30 AI Prompts for SMEs: Email, Quotes, Admin and Customer Service + Free 30 SME AI Prompt Templates (Word)
A useful prompt is not a magic phrase. It is a clear work order: what must be done, which information is authoritative, what must not be assumed, the required output and who reviews it. “Make this more professional” is rarely enough if the model does not know the customer stage, approved product facts, pricing authority or next action.
The 30 templates below are designed for everyday SME work. Replace every bracketed field with real information. When evidence is missing, instruct the model to list the gap instead of inventing an answer. Prices, contracts, refunds, personal data and public commitments require review by an authorised person.
Six parts of a reusable prompt
Part | Question to answer | Example |
|---|---|---|
Role | Who owns the work? | You are an e-commerce manager |
Objective | What outcome is required? | Turn enquiries into an assignable list |
Inputs | What evidence is available? | Provide product, audience and channel data |
Constraints | What must not happen? | Do not invent prices or commitments |
Format | How should output be delivered? | Table: task, owner, deadline |
Review | How will it be checked? | List assumptions, gaps and human checks |
How to use the templates
Provide the necessary context first, then the relevant source material, and finally the output format. Use the first pass for a draft and the second pass to identify omissions, risk and uncertainty. Do not ask one prompt to research, price, approve and send. OpenAI's official guidance likewise emphasises clear, specific instructions, sufficient context and an explicit format.
Free resource: 30 SME AI Prompt Templates
Adapt the input fields for email, quotations, administration and customer service. Get an editable Word version to adapt and reuse for your team.
Enter your email, choose how you’ll use the kit and confirm your consent to receive it by email.
Email and daily communication｜01–05
01｜Reply to a first enquiry: You are a customer service colleague at [company/brand]. Using the customer's message “[paste original]” and approved service information “[paste source]”, write a reply of no more than 150 words. Confirm the need, ask no more than three essential questions and give one clear next step. Do not invent price, timing or functionality. Mark missing information as [TO CONFIRM]. Output: subject, body and internal review points.
02｜Acknowledge receipt: In a [professional/friendly/concise] brand voice, write a 60–90 word acknowledgement for [document/enquiry/complaint]. State which role will follow up without promising an unapproved response time. Output Chinese and English versions plus one internal note.
03｜Request missing information: Read “[customer message]” and “[required information list]”. Ask only questions that affect scope or quotation, with a maximum of five in priority order. Give an easy example or choice for each question and do not repeat information already supplied.
04｜Delay notice: Using [original date], [current situation], [confirmed new arrangement] and [items we cannot promise], write an accountable delay notice. Include what happened, customer impact, action taken and the next update point. Do not invent a cause.
05｜First complaint response: Organise “[complaint]” into facts, impact, request and risk, then draft a response that acknowledges impact before explanation. Promise only [authorised scope]. Mark refunds, legal issues, safety or privacy as [HUMAN ESCALATION REQUIRED].
Quotations and sales｜06–10
06｜Qualify an enquiry: Using [industry, objective, budget, timing, decision maker and existing assets], identify missing information required for a proposal. Output three columns: known, unknown and the single most useful next question. Do not silently turn unknowns into assumptions.
07｜Quotation summary: Turn “[quotation]” into a client-friendly summary with objective, scope, deliverables, timeline, client inputs, fees, exclusions and next step. Preserve every number exactly. Flag contradictions as [CHECK REQUIRED].
08｜Three options: Using [objective], [budget range] and [non-negotiable limits], design Basic, Recommended and Growth options. For each, state the fit, inclusions, exclusions, expected timing and main trade-off. Do not use fake discounts or fear-based claims.
09｜Respond to “too expensive”: Using “[client wording]” and approved value evidence “[evidence]”, create three replies: clarify budget, reduce scope and phase the project. Keep each under 100 words. Do not attack competitors or promise an unapproved discount.
10｜Post-meeting follow-up: Using “[meeting notes]”, draft an email with confirmed points, open points, actions, dates and risks. Put contradictions in an internal review section rather than choosing a version.
Administration and operations｜11–15
11｜Meeting summary: Turn “[transcript/notes]” into decisions, actions, responsible roles, deadlines and open questions. Record only explicit information. Use [NOT SPECIFIED] when an owner or date is missing.
12｜Task plan: Break “[project context]” and “[existing tasks]” into the smallest executable steps. Give each a completion definition, dependency, responsible role and risk. End with one unique next action.
13｜SOP draft: Convert “[current practice/interview notes]” into an SOP with purpose, trigger, inputs, steps, exceptions, approvals, record location and evidence of completion. Mark unverified steps [TO TEST].
14｜Weekly report: Using “[completed]”, “[data]”, “[blockers]” and “[next priorities]”, write a one-page management update. Separate activity from business outcome and end with only one decision required.
