Most guidance about AI in business is a list of capabilities. That's the wrong frame, because it can do almost anything badly, and the list tells you nothing about what's worth doing.
This guide is about the narrower, more useful question — and about the specific ways it costs small businesses money when it goes wrong.
The sequence
| Step | Question | |---|---| | 1. Choose the task | Does verifying cost less than doing? | | 2. Brief it properly | Would a competent freelancer manage from this? | | 3. Know the failure modes | What does it cost when this is wrong? | | 4. Choose tools carefully | Does this add anything the model alone doesn't? | | 5. Write the policy | Has anyone said what's allowed? |
Step 1 — Apply the verification test
Not "can AI do this." Does checking the output cost less than producing it yourself?
Two variables: how expensive a wrong answer is, and how quickly you'd spot it.
| | Error cheap to spot | Error hard to spot | |---|---|---| | Wrong answer cheap | Use freely | Use with spot checks | | Wrong answer expensive | Use with review | Don't |
That bottom-right cell is where most disappointment comes from. It looks like a time saving and it's a liability.
Strong: first drafts you'd edit anyway, summarising material you can check, reformatting, categorising at volume, variations on something you wrote, explaining an unfamiliar topic as a starting point.
Weak: anything with figures you can't verify, anything a client sees unreviewed, judgement calls with real consequences, anything needing current specifics about your business, and regulated advice.
And count the real number: time saved minus time spent verifying. A task that takes 40 minutes, drafted in 2, checked in 25, saves 13 — not 38. Still worth it, and a third of what gets claimed. Right 90% of the time still means checking 100% of it, because you don't know in advance which tenth is wrong.
Full test: where AI actually earns its keep.
Step 2 — Fix the brief before blaming the model
Most complaints — generic, bland, missed the point — are briefing problems.
The test: would a competent freelancer who has never seen your business produce good work from what you typed?
Five parts: context, task, constraints, format, and an example. Most people supply only the task.
One example beats three paragraphs of instruction, because it carries tone, structure, vocabulary and length simultaneously. Two examples beat one, because the model can infer what's consistent rather than copying incidental details.
State what to avoid. Negative constraints rule out the defaults you'd otherwise keep correcting — no preamble, don't hedge, and don't invent figures, flag gaps instead.
Supply the raw material. Asking for a summary of a document you haven't pasted produces invention, and invention is the dangerous mode because it looks like the real thing.
Then iterate rather than restarting. Specific correction — what was wrong, what to change — beats three attempts from scratch.
Full method: why your AI output is mediocre.
Step 3 — Know what actually costs money
Five failure modes with a price:
Fabricated specifics. The expensive one. Figures, citations, regulations and technical claims that are confidently formatted and wrong. Never use a specific you haven't verified.
Pasting things you shouldn't. Client data, contracts, employee information, credentials. Data handling varies by product and by plan — check the tier you're actually on, and set a rule before someone needs one.
Output that reads as automated. Not a technical failure, a relationship one. The cost isn't the email — it's what it signals to a client about how much you value them. The test: would this person be annoyed to learn how it was produced?
Assuming it knows your business. Its confidence is uncorrelated with whether it has the information.
Accountability. If you send it, you said it. "The AI wrote it" is not a position anyone accepts, and offering it makes things worse.
And the slow one: skill erosion. Use it for the parts of your work that aren't the point, and stay sharp on the parts that are.
Full detail: the AI mistakes that cost money.
Step 4 — Don't buy six tools
Try the task in a general assistant you already pay for, with a proper brief. A large share of specialist tools are a form and a prompt around a model you already have access to.
A specialist tool earns its price through integration, workflow, domain data, or volume handling — not through the model underneath, because almost everyone uses the same handful.
And check what your existing software already includes. Most major business platforms have added AI features you're already paying for — the same work down from what you own discipline as any tool decision.
Before subscribing, ask what happens to your data, how you get it out, what it costs at three times current usage, and whether the vendor will exist in two years. Then trial for a fortnight on real messy work, tracking time saved after verification.
Audit quarterly. AI subscriptions accumulate faster than most because they're individually cheap.
Full method: how to choose AI tools.
Step 5 — Write the policy before you need it
Your team is already using it. The only question is whether anyone has said what's allowed.
Write permissions, not prohibitions — a policy that bans everything gets ignored and drives usage underground, which removes your only visibility.
Four things to cover: what data can go in, what must be verified, what gets disclosed, and who owns the output. Plus an approved tool list and a named person to ask when it's unclear.
One page. Reviewed every six months, because this changes faster than most policies you'll write.
Full template: a one-page AI policy.
What this connects to
- Process automation — AI widens what's possible to automate; it doesn't change what's worth automating. The frequency × stability test still governs, and an AI step that's occasionally wrong needs a review step you must count.
- Documentation — a brief that works reliably is a process worth writing down and reusing, rather than retyping forever.
- Where automation stops — the assist rather than automate setting fits AI better than full automation for anything with consequence.
The mistakes, collected
- Asking what it can do rather than what's worth doing.
- Not counting verification time. Overstates the saving by roughly a third.
- Task-only briefs. The commonest cause of generic output.
- Using specifics you haven't verified. The expensive failure.
- Pasting confidential material without checking what happens to it.
- Buying a tool before trying the task directly.
- No policy. People are using it anyway, without guidance.
Where to start this week
Two things, both about an hour.
Write the two-line data rule — what can and can't go into external tools, and who to ask. It prevents the failure that's hardest to undo.
Pick one task you do weekly whose output you could verify in under a minute, brief it properly, and run it for a month tracking time saved after checking. That real number is what should decide whether you do a second one.
The Newsletter
WealthLink Weekly
Business. Money. Marketing. Real Estate. Technology. One email.
One email a week. Unsubscribe anytime.