Each department has two or three jobs AI already does well. Sales can research prospects and draft proposals, marketing can repurpose content, client service can sort emails and draft replies, finance can code transactions and chase payments, operations can turn know-how into procedures, and HR can draft adverts and onboarding plans. Start where repetitive writing eats the most hours.
In a small firm a "department" may be one person, or half of one. That doesn't matter; the tasks are the same. Each use case below says what it is, how it looks in an illustrative 20-person insurance brokerage, how to start and what to watch. There's a table at the end showing where to begin in each area and what to measure.
Sales and new business
Prospect research before a meeting. A chat assistant with web search can turn a company name into a one-page brief: what the business does, recent news, how it's structured and questions worth asking. In the brokerage, an account executive meeting a haulage firm about fleet cover gets a brief in five minutes covering fleet size clues, recent contracts and the questions to ask about drivers and routes. Start with a saved prompt that asks for sources beside each claim. Watch: web results can be out of date, so check anything you'll repeat to the prospect.
Two lines from an illustrative brief show why the sources matter: "Operates around 40 vehicles (source: the company's fleet page)" and "Recently won a large distribution contract (source: trade news article)". Clicking through, the fleet page belongs to a sister company with a similar name, and the "recent" contract was announced three years ago. Neither is a hallucination as such; both are real pages, misread. Opening a meeting with either would have told the prospect the broker hadn't done their homework, which is the opposite of what the brief was for. Two clicks per claim is the whole check.
Recommendation letters from quote comparisons. The broker has done the work: three quotes compared in a spreadsheet and notes on the client's needs. AI turns that into a clear letter explaining the options and the recommendation in plain English. Paste the figures in rather than letting the AI supply them, and check every number against the spreadsheet. The recommendation itself stays the broker's.
A paragraph from an illustrative draft: "Option B costs $1,120 a year more than your current policy, but it raises the goods-in-transit limit from $50,000 to $150,000, which matters now that you carry electronics loads twice a week." That's the useful kind of sentence, built entirely from the broker's notes. The one to delete is the line AI likes to add at the end, such as "This policy will give your business complete protection", because no broker should put that promise in writing.
Follow-up that actually happens. A common leak in small firms is the quote nobody chased. AI inside the CRM can flag quotes with no activity after a set number of days and draft the follow-up. HubSpot's Prospecting Agent, for example, is available from Starter upwards on most hubs (HubSpot's page of 16 September 2026) and uses 100 HubSpot Credits, about $1, per lead it recommends for outreach. Watch: automated outreach at volume damages your reputation faster than it wins business; keep a person approving each message.
In the brokerage, chasing existing quotes looks like a saved Monday view in the CRM: 23 quotes issued last month, 9 with no contact for more than ten days. A drafted chaser for each, referring to the actual quote ("the fleet quote from the 4th, which included the two new vans"), takes an account executive about 15 minutes to approve instead of a morning to write. That part needs only the saved view and a drafting prompt. The Prospecting Agent is for finding new leads and is available from Starter upwards on most hubs, on paid credits; at 100 credits each, Professional's 3,000 monthly credits would cover about 30 recommendations if nothing else drew on them, so treat it as a targeted extra rather than a pipeline.
Marketing
One piece, many formats. Write one solid guide a quarter, say "What to check before your commercial property renewal", and let AI turn it into a newsletter article, three social posts, a short FAQ for the website and a script for a one-minute video. That's where AI saves the most marketing time in a small firm: the reformatting, not the original thinking.
Here is a guide we wrote for clients (pasted below). Turn it into:
1. A 250-word newsletter article with one clear call to action
2. Three LinkedIn posts, each built around one point from the guide
3. Five FAQ entries (question plus a 60-word answer) for our website
Rules: use only facts that appear in the guide. Don't promise cover,
prices or outcomes. Plain English, no exclamation marks.
Mark anything you're unsure of with [CHECK].One of the LinkedIn posts might come back like this (illustrative):
Renewing your commercial property cover next month? Check one thing first: is your rebuild value still right? Building costs have moved a lot, and being underinsured can cut a claim payout [CHECK: does the guide explain how a shortfall reduces the payout?]. Our guide covers the three numbers to confirm before you sign.
The [CHECK] tag did its job: the guide mentions underinsurance but never explains how a shortfall is worked out, so the broker either adds a line to the guide or cuts the claim. "Building costs have moved a lot" needs the same treatment. It isn't in the guide, so it's exactly the kind of borrowed fact the prompt ruled out, and it slipped through without a tag.
The monthly newsletter from the month's notes. Collect what happened (a claims tip, a market change, a team hire) in one running document, then ask AI for a first draft. It turns a half-day job into an hour of editing.
The running document can be scrappy. An illustrative month: "3rd: two clients with the same roof-leak claim, neither had photos of the premises; tip on photographing your building. 11th: new account handler starts, covering fleet. 19th: one insurer adding three questions to its cyber questionnaire at renewal." The prompt asks for a 400-word newsletter with each note as a short section and a closing reminder about photos. Most of the editing hour goes on the cyber section, where the AI will fill in what the new questions are unless you paste the insurer's actual notice.
