When I work with small services firms that have tried ChatGPT and given up on it, the first thing I ask to see is the prompts. Almost always, they’re single sentences. “Write something about our new flexible working policy.” “Draft a client update about the delay.” The tool delivers exactly what those instructions deserve, which is text that is plausible, generic, and needs rewriting before it can be used.
When the same task gets a properly structured instruction, one that names the audience, the format, the key points, and the tone, the second draft goes out with two or three edits. The change is entirely in how the request was framed.
That gap, between a vague ask and a structured one, is where much of the accessible value in ChatGPT sits for a 5 to 50 person services firm. The habits below close it.
What makes a prompt actually work?
A prompt works when it gives ChatGPT enough structure to narrow the task. The core elements are: a role (what expertise to adopt), an audience (who is reading the output), a format (email, policy, bullet list), a length, and any constraints such as “UK English only.” Anthropic’s 2024 research found that prompts built around those elements reduced factual errors by around 28% compared with vague instructions.
OpenAI’s guidance for business users recommends the same approach. Specify what expertise the model should adopt, who it is writing for, what format the output should take, the approximate length, and what it should avoid. A prompt for an internal email might read: “You are an HR adviser at a UK professional services firm. Write a 300-word staff email announcing our new flexible working policy. Key points: [three bullets]. Tone: positive but direct. UK English throughout.”
The discipline involved is the same as writing a good brief for a junior colleague. Naming the audience, the objective, the format, and the constraints produces a better prompt for the same reason it produces a clearer written brief. Firms that send ChatGPT one-sentence instructions are giving it less direction than they would give a work-experience student, then wondering why the output needs such heavy editing.
Why does prompt quality matter for a 5 to 50 person firm?
McKinsey’s 2023 research estimated that generative AI could automate 60 to 70 per cent of the time employees spend on tasks such as drafting emails, summarising documents, and basic analysis. That figure assumes the tool is being used well. A vague prompt to a premium model produces output that needs as much editing as a strong prompt to a free one.
Deloitte’s 2023 survey found that around 29 per cent of UK SMEs were already experimenting with generative AI, with professional services firms leading adoption. Data protection and output quality were the two main barriers identified. Both of those concerns respond directly to better prompting discipline. Higher-quality output means less time correcting drafts, and prompts that keep personal data out of the conversation address the data protection concern before it becomes an issue.
For a small firm, the arithmetic is straightforward. If three people each spend 45 minutes a week rewriting AI drafts that came back too generic, the team is losing over two hours a week to poor prompting habits. A shared template and a simple iteration rule reclaims a good portion of that, without any new software or training budget.
Where does prompt quality show up in a firm’s day-to-day work?
The time cost of weak prompts shows up most clearly in three types of work: drafting client-facing documents (proposals, updates, reports), producing internal materials that need to reflect the firm’s specific context (policies, onboarding guides, briefing notes), and asking factual questions that require UK-specific guidance. In all three, a vague prompt returns something that looks usable but needs substantial reworking before it can go anywhere near a client.
For client-facing work, the fix is to include the client’s sector, the purpose of the document, and any constraints on tone or terminology in the prompt. “Write a project update email for a law firm client concerned about timeline” will produce something more useful than “write a project update email.”
For internal documents, adding the firm’s context removes the generic quality that makes AI-drafted policies feel borrowed from a downloaded template. Specifying that the firm is a 20-person UK accountancy practice, the audience is all staff, and the subject is AI tool use in client work gives the model enough to work with.
For factual questions, the UK Department for Education’s 2023 guidance on generative AI makes the point plainly: use AI as a drafting assistant and ensure that humans review and refine output to fit local context. The NCSC’s guidance on large language models adds that AI-generated advice can sound plausible and still be wrong, particularly on technical or regulated questions. Specify the UK context in the prompt and build a verification step in afterwards.
When should you build a shared template, and when can you just type?
A prompt template earns its place when the same type of task comes up more than once a week and involves more than one person on the team. The time investment in writing a shared template is small. The return, fewer rewrites across everyone who uses it, compounds quickly. For one-off or exploratory tasks, a thoughtful typed prompt works fine. The template is for the repeatable work.
A basic template for a UK services firm might read: “You are a [role] advising a UK [sector] firm of [size]. The audience is [clients / staff / a regulator]. Write a [format] of around [length]. Tone: [tone]. Constraints: UK English, reference UK law and official guidance only, no US references unless asked.”
That can sit in a shared note and be pasted into the first line of any prompt. Staff who use it regularly adapt it quickly to their own situations. The habit of including the elements reliably matters more than any particular wording. Firms that make this the default tend to see the quality of AI output improve across the team within days.
Two situations where a template does not help: genuinely exploratory tasks where the question is still forming, and requests so specific to one client or situation that pre-written framing adds more friction than it removes. For those, a well-thought-out typed prompt and a round of refinement is the better approach.
What else do better prompts need to account for?
Three things consistently undermine the value of better prompts once a small firm moves past individual use: treating the first answer as the final answer, using AI-generated factual claims without checking them, and putting personal or client data into a public tool. A structured initial prompt addresses the quality of the first response. These three habits address what happens next.
Iterate rather than re-prompt
OpenAI’s guidance emphasises that ChatGPT is built for multi-turn conversations. Ask for the first draft, then ask for one or two refinements: “Make this shorter and more conversational” or “Add a paragraph addressing the concern about costs.” That loop is almost always faster than trying to write the perfect initial prompt. The DfE’s 2023 generative AI guidance makes the same point: treat AI as a drafting assistant, review the output, and refine it before use.
Verify before you act on it
For anything involving law, regulatory requirements, tax, or statistics, ask ChatGPT to list the UK sources it relied on, then check those sources directly. The ICO notes that generative AI systems may produce inaccurate or invented information and says organisations must address these accuracy risks under UK GDPR. The FCA has confirmed that regulated firms remain responsible for the accuracy of their communications even when AI assisted in drafting them.
Keep personal data out of public tools
The ICO’s guidance on generative AI under UK GDPR is explicit: organisations must minimise personal data shared with third-party AI systems and have a lawful basis for doing so. The NCSC adds that prompts to online AI tools may be logged and visible to the service provider. The practical rule: use “Client A” and “£X turnover” rather than real names and figures. Write this into any AI guidance you give staff before they start using these tools in client work.
