You’ve been sitting on the same idea for two weeks. You’ve sketched it out, talked it through with someone you trust, and still can’t name what’s bothering you. The idea might be genuinely good. Or there might be a flaw you haven’t found yet.
Many founders in that situation either push ahead and hope, or put the idea down until the moment passes. A third option has become practical in the last two years: using a generative AI tool as a structured thinking partner. The aim is to get better questions, not to outsource the decision.
What does using AI as a thought partner actually mean?
Using AI as a thought partner means prompting it to surface your assumptions, argue against your idea, and generate cheap tests, not to hand it the decision. A 2023 Boston Consulting Group study of 758 consultants found that access to GPT-4 raised creative problem-solving performance by 40%, but damaged analytical accuracy when people deferred to the model’s reasoning rather than checking it themselves. The value is in the challenge, not the answer.
The practical version of this looks like a structured prompt rather than a casual question. The UK Government’s AI Playbook for public bodies advises teams to specify their user needs, context, and constraints when using generative AI, and to test outputs rather than accept them. For a founder testing a business idea, that translates into something like this:
“You are acting as a critical friend to a UK professional services firm with twelve employees. I’ll describe a service idea. Your tasks: identify at least five implicit assumptions I’m making, list concrete ways each assumption could be wrong in practice, and suggest one low-cost test for each.”
That prompt works because it forces the AI to work against your idea, which is exactly where the value sits. You can also spot immediately when the challenge is off-target, because you know your own business and market better than the model does.
Why does this matter for how your business runs?
For an owner-managed firm, decision quality matters more than decision speed. The UK’s Department for Science, Innovation and Technology reported in 2024 that 68% of UK businesses using AI apply it for data management and analysis, and 52% for predictive analytics, compared to 39% using it for chatbots. AI is already showing up as a thinking tool first, and the founders who learn this early get better options in front of them faster.
The business case sharpens when you’re facing a decision with long consequences and limited data. Pricing changes, new service lines, key hires, whether to pursue a particular contract: these are the decisions where confirmation bias bites hardest. A 2024 McKinsey survey found that organisations using AI in forecasting and budgeting reported a three to five percentage-point improvement in forecast accuracy on average, with the leading firms seeing gains of up to fifteen points. That accuracy gap closes in part because AI-assisted analysis is harder to unconsciously slant in a preferred direction.
For a founder working through a difficult decision alone on a Tuesday afternoon, that is not a small thing.
Where will you actually use this in your working week?
The useful moments are the decisions you’d normally wrestle with alone: a pricing change, a hire you’re unsure about, a new service offering that keeps stalling. Pull up ChatGPT or Copilot, describe the situation in a paragraph, and ask it to list five assumptions you’re making and how each could be wrong. You get in two minutes what might take a full coaching session to surface.
Three prompt patterns work well in practice. The assumption audit asks the AI to list five assumptions behind your idea, explain how each could be wrong, and suggest one thing you could do in a week to test it. This is a direct application of what the UK Government’s AI Playbook calls testing with real users and measuring performance, applied to your own thinking rather than to a product.
The steel-man and straw-man prompt is useful when you’re already attached to an idea. You ask the AI to make the strongest possible case for your proposal, then the strongest possible case against it, then finish with a list of the key uncertainties that decide which side is right. The structure surfaces specific risks rather than a generic list of concerns, which is where confirmation bias typically lives.
The scenario test takes a decision you’re close to making and asks: what would have to be true for this to go wrong? Adding a UK market or sector frame to the question tends to improve the quality of the challenge significantly.
When should you push back on what AI tells you?
The 2023 BCG study of 758 professional consultants is the clearest warning: performance on complex analytical tasks dropped when consultants trusted the model’s reasoning rather than their own. On decisions that depend on sector knowledge, client relationships, or regulatory compliance, your judgement outperforms anything the model can generate. Treat every AI output as a hypothesis you bring to your own assessment, not a verdict to act on.
The situations where AI-generated challenge is least reliable include anything touching FCA-regulated advice, because the model’s grasp of current regulatory expectations is imperfect and you carry professional responsibility. Legal matters fall into the same category. And any claim the AI makes that cannot be sourced. The Competition and Markets Authority has warned that foundation models can confidently produce incorrect information. The Levidow, Levidow and Oberman case in the United States, where lawyers submitted court filings containing fabricated AI-generated case citations, became a widely-cited caution in UK legal commentary for exactly this reason.
The check is straightforward: if you’re going to act on something the AI surfaced, find the primary source. If the source doesn’t exist, the claim doesn’t count.
What should you have in place before you start?
Two things catch founders out when they start using AI for sensitive decisions. The first is putting client data, pricing strategies, or HR matters into a consumer account where the provider’s terms allow training on your prompts. The second is not recording what the AI surfaced and what you decided. The UK’s National Cyber Security Centre advises classifying data before it goes near any AI tool.
For a firm of five to fifty people, the practical setup is accessible. Microsoft 365 Copilot, priced at £27.30 per user per month in the UK, processes prompts within your Microsoft tenant under commercial data protection terms, which means your strategy conversations stay separate from the model’s training data. Google’s Gemini for Workspace starts from around £16 per user per month and offers similar separation. Either option gives you a credible enterprise-grade thinking aid within a normal software budget.
The second piece is a lightweight decision log: the question you brought to the AI, the assumptions it surfaced, what you decided, and when you’ll revisit it. The Information Commissioner’s Office expects firms to be able to explain AI-influenced decisions that affect people, whether employees, clients, or suppliers. A short log entry after each substantive session covers that and makes it easier to revisit a decision if the context changes.
If you want to go further, the UK Government’s AI Playbook provides a framework for structured human oversight, and the ICO’s guidance on AI and data protection sets out the specific obligations that apply when personal data is involved.
