Most of the AI coverage in business media is either too abstract ("AI will transform your operations") or too narrow ("here's how to use ChatGPT for your social media"). Neither is useful if you're trying to identify what's actually worth building right now.
This is the practical version: five specific AI tasks that have proven value in service business operations, with enough detail to know whether they apply to your situation and how to start.
1. Enquiry classification
What it does: Reads an incoming message or email and categorises it — booking request, general enquiry, complaint, existing customer question, spam.
Why it matters: Most inboxes mix everything together. A human has to read every message to decide what it is before they can decide what to do with it. Classification does that automatically, so the routing step (who handles it, and how urgently) can happen without manual triage.
How to start: You don't need a fine-tuned model. A well-written prompt sent to a standard model (GPT-4o, Claude) with a few examples of each category produces reliable classification at scale. Start with three categories: "booking request", "existing customer", "other". Expand from there once the three-way split is working.
Realistic accuracy: Expect 85–90% accuracy with a good prompt. Build a human review step for the "other" category and anything flagged as low-confidence.
2. Enquiry summarisation
What it does: Reads a long, unstructured customer message and extracts the key fields: what they want, when they want it, any constraints mentioned.
Why it matters: Customers don't write structured intake forms. They write messages. "Hi, I'm looking for someone to sort out my garden, it's quite overgrown, maybe 3-4 hours work, I'm in Brixton, hoping to get it done before my daughter's birthday which is the 15th" contains all the relevant information but requires reading and interpretation to extract it.
How to start: Write a prompt that asks the model to extract specific fields from the message text. Pass the output to your CRM or booking system as structured data rather than free text. Start with the fields your team manually copies today.
3. Draft reply generation
What it does: Given an incoming enquiry and some context about the business (services, pricing, availability), generates a draft reply for a staff member to review, edit, and send.
Why it matters: The time cost of customer replies isn't reading the message — it's composing a response that's accurate, professional, and specific to the enquiry. A draft reduces that to editing, which is significantly faster.
Where it works well: Standard enquiries with predictable structure — pricing questions, availability checks, service descriptions. Where it works less well: complaints, edge cases, relationship-sensitive communications.
How to start: Keep the human in the loop. Generate drafts; don't send automatically. The goal is to halve the time to compose a response, not to remove the human entirely. Most teams see 40–60% time reduction on standard replies once they've tuned the prompt.
4. Lead triage
What it does: Scores an incoming enquiry against your ideal customer profile and flags it as high, medium, or low priority before it reaches a salesperson or account manager.
Why it matters: Not all leads are equal, and most teams don't have a consistent way to decide which ones to prioritise. A triage system means the best leads get a faster, better response — which is where your conversion rate improvement actually comes from.
How to start: Define your ideal customer profile clearly: what size business, what type of problem, what budget signals, what urgency. Write a prompt that evaluates an enquiry against each criterion. Start with binary (worth pursuing / not worth pursuing) and add nuance once you trust the output.
5. Document Q&A
What it does: Allows staff (or customers) to ask questions in plain language against a set of documents — policies, product specs, pricing guides, FAQs — and get accurate answers without searching manually.
Why it matters: Knowledge is often trapped in documents that are too long to read quickly. A customer asks "what's included in your standard package?" and the honest answer is "look at page 4 of the service agreement." With document Q&A, the system reads page 4 and answers directly.
Where it works well: Well-structured documents with clear factual content. Where it works less well: documents with ambiguous language, heavy legal interpretation required, or frequently changing content.
How to start: Pick one document — your most-referenced FAQ or pricing guide. Build a simple retrieval-augmented generation (RAG) system that chunks the document, embeds the chunks, and retrieves relevant chunks at query time. Most cloud AI providers now have managed RAG services that reduce the build time to a day or two.
The pattern across all five
None of these are about replacing human judgment. They're about removing the mechanical steps that precede it — reading, categorising, extracting, drafting, searching — so that the human time that remains is spent on the work only humans can do.
The return on each is different, but the principle is the same: identify where your team is doing repetitive cognitive work, and ask whether a well-designed AI step could do most of it, most of the time, well enough.