I work with small businesses. I have watched the AI sales pitch unfold in real time: every tool is now "AI-powered," every workflow can be "transformed," every owner is one subscription away from 10x productivity.

Most of it is noise. Some of it is genuinely useful. Here is my honest read on both.

Where AI actually earns its keep

First-draft writing for things you do repeatedly. Proposals, job postings, onboarding emails, FAQ responses. Not the final version — AI is not better at your voice than you are — but the first pass. If you write the same type of document six times a month and each one takes 45 minutes to start from a blank page, the arithmetic on a good AI assistant is easy. Give it your best previous version, tell it what's different this time, edit the output. Total time: 15 minutes instead of 45.

The key word is "repeatedly." One-off, highly specific writing doesn't benefit much. Recurring formats with predictable structures do.

Summarising things you have to read. Long contracts. Meeting transcripts. Email threads that ran to 40 messages. Dense reports from vendors or accountants. AI is genuinely good at extracting the relevant points from text you'd otherwise have to read carefully yourself. Not perfect — read the summary critically, then skim the source if something seems off — but useful enough to save real time.

Answering questions your team asks you repeatedly. If you answer the same five questions in Slack every week, there is a version of this problem where you tell an AI assistant the answers once and it responds to future questions instead of you. This takes more setup than the previous two and the results depend heavily on whether the questions are actually standardised. But when it works, it genuinely removes a category of interruption.

Where AI does not earn its keep

Client-facing communication that requires relationship. AI can draft it. AI cannot send it without review. And the review often takes as long as writing it would have. The places where clients can tell the difference between a response that sounds like you and one that was drafted by a system — price negotiations, difficult conversations, emotional moments in a project — are exactly the places where you shouldn't be using AI to reduce your involvement.

Anything where the output can't be verified by a non-expert. If AI tells you something about tax law and you don't know enough to check it, you have a problem. If AI writes code for a system you'll rely on and you can't read the code, you have a problem. The rule is: AI is a tool, not an authority. Every output should be reviewed by someone who can tell whether it's correct.

Processes that aren't yet defined. AI is an accelerant. If the underlying process is broken or undefined, AI accelerates the chaos. Before asking "how can AI help with this," ask "do I have a clear process for this that I could explain to a new employee?" If the answer is no, document the process first.

Replacing the judgment calls that actually matter. Which client to take on. Whether this hire is right for the team. Whether the business should expand into a new service. These require judgment that is specific to your business, your clients, your goals. AI doesn't have that context. The danger isn't that AI gives you bad answers — it's that it gives you confident-sounding answers that feel like they've resolved the question when they haven't.

Data entry into systems it can't connect to. I see a lot of small businesses trying to use AI chatbots to do data entry work: transcribing call notes into a CRM, moving information between spreadsheets. Usually this is slower than a human doing it and produces errors that are hard to detect. If the data entry problem is real, the answer is usually a proper integration, not a language model.

The rule I use

AI is useful when the task is structured, repeatable, and the output can be checked by someone with domain knowledge. It is not useful when the task is relational, context-dependent, or when the output can't be meaningfully verified.

Most of the breathless promises fail on one of those three criteria. Most of the genuine wins satisfy all three.

The [AI, Used Carefully](/services#ai) service is one of five things I offer — it's specifically about finding the two or three places in your business where AI saves real time, setting it up properly, and training the people who use it. Without the hype.