15｜File organisation: Using “[file list]” and “[folder, naming and retention rules]”, propose a move and rename plan. Do not delete files. Identify duplicates, recommend which to keep and list unresolved items.
Marketing content｜16–20
16｜Social post: Using [audience], [offer], [evidence], [voice] and [CTA], create informative, proof-led and action-led versions. Each includes an opening, body, CTA and visual suggestion. Do not invent customer results.
17｜Thirty-day content plan: Using [brand position], [three customer questions], [available assets] and [main channels], propose 30 non-duplicated topics across education, proof, product, behind-the-scenes and interaction. State the search or decision intent for each.
18｜Ad copy variants: Using [audience], [pain], [benefit], [evidence] and [limits], produce 12 headline/body pairs. Change one main hypothesis per pair and name the test variable. Do not claim unsupported performance.
19｜Repurpose long content: Turn “[article/video script]” into a LinkedIn post, five-card carousel, 30-second video script and email. Preserve facts and numbers, identify statements that need source links and add no new case study.
20｜FAQ: Read “[sales conversations/support logs]” and identify ten repeated questions across pre-purchase, active use and after-sales. Answer only from [approved knowledge base], and send evidence gaps to the internal team.
Customer service and knowledge｜21–25
21｜Classify an enquiry: Classify “[message]” by service, location, urgency, sentiment, next action and human-escalation need. Use only [approved categories]. Return [UNCERTAIN] with a reason when confidence is insufficient.
22｜Knowledge-base answer: Answer “[question]” only from “[knowledge-base extract]”. Give the direct answer, steps and limitations. If the source has no answer, identify the gap and the role to receive the handoff. Do not fill company policy with general knowledge.
23｜Human handoff: Read “[conversation]” and write a handoff summary covering customer objective, verified identity data, attempts made, unresolved issue, risk, sentiment and suggested next sentence. Remove unnecessary personal data.
24｜Tone correction: Rewrite “[reply]” in [brand voice] while preserving every fact, price, date, limitation and legal phrase. Output the rewrite and a change log. Friendliness must not weaken a required term.
25｜Feedback synthesis: Group “[anonymised feedback]” by theme, frequency, severity, representative sentence, improvement and unknowns. Do not call one comment a trend. Remove personal information first.
Management and decisions｜26–30
26｜Decision memo: Using [options], [objective], [constraints], [cost] and [evidence], write a one-page memo with recommendation, rationale, trade-offs, risks, validation needs and reversal conditions. Separate fact, assumption and inference.
27｜Pre-mortem: Assume “[plan]” has failed in three months. Generate ten plausible reasons across people, process, data, technology, customer and vendor. For each, give an early signal, prevention and responsible role.
28｜Vendor questions: Using [procurement purpose] and [data/systems involved], create 12 questions covering evidence, data handling, permissions, exit, support, total cost and acceptance. Add acceptable evidence to each question.
29｜Training plan: Create a two-week AI work programme for [role], limited to [three workflows]. Each lesson includes demonstration, practice, common errors, review criteria and a real work output. Course completion is not proof of competence.
30｜Monthly review: Using [usage logs], [errors], [time saved], [quality checks] and [feedback], compare this month with last month. Separate activity, quality, cost and business outcome. Recommend no more than three improvements, each with one measurable next validation.
Turn prompts into a company asset
Start with ten recurring workflows. Assign an owner, required inputs, approved sources, prohibited commitments, review role and version date. Every failure should lead to a prompt, data or process correction—not another reminder to “ask the AI better”.
For planning, a first Hong Kong SME prompt library may require HK$4,000–12,000 for ten workflow templates. Adding a knowledge base, forms, permissions and systems integration may start around HK$15,000–60,000. Allow one to four weeks depending on source quality, approval levels and real-system testing. These are Allua planning ranges, not software subscriptions or fixed quotations.
Frequently asked questions
Can the whole company receive all 30 prompts at once?
Yes, but smaller role-based sets work better. Customer service, sales and management need different data and approval lines. A huge library encourages people to return to improvised prompts.
Does fluent output mean correct output?
No. Fluency is presentation. Price, policy, dates, contracts, personal data and customer commitments still require authoritative verification.
How should we measure prompt improvement?
Use the same de-identified cases and compare omissions, edit count, handling time and quality-check pass rate. A before/after comparison is not a randomised A/B test; staff learning and workload may differ.
Want governed prompts for real company workflows?
Share your ten most frequent tasks, current forms or SOPs, brand voice, approval rules and failure examples with Allua. We can convert them into a prompt library, review checklist and pilot plan.
Information checked on 2 October 2026. AI output can be wrong. Financial, legal, medical, personal-data, payment and irreversible actions require qualified human review and access controls.
Comments