A founder runs the question through ChatGPT. He’s deciding whether to raise the firm’s minimum project size and wants a sanity check before talking to clients. The model produces two confident paragraphs that mirror his own thinking, stated more fluently. He reads it, nods, and sends the announcement.
Three months later, two long-standing clients have walked and the pipeline is thinner than it was.
The issue sat in the prompt. He’d asked for validation of a direction he’d already picked, and a large language model will almost always provide it, fluently and without hesitation.
What’s the choice you’re actually facing?
Founders usually mean one of two things when they talk about using AI for decisions. The first is using AI to challenge a plan: asking it to surface objections, model alternative scenarios, or argue the opposite case. The second is using AI as a productivity shortcut: summarising research, drafting options, or turning existing thinking into communication. Both are legitimate. They suit different situations.
The distinction matters because large language models are designed to produce plausible text, not to verify truth. The UK’s National Cyber Security Centre warns that LLMs “can produce false, misleading or biased information with confidence, even where they have been trained on high-quality data.” Ask a model to validate your thinking and it will produce confident validation. Ask it to challenge your thinking and it will produce confident challenge. The design of the prompt determines what you get.
That job distinction carries more weight than founders typically give it at the moment they open the chat window. Calling AI in as a challenger is a different act from calling it in as a scribe, and confusing the two is where many decisions quietly go wrong.
When should you use AI to challenge your decisions?
AI works well as a decision-challenger when the decision is material but not existential, you have some concrete data to anchor the model, and you can run several prompts before you act. Pricing changes, service restructuring, early-stage hiring plans, and marketing positioning all sit in the right territory. You’re stress-testing a direction, not asking the model to set one.
The Competition and Markets Authority has noted that AI is well-suited to analysing large information sets and generating options, provided humans stay meaningfully involved in the final call. Prompts like “argue against this plan”, “list the operational risks I haven’t considered”, or “give me the pessimistic scenario for this decision” pull significantly more value from a model than “is this a good idea?”
Running three to five independent prompts on the same decision gives you a range of perspectives rather than a single plausible view. NCSC advises treating model outputs as one input among several, cross-checked with other sources. You’re using AI the way you’d use a well-briefed sceptic: give it your plan, your assumptions, and your constraints, then ask it to find the cracks.
A 2025 Enterprise Nation survey of 1,000 UK SME decision-makers found that 39% of AI-using SMEs use tools for brainstorming, a function that maps directly to structured assumption-testing when the prompts are designed for challenge rather than confirmation.
When should AI stay as a background assistant?
Keep AI in assistant mode when the primary goal is speed, clarity, or documentation rather than challenge. Summarising research, turning verbal reasoning into a structured memo, drafting communications, or cleaning up internal documents are lower-risk uses that suit AI’s strengths without the verification overhead that structured decision-challenging requires. A 2025 Enterprise Nation survey found 47% of AI-using SMEs use it primarily for research and summarisation.
The clearest case for staying in assistant mode is when the topic touches regulated territory. For financial promotions, employment decisions, or contractual negotiations with significant consequences, both the ICO and the FCA expect genuine human judgement rather than algorithmic recommendation. AI can draft options and summarise information in these contexts, but it is not a substitute for qualified advice.
The second situation is data sensitivity. NCSC guidance is direct on this: treat prompts to public AI models as leaving your network. If stress-testing a decision requires sharing confidential client data, commercially sensitive pricing, or unannounced product plans, the risk of data leakage outweighs the analytical benefit unless you have contractual protections and data-processing terms confirmed.
The third is where you are early in your AI literacy. The Competition and Markets Authority has flagged evidence that users tend to follow AI recommendations even when those recommendations conflict with their own judgement, particularly under time pressure. If you have not yet built the habit of interrogating and verifying model outputs, lower-risk uses are the right starting point.
What does it cost to get this call wrong?
The costs of misusing AI in decision-making sit in two categories. Strategic mis-use shows up as decisions shaped more by plausible-sounding text than by your actual situation. Regulatory mis-use, specifically sharing personal or sensitive data with a public AI service without proper controls, can breach UK GDPR and carry fines of up to £17.5 million or 4% of global annual turnover.
The strategic cost tends to show up slowly. A founder who uses AI to confirm a pricing decision rather than challenge it may not notice the problem until the market responds. The Alan Turing Institute has documented a recurring pattern of “automation bias,” where decision-makers defer to algorithmic advice, especially when the AI presents outputs confidently and the decision carries time pressure. A mis-timed hire, a mis-priced service, or a mis-targeted campaign can each cost an SME tens of thousands of pounds in lost margin over a year without anyone tracing the root cause back to an AI interaction.
The data-compliance cost is more immediate. ICO enforcement fines can reach £17.5 million or 4% of global annual turnover for serious UK GDPR infringements. Remediation, legal advice, security reviews, and client communications often cost more than the original tooling saved. NCSC is explicit that feeding sensitive information into public AI tools without appropriate contractual and technical safeguards is a supply-chain risk.
The softer cost is decision paralysis. Generating large volumes of AI-produced analysis without a clear verification and decision process creates noise rather than clarity. For a small firm, delayed decisions can mean missed seasonal demand or a slower response to competitors, and neither shows up in any AI session log.
What should you ask before you decide?
Before committing to how you will use AI on a significant decision, four questions will quickly tell you whether you are in challenger or assistant territory. They take five minutes and they prevent the kind of early commitment to a direction that becomes harder to question once you have already received a confident-sounding answer and started acting on it.
The first question is what happens if this decision is wrong. If the answer involves regulatory breach, serious harm to clients, or financial existential risk, treat AI as a background assistant and get qualified human input. The Government’s AI Opportunities Action Plan is explicit that AI tools deliver real value for well-defined, bounded use cases, and that vague objectives and poor problem framing are major causes of AI project failure.
The second is how clearly you can state the decision in a single sentence. “Should we raise our minimum project size from £5,000 to £10,000 this quarter?” is a question AI can help you challenge. “What should we do about pricing?” does not give the model enough to work with. The sharper the question, the more productive the challenge.
The third is what data the model would need, and whether it is lawful and safe to share. If the decision involves personal data, identify your lawful basis before anything goes near a public AI service. The ICO’s guidance on automated decision-making is clear that using AI to inform decisions about individuals triggers specific rights and safeguards under UK GDPR.
The fourth is how you will verify the output before acting on it. Set the standard before you read the answer, not after. Identify what you will check manually: sample calculations, sector data from a trade body, or a sense-check with your accountant. NCSC guidance is clear that model outputs should be verified before being relied on for any decision that carries material consequences. Deciding on the verification method in advance is what separates structured challenge from sophisticated confirmation bias.
