A business owner came to me recently after pricing a new service based on a market analysis her chatbot had produced. The output was confident, well-organised, and wrong in two material places. The citations looked credible. Neither existed. She had not thought to verify them because the text surrounding them sounded authoritative. The pricing decision cost her three months of rework. She had the right tool for a different job.
What choice are you actually facing?
For many owner-managed businesses, the real question is how to match tool type to task type. Three categories cover the main jobs: general-purpose chatbots for framing problems, evidence-first engines for grounding factual claims, and structured environments for making sense of your own qualitative data. The UK government’s 2024 business survey found that over 70% of AI-adopting businesses already use generative AI for research or content tasks. The category choice is where the real decision sits.
A 2023 McKinsey survey found that 79% of businesses using generative AI had deployed it in at least one function, with strategy and R&D among the common applications. That is a large and growing number. What the survey does not tell you is whether those businesses were using the right category of tool for each type of task. The mismatch is common, and it tends not to announce itself. Fluent, well-organised AI output looks the same whether the underlying facts are accurate or not.
When does a general-purpose chatbot earn its place?
A general-purpose chatbot earns its place when you are still working out what you need to know. These tools are built for following complex instructions and generating well-structured text from them. They handle brainstorming, decision-framing, and turning your own verified notes into a structured memo. The key point is “verified notes”: chatbots produce fluent text, but they do not reliably verify the facts they draw on.
Use a chatbot when the task is about structure, not source-finding. Explaining how a regulation works in plain terms, setting out the options in a pricing decision, or drafting the structure of a board memo from points you already know well, these are all good fits. The chatbot is doing the organising.
Where chatbots fall short is when the facts need to be accurate and easily verified. Their default behaviour is to produce confident text. That text may contain errors that look identical to correct information, including fabricated citations that appear real. In regulated sectors, or when a client or adviser will scrutinise your work, output without verifiable sources is a liability before it is an asset.
The practical boundary is worth keeping in mind. If you would be comfortable saying “I worked this out myself; the AI helped me structure it”, the chatbot is probably the right tool. If you would need to say “the AI researched this”, reach for something with visible sources.
When does an evidence-first or structured tool matter more?
An evidence-first engine starts from sources rather than generating text freely. Tools like Perplexity, Elicit, and Consensus pull from specified corpora and show where each claim originates. A 2023 Stanford study found that large language models frequently hallucinate citations, producing plausible-sounding but non-existent references. Evidence-first tools reduce that risk by grounding answers in documents the tool has actually retrieved, rather than inferred.
Use an evidence-first engine when you are asking factual questions where you will act on the answer. What are the main UK data retention requirements for HR records? What does the evidence say about the impact of flexible working on productivity? What are competitors in this category charging? These are research questions that need real sources, not fluent summaries of unverifiable claims.
A second category worth knowing is structured analysis environments. If the data you need to make sense of is yours rather than external, tools like NVivo or a well-configured Notion with AI features let you organise interview notes, client feedback, and staff surveys while preserving an audit trail. The output is traceable to a source document, which matters when a conclusion will inform a board decision or a regulatory response.
The practical starting point for an owner-managed business is two tools: one chatbot on a business plan with appropriate data controls, and one evidence-first engine for questions where accuracy matters. Add structured analysis tools when you have a repeating need to code and categorise your own documents.
What does it cost to get this wrong?
A tool mismatch tends to produce a confidently delivered wrong answer, which is harder to catch than an obviously uncertain one. A business owner who asks a chatbot for regulatory guidance and receives well-written but inaccurate output may act on it without realising it needs checking. The ICO has confirmed that organisations remain accountable for decisions made with AI assistance, even when the tool itself generated the error.
Three costs tend to compound when the mismatch goes unaddressed.
The first is wasted time. An AI-generated market analysis that later proves unreliable requires significant rework, often across multiple functions. The cost is not just the hours spent on the initial analysis; it is the downstream decisions made before anyone noticed the problem.
The second is data protection risk. The ICO’s guidance on generative AI is explicit: organisations must have a lawful basis for processing personal data in AI prompts and a data processing agreement with the tool provider. UK GDPR fines can reach £17.5 million or 4% of global turnover. UK law firm Fieldfisher has noted that careless use of public AI tools has already led to accidental disclosure of confidential material from several businesses.
The third applies specifically to regulated businesses. The FCA’s discussion paper on AI confirmed that using AI for research or decision support does not reduce a firm’s responsibilities under conduct rules, including the Consumer Duty. If an adviser uses a chatbot for product research and the output is inaccurate, the liability for unsuitable advice remains with the firm. The CMA has also noted that businesses relying on similar uncustomised models risk producing homogeneous analysis, which is a problem for competitive differentiation as much as accuracy.
What to ask before you commit to a tool?
Before settling on any AI for research and thinking tasks, three questions cut through much of the noise. Where is the data stored, and is there a data processing agreement that satisfies UK GDPR? Can you see the sources behind each answer, not just the answer itself? And who in the business reviews AI-generated research before it becomes a decision? Those three questions tell you more than any feature comparison.
On data storage: the country and cloud provider matter, as does whether your data is used to retrain the model or shared across other users. Business-tier subscriptions for tools like ChatGPT and Claude typically offer better data isolation than free tiers, along with proper data processing agreements. Using free-tier public AI tools with client information is not advisable under current ICO guidance.
On source visibility: if a tool will not show you what it drew on to reach an answer, you cannot verify that answer. Perplexity cites sources by default. General-purpose chatbots do not, unless prompted specifically. For board papers, risk registers, or any document that will support a decision, visible and clickable sources are a practical requirement, not a nice-to-have.
On review governance: AI-generated research requires a named person to check it before it becomes a decision. The Bank of England and FCA’s joint AI forum report noted that using AI for decision support still raises model risk and accountability concerns, even when the AI is not making fully automated decisions. A short written policy on what staff may and may not input into external AI tools closes most of the exposure without slowing anyone down.
If you want help working out which tools fit your business and what governance should sit around them, book a conversation.
