Skip to content

How to stop AI producing rubbish for your business

6 min read
Image of Stop ai producing rubbish featured

Most business owners have had the same disappointing AI experience by now.

You ask a tool to write something useful. It comes back quickly. It sounds confident. It uses the right kind of words. Then you read it properly and realise it is vague, generic, slightly wrong, or simply not something you would ever send to a customer.

That does not mean AI is useless. It usually means the AI was given a weak job.

Good AI output does not happen because someone has found a magic prompt. It happens because the system has enough context, clear examples, sensible rules, human review and feedback. In other words, it needs managing.

Why AI produces rubbish

AI tools are very good at producing plausible-looking output. That is useful when the output is based on good instructions and good source material. It is risky when the tool is guessing what you mean.

If you ask AI to “write a proposal”, “summarise this”, “create some social posts” or “draft a policy” without enough detail, it will fill in the gaps. Sometimes that is fine. Often it is where the rubbish creeps in.

For a business, the problem is not just poor writing. Bad AI output can waste staff time, create confusion, annoy customers, misrepresent your service, or quietly introduce factual errors into work that looks polished.

The answer is not to ban AI. The answer is to stop treating it like a mind reader.

Start with context, not clever wording

The most useful improvement is usually the simplest: give the AI the context a competent member of staff would need.

That might include:

  • who the work is for
  • what the customer already knows
  • what your business does and does not offer
  • your tone of voice
  • what information must be included
  • what information must not be invented
  • where the source material comes from
  • what a good finished answer looks like

For example, “write an email chasing an overdue invoice” is open to interpretation. “Write a polite but firm email to a long-standing local client whose invoice is 14 days overdue, keeping the relationship positive, mentioning the invoice number below, and not threatening legal action” is far more likely to produce something usable.

The AI has not become magically smarter. You have given it the shape of the job.

Use examples of good work

One of the most overlooked ways to improve AI output is to show it examples.

If you have three good proposals, two good customer replies, a strong project update, or a well-written report, those are valuable. They show the AI what “good” means in your business.

This is especially important for professional services firms, where the difference between acceptable and unacceptable output is often judgement rather than grammar. A solicitor, accountant, consultant, architect or adviser may not want “friendly and punchy”. They may need careful, measured, accurate and appropriately caveated.

Examples help with that. They also reduce the temptation for staff to keep rewriting AI output from scratch because it never quite sounds like the business.

Set rules before you scale it

If AI is being used casually by one person, a quick review may be enough. If it is being used across a team, or feeding into customer-facing work, you need rules.

These do not have to be complicated. Useful rules might include:

  • never invent prices, dates, legal advice or guarantees
  • only use approved source documents
  • flag uncertainty instead of pretending to know
  • keep drafts in review until a human approves them
  • do not paste confidential client information into public tools
  • use a fixed checklist before anything is sent or published

Those rules are not there to slow everything down. They are there to make AI safe enough to be useful.

A real example: the Cheltenham Times experiment

In my Cheltenham Times AI agent experiment, I gave AI agents a real job: help build and run a local news and information website for Cheltenham.

The early results were fast, but not perfect. Some posts were useful. Some needed steering. Some were not worth publishing. I had to teach the system what counted as Cheltenham, which sources were reliable enough, what tone was acceptable, when a story was too thin, and whether an image genuinely matched the subject.

That is the practical lesson. AI can research, draft, categorise and suggest. It can also make mistakes quickly. The value comes from putting structure around it: clear geography rules, source checks, image checks, draft and review stages, and regular feedback when something goes wrong.

The same pattern applies in an ordinary business. If you want AI to help with sales admin, customer replies, reporting, research, marketing or internal knowledge, the first version will need supervision. You improve the system by correcting it, tightening the rules and saving what works.

Build a feedback loop

Many businesses try AI once, get a mediocre answer, and give up. Others use it every day but never improve the process.

The better approach is to create a simple feedback loop:

  • What did we ask the AI to do?
  • What did it get right?
  • What did it get wrong?
  • Was the source material good enough?
  • Did it need a rule, an example, or a human decision?
  • Should this become a repeatable workflow?

This is how you move from random prompting to a useful business process. You are not just asking AI questions. You are designing a small system that produces better work over time.

Review is not a failure

There is a strange idea that AI is only worthwhile if it works without human involvement. For most SMEs, that is the wrong test.

A junior member of staff needs review. A new supplier needs review. A new process needs review. AI is no different.

The useful question is not “can we remove the human?” It is “can we reduce the boring, repetitive or messy part of the work, while keeping human judgement where it matters?”

That might mean AI drafts the first version of a customer response, but a person approves it. It might mean AI summarises documents, but a professional checks the conclusion. It might mean AI monitors sources and suggests opportunities, but a human decides what to act on.

That is sensible AI implementation: faster, more consistent, but still accountable.

A simple next step

If your business is getting poor results from AI, do not start by buying another tool. Start by picking one repeatable task and writing down what good output should include.

Then gather two or three examples, define the rules, decide who reviews the output, and test it on real work.

If you would like help identifying where AI could produce useful, safe and measurable results in your business, I can help. Book a short AI consultancy call and we can look at one practical workflow where AI might save time, improve consistency, or create something useful without adding unnecessary risk.

In Blog

Want help putting AI to work?

If this article sparked an idea, let's have a short call about what AI could do in your business — no jargon, no obligation.