Every week there's a new headline about AI replacing a job. If you run a small business, it's hard to tell what's real and what's a sales pitch.
I spend most of my working days building AI into real businesses. Here's the part that doesn't make headlines: the list of things it still does badly. Knowing what AI can do matters. Knowing what it can't do matters more, because that's where the expensive mistakes happen.
1. It can't be accountable
AI can draft the quote, the refund decision or the reply to an unhappy customer. It can't own the consequences. When something goes wrong, a person has to explain it, fix it and decide what changes.
That's why every system I build has a named human checkpoint for anything that commits money, makes a promise or goes to a customer in a sensitive moment. The AI does the preparation. A person makes the call.
2. It can't tell you when it doesn't know, reliably
Models are fluent even when they're wrong. A made-up invoice number reads exactly like a real one. A confident answer about your return policy can be a guess dressed up as a fact.
You can reduce this a lot with good design: give the model your real data instead of asking it to remember, make it quote its sources, have a second check compare its output against the original. But you can't assume it will flag its own mistakes. The checking has to be built in.
3. It doesn't know your relationships
Your best customer gets a call, not an email. The supplier who's been with you fifteen years gets some slack on a late delivery. The contractor who did sloppy work last time doesn't get the next job, even though they're cheapest.
None of that is written down anywhere, and most of it shouldn't be. AI can draft the email, but the judgment about who gets what kind of treatment is yours.
4. It can't be on site
Anything physical is out: inspecting a roof, fitting a ski boot, checking whether a unit was left clean, meeting a client for coffee. AI can schedule it, prepare the paperwork and write up the notes afterward. Someone still has to show up.
5. It struggles with messy, unwritten processes
If the way a job gets done lives only in one long-time employee's head, AI can't do it yet. Not because it isn't capable, but because nobody can tell it what "done right" looks like.
This is the most common blocker I see, and the good news is that fixing it helps the business whether you use AI or not. Writing the process down is the first step of every project I take on.
6. It shouldn't make decisions about people
Hiring, firing, performance reviews, deciding who gets credit: these involve fairness, legal obligations and trust. AI can help organize information, but letting it decide is a mistake, both ethically and practically.
7. It can't replace knowing your numbers
AI can build the report. It can't tell you whether a 4% margin is fine for your business this quarter or a sign of trouble. It doesn't know your plans, your bank's patience or what you're willing to risk. Interpreting the numbers stays with you.
Why this list is good news
Look at what's left once you remove those seven things. Reading documents and pulling out the details. Drafting emails, reports and summaries. Answering routine questions from your own information. Checking one list against another. Formatting, filing, chasing, reminding, reconciling.
That's a big share of the hours in most office jobs. It's repeatable, it has clear right answers, and it's exactly the work AI handles well when it's set up properly.
The businesses that get real value from AI aren't the ones that believe it can do everything. They're the ones that draw a clear line: AI prepares, checks and drafts; people decide, relate and take responsibility. Draw that line first, and the rest becomes a practical question of which jobs to hand over.
A useful exercise
List the ten tasks that take the most hours in your office each week. For each one, ask two questions:
- Is there a clear right answer, or does it depend on judgment about people and risk?
- Could you write down the steps for a new employee?
Anything with a clear answer and writable steps is a good candidate for AI. Anything else stays with a person, maybe with AI doing the prep work. Most owners find more of the first kind than they expected.