You cannot eliminate hallucinations entirely, but you can design work so they are less likely and easier to catch.
The Wrong Aim
Some people look for a perfect prompt that will stop AI making things up. It is an understandable hope, but not a safe operating principle. Better prompting can reduce hallucinations. Source material can reduce them further. Clear instructions can help. None of this removes the need to check important work.
The right aim is not to pretend hallucinations can be abolished. The aim is to make them less likely, more visible and less damaging. It is the natural follow-up to recognising that hallucinations are not rare enough to ignore.
Give the AI Source Material
The safest factual tasks are usually source-bound tasks. Instead of asking AI to explain a policy from memory, provide the policy and ask it to summarise only what appears in the text. Instead of asking for the details of a product, provide the product sheet. Instead of asking for an answer about a contract, provide the relevant clause and ask for a plain-English summary with a warning that it is not legal advice.
A useful instruction is: “Use only the information provided. If the answer is not in the material, say so.” This does not guarantee perfection, but it reduces the temptation to fill gaps.
Separate Facts, Assumptions and Suggestions
One of the most effective review techniques is to ask AI to label its output. For example: “Separate your answer into three sections: facts from the provided material, assumptions you are making, and suggestions for next steps.”
This makes the answer easier to inspect. Facts can be checked against the source. Assumptions can be accepted or corrected. Suggestions can be judged commercially. The business then reviews three different kinds of content instead of one blended, authoritative-looking paragraph.
Ask for Uncertainty
AI often appears more certain than it should. You can counter this by asking it to identify uncertainty. Prompts such as “State where the information is incomplete”, “Tell me what would need verifying”, and “Do not guess if the answer is not in the text” help create a more cautious output.
This is particularly useful when summarising documents, preparing decision notes or drafting advice-like content for internal use.
Create a Verification Workflow
A simple workflow works well for small businesses. First, provide safe source material where possible. Second, restrict the task: summarise, extract, compare or draft from the material rather than inventing from memory. Third, require AI to list claims and assumptions. Fourth, have a human check the important points. Fifth, record approval if the output is external or consequential.
This is not bureaucracy. It is quality control.
Use AI to Help With Checking
AI can help prepare the checking work, even though it cannot replace the checking itself. Ask it to produce a table of claims that need verification. Ask it to identify statements involving numbers, dates, policies, prices, legal duties, promises or third-party facts. Ask it to flag wording that sounds like a guarantee.
This turns review from an open-ended reading exercise into a targeted check — and pairs naturally with breaking work into staged prompts.
Know When Not to Use AI
There are times when reducing hallucination risk is not enough. If the issue is legal, financial, medical, safety-critical, disciplinary or commercially sensitive, AI should not be the final source of truth. It may help draft questions for an adviser, summarise documents for internal preparation or create a checklist of points to discuss. It should not replace qualified advice or accountable management.
A mature AI user knows when to stop.
Practical Steps You Can Take Today
- Add this line to factual prompts: “Use only the source material I provide. If the answer is not there, say it is not there.”
- Ask AI to mark facts, assumptions and recommendations separately.
- Ask for a verification checklist before using any external-facing factual content.
- Use sample or anonymised material when testing workflows.
- Decide which categories of work always require expert or senior review.
Design for Fallibility
A reliable business process does not depend on every person or tool being perfect. It builds in checks where errors would matter. AI should be treated the same way. By providing sources, limiting the task, surfacing assumptions and checking claims, a small business can use AI more safely without becoming paralysed by risk.
The principle is simple: use AI where it helps, but design the workflow as though it can be wrong.
