The dangerous thing about AI is not that it is wrong. Plenty of things are wrong. The dangerous thing is that it is wrong in precisely the tone it uses when it is right, so nothing in the output signals that this is the paragraph to check.
After a couple of years of watching this happen in real businesses, the failures fall into five recognisable patterns. Each has a specific tell and a checking habit that catches it in seconds.
1. Invented specifics
The classic. A price, a date, a part number, a percentage, a section of legislation, a British Standard number. It is precise, it is plausible, and it does not exist.
This happens because a language model generates the most likely continuation of the text rather than retrieving a verified fact. Where it does not have the specific, plausible and correct come apart and nothing in the mechanism flags the difference.
The tell: unusual precision on something the tool had no way of knowing.
The habit: every number, date, name and reference gets checked against a source before the document goes out. Make it a rule rather than a judgement call, because judgement is exactly what fluent output disarms.
2. Stale knowledge
A model trained months ago tells you about a product tier that has been renamed, a threshold that has changed or a plan that no longer exists.
This has been especially costly during 2026 in the Microsoft world, where plan names, bundling and pricing all moved in July. An AI answer about Microsoft 365 licensing written from a training set that predates that is confidently describing a world that no longer exists.
The tell: any answer about pricing, legislation, product tiers or government schemes.
The habit: for anything time-sensitive, ask the tool to search rather than answer from memory and check the source it returns. If it cannot cite, treat the answer as a hypothesis.
3. Confidently summarising the wrong document
This is the failure that grows as you connect AI to your own data, and it is the one most businesses do not see coming.
Grounded AI does not invent. It quotes. The problem is that it quotes whatever it found, and what it found was the 2023 price list somebody left in a personal folder, or the draft of the policy rather than the signed version.
The tell: an answer that is oddly specific and slightly off, particularly on prices, terms and lead times.
The habit: ask which document the answer came from, every time. Then fix the underlying cause, which is document sprawl rather than the tool. That is what a data readiness review is for.
Ungrounded AI invents facts. Grounded AI quotes the wrong file with complete confidence. The second is harder to spot because it is technically true of something.
4. Agreeing with you
Ask a leading question and you will get a supportive answer. Explain why this approach is the best option produces a case for it, not an assessment of it. Models are trained to be helpful, and helpfulness reads as agreement.
This matters most in exactly the situations where you wanted a second opinion: reviewing a decision, sense-checking a plan, deciding between suppliers.
The tell: the output agrees with the premise of your question without qualification.
The habit: ask the opposite question as well. Make the strongest case against this. If the objections are weak, that is evidence. If they are strong, you have just avoided something.
5. Tone drift
The least dangerous and the most pervasive. Left unchecked, everything the business writes converges on the same smooth mid-Atlantic corporate voice, and customers notice long before anyone internally does.
The tell: words your business does not use. Leverage, seamless, elevate, reach out, in today's fast-paced landscape.
The habit: keep three examples of your own best writing and give them to the tool as the reference for voice. Then read the output aloud. Anything you would not say, you do not send.
The whole checking discipline in four lines
You do not need a quality framework. You need four habits that take under a minute each:
- Check every specific against a source.
- Ask which document a grounded answer came from.
- Ask the opposite question whenever the output is advising you.
- Read it aloud before it goes to a customer.
Businesses that do these four things get almost all of the value with almost none of the exposure. Businesses that skip them eventually send a proposal containing a price nobody has ever charged.
If you would like help putting checking habits and the technical guardrails behind them into your business — approved tools, grounded data, retention and audit — that is standard work for our IT consultancy team. Start a conversation.
