How to write prompts that actually work for business tasks

To write prompts that work for business tasks, give the AI the same things you would give a capable new starter: context about your business and the reader, a clearly defined task, the format you want back, the constraints it must respect, and one or two examples of work you already consider good. Then treat the first output as a draft and iterate in the same conversation rather than starting again from scratch. That five-part structure, context, task, format, constraints and examples, fixes most of the vague, generic output that puts businesses off AI.

Why do most business prompts fail?

UK businesses have adopted AI faster than almost any technology before it. The British Chambers of Commerce found that 54% of SMEs are now actively adopting AI, up from 35% a year earlier and 25% the year before that. The same research notes that most firms are using generic tools such as chatbots and writing assistants to draft content, summarise information and conduct research.

So plenty of businesses are typing prompts. Far fewer are getting consistently usable results, and the difference is almost never the tool. It is the brief.

"Write a marketing email for my business" gives the model nothing to work with. It does not know who your customers are, what you sell, what tone your brand uses, or what you want the reader to do. So it falls back on the average of every marketing email it has ever seen, which is exactly why so much AI output reads like everyone else's.

The most common failure modes are easy to spot once you know them:

• No context. The model does not know your business, your customer or the situation.
• No defined task. "Help me with this proposal" is a topic, not a job.
• No format. You wanted a tight paragraph; you got 600 words of bullet points.
• No constraints. Nothing told it to avoid jargon, stay under a word count, or skip the hard sell.
• No example. It has never seen what good looks like for you, so it guesses.
• Vague verbs. "Improve", "polish" and "make it better" mean nothing without a definition of better.

This briefing gap also explains why so many firms struggle to find value at all. When the Office for National Statistics asked UK firms about barriers to AI, the most common answer was difficulty identifying activities or business use cases, cited by 39% of firms, ahead of cost at 21%. A few disappointing generic outputs early on, and the conclusion becomes "AI does not work for us" when the real problem was a two-line prompt.

"A prompt is a brief, not a wish. If you would not hand the same instruction to a new starter and expect a usable result, do not hand it to an AI."

Dean Cookson, founder, Operosus

What should every business prompt include?

Five ingredients cover almost every business writing task. You will not need all five every time, but when an output disappoints, the fix is nearly always one of these.

How much context is enough?

Two or three sentences usually do it: who you are, who the reader is, and what outcome you want. More context helps up to a point, but paste with care. Client names, financial details and anything personal should be anonymised before they go into a general-purpose chatbot, and if you are handling customer data at any scale, check the tool's data terms first. The ICO's guidance on AI and data protection is the primary reference for UK businesses here.

A practical habit: write a short "about us" paragraph once, covering what you do, who you serve and how you sound, and paste it at the top of any prompt that produces customer-facing words. It takes ten minutes to write and improves everything after it.

What does a good prompt look like in practice?

Take invoice chasing, a task most SMBs do badly because it is awkward (we have a full guide to automating invoice chasing if it is your bottleneck). Here is the weak version most people type:

Write an email chasing an unpaid invoice.

And here is the same request as a proper brief:

We are a small web design studio. The reader is a marketing manager at a client we like and want to keep. Invoice 1042 for £3,400 was due 21 days ago; we sent one gentle nudge a week ago and heard nothing.

Draft a follow-up email. Under 100 words. Warm but unambiguous: we need a payment date. Reference the previous nudge without sounding wounded. British English, no exclamation marks, sign off as Dan. Do not offer a discount and do not mention late fees yet.

The second prompt takes ninety seconds longer to write and produces something you can send with light edits instead of a rewrite. That is the entire economics of prompt writing: a minute of briefing buys back ten minutes of editing, every single time the task recurs.

Why do examples beat instructions?

Adjectives are the weakest tool in a prompt. "Professional but friendly" means something different to every business, so the model averages across all of them. A single real example pins it down instantly: paste one email, one tender answer or one product description you are proud of, say "match this tone and structure", and the output jumps in quality more than any list of style instructions can manage.

We see this constantly in our own builds. Bidwell, our AI tender-writing product, leans on a business's past bids and real evidence rather than asking the model to "write persuasively", because a model shown a genuine winning answer produces something in the bidder's voice, while a model given adjectives produces boilerplate that evaluators have read a hundred times. The same pattern holds at every scale, from a sole trader drafting quotes to a firm answering 40-page frameworks.

If you have no example to hand, create one: edit the AI's first attempt until it sounds like you, then feed that edited version back in as the reference for next time. You only have to do this once per task type.

How do you iterate without starting again?

The first output is a draft, and the conversation is the edit. Most people either accept a mediocre first answer or delete everything and rewrite the prompt from zero. Both waste the model's biggest advantage: it remembers the conversation and responds well to targeted feedback.

Iterate the way a good editor gives notes:

1. Be specific about what stays. "Keep the structure and the second paragraph."
2. Be specific about what changes. "Cut the first sentence, it restates the subject line."
3. Name the problem, not just the symptom. "This sounds like it is apologising for existing. Be more direct in the ask."
4. Change one thing at a time when output quality is close, so you can tell what worked.
5. Save the winning prompt. When a version lands, copy the full prompt that produced it into a shared document. That document becomes your prompt library, and it is worth more than any course.

One warning sign worth knowing: if you have iterated five or six times and the output is still drifting, the conversation has usually accumulated contradictory instructions. Start a fresh chat and paste in your refined prompt plus the best draft so far. You lose nothing and the model stops trying to satisfy six rounds of conflicting notes.

When does a prompt stop being enough?

A prompt you type weekly should be a saved template. A template your whole team uses, fed with the same customer data every time, should be a system. The BCC research found that only around one in ten firms have adopted bespoke AI systems integrated into core functions like workflow automation and customer service, which means most businesses are still hand-typing briefs for work that has long since become routine.

The threshold is repetition plus data. Our appointment-booking work for a national at-home vet service is a useful illustration of the shape: the messages that confirm bookings and keep customers informed are not typed into a chatbot by a person each time. The context a human would have pasted in, who the customer is, what they booked, where the enquiry came from, flows in automatically from the booking system, and the communication goes out with no one writing a prompt at all. The same principle drove our facility-hire work for schools: once listing descriptions follow a proven pattern, generating them belongs in a pipeline, not a chat window.

You do not need to start there. You need to start with prompts, because the prompts you refine by hand are the specification for the system you eventually build. Every well-briefed, well-iterated prompt in your library is a documented, tested piece of how your business communicates.

Where to start

Pick one recurring writing task this week, ideally one someone mildly dreads: invoice reminders, enquiry replies, quote follow-ups, meeting summaries.

1. Write the five-part brief: context, task, format, constraints, and one real example of the task done well.
2. Run it, then iterate with specific editor's notes rather than "make it better".
3. When the output is sendable with light edits, save the full prompt in a shared document with a clear name.
4. Repeat for the next task. After four or five, you have a prompt library that makes every new hire and every busy Monday faster.
5. When a saved prompt is being used weekly with data copied out of another system, that is your signal it should become an automated workflow.

One prerequisite worth a few minutes: if you are still deciding which assistant to write these prompts in, our comparison of ChatGPT, Claude and Gemini for small business settles it faster than trial and error.

The first four steps cost nothing but attention. The fifth is where we come in: Operosus designs and builds AI systems for UK businesses that have proven the value by hand and want it running without the typing. If you have a prompt library and a bottleneck, talk to us.