company can read my data, and under what conditions?
- Do you have a data processing agreement I can sign, and does it cover your own sub-processors?
- What happens to my data if you are acquired or go out of business?
A vendor that answers these clearly is usually a vendor you can work with. A vendor that gets vague, or points you at a marketing page, has told you what you need to know.
What this looks like in practice
The businesses I see getting real value from AI are not the ones with the most ambitious plans. They are the ones that picked one boring, high-volume task, put AI in front of a person rather than instead of one, measured the time saved, and only then looked for the next task. That is not a cautious approach; it is the fast one, because nothing has to be rolled back.
If you are weighing an AI tool and want a second opinion on where it is safe to use and what the contract should say, that is exactly the kind of conversation I am happy to have.