Case studies from call notes. With the client's permission, a transcript of a short interview becomes a draft case study. The failure to look for is the invented quote. An illustrative draft for a restaurant client ends with the owner saying "They made the whole renewal effortless", and nothing like that sentence appears in the transcript; the AI wrote what case studies usually say. Add "use only quotes that appear word for word in the transcript" to the prompt, and send the finished draft to the client to approve before it goes anywhere public. Watch across all marketing: claims about what a policy covers must be checked by someone who knows, and AI drafts drift towards a bland house style unless you give it examples of your own writing.
Client service and claims
Email sorting. AI reads each incoming email, labels it (new claim, renewal, mid-term change, complaint, general question), pulls out the policy number and routes it. You can do this with the AI already in Outlook or Gmail, or with an automation platform. Adding AI steps to Zapier shows the classify-and-route pattern. Watch: set a rule that anything that looks like a complaint or a claim goes to a person, even if the AI isn't sure, because a misfiled complaint is the costliest error here.
Ask the AI to return the same fields every time, so the routing step can read them. An illustrative output for one email:
category: mid-term change
policy_number: [as written in the email, or "none found"]
urgency: normal
needs_person: yes
reason: client wants to add a driver from Monday; no claim mentionedThe reason line earns its place in the weekly check. Say an email about a van accident gets tagged "mid-term change" because the client also asks to add a driver. The reason line shows exactly why, and the fix is one more rule: any mention of an accident, damage or theft means "claim", whatever else the email asks for.
Draft replies from a knowledge base. Routine questions (how do I make a claim, when is my renewal, can I add a driver) can get a drafted reply built from your own approved answers. If you're on HubSpot Professional or Enterprise, its Customer Agent costs 50 credits, about $0.50, per resolved conversation; it isn't available on Starter. Watch: the AI must never tell a client they're covered for something. Coverage questions go to a broker.
The approved answer is the source; the AI only adapts it. The brokerage's stored answer for "can I add a driver?" might say: "Yes, usually the same day. We need their full name, date of birth, licence details and any claims or convictions in the last five years. The insurer may change the premium." The drafted reply to a client asking about their 19-year-old son keeps every one of those points, uses the son's name from the email, and ends with "we'll confirm any change to your premium before it applies". What it must never add is "that should be fine", which is a coverage judgement dressed up as reassurance.
Call notes without typing. A note-taker records and summarises client calls. Otter.ai Pro is $8.33 per user a month billed annually ($16.99 monthly), and Business is $19.99 annually ($30 monthly). Watch: tell clients the call is being recorded and why, and check the summary against what was actually agreed before it goes on file.
A typical slip, in illustration: the summary records "client agreed to raise the excess to $1,000 to reduce the premium" when the client actually said "let me think about raising it". Filed unchecked, that's a change the client never approved. The fix is a two-minute read of every summary on the same day, and a summary template with separate headings for "agreed" and "discussed, not decided".
Finance and accounts
Transaction coding and reconciliation. Your accounting software may already do this. QuickBooks Online's Accounting AI agent categorises transactions and helps with reconciliation; in Xero, the JAX assistant takes on work such as bank reconciliation. Switch on what your software has and measure how many items still need a person each month. In an illustrative month, the brokerage's bank feed has 340 lines; rules and AI suggestions code 290, and the bookkeeper handles the other 50, mostly one-off suppliers and split payments. Track that 50. If it hasn't shrunk after three months, the rules need attention, not more AI.
Credit control. AI drafts payment reminders that change tone with the age of the debt: friendly at 7 days, firm at 30, formal at 60. The brokerage's credit controller reviews and sends a batch in 20 minutes instead of an afternoon. Chasing late payments with AI reminders has wording for each stage. At 30 days, an illustrative draft reads: "Our invoice [number] for [amount], due on [date], is now 30 days overdue. Please arrange payment this week or reply with a date you can pay by. If something's wrong with the invoice, tell us and we'll put it right." The one line the controller adds herself, for premiums paid by instalment, is what happens to cover if payment stops, taken from the insurer's terms and never from the AI.
Supplier invoice capture. Emailed PDF invoices are read, the supplier, date, amount and reference extracted and posted as draft bills for approval. Watch across finance: AI doesn't spot payment fraud by default. Any change to a supplier's bank details must be confirmed by phone on a known number, whatever the email or the AI says.
Here's how that shows up in practice, as an illustration. An email that appears to come from the office-cleaning contractor says "we've changed banks, please use the details on the attached invoice". The capture tool reads the new account number perfectly and posts a tidy draft bill. Nothing looks wrong because, to the AI, nothing is wrong: it extracted exactly what was on the page. Only the call to the number already on file shows the contractor never sent it.
Operations and compliance
Procedures from people's heads. Record your most experienced broker talking through how they handle a mid-term adjustment, then ask AI to turn the transcript into numbered steps with the checks at each stage. It's the fastest way to write down what only one person knows. Documenting your processes before adding AI explains why this comes before any automation.
From a 12-minute recording, an illustrative first draft might open: "1. Log the change request in the policy system with the date received. 2. Check whether the change affects the premium; if it does, get the insurer's figure before telling the client. 3. Confirm the change to the client in writing, with any premium difference and the date it takes effect." The senior broker's review usually adds what the recording skipped because it's second nature to her, such as "first check the client has no missed instalments".
A first pass over file notes. Give AI your internal checklist for what a client file note must contain (the client's needs, the options considered, the reason for the recommendation) and ask it to list what's missing from each note. It's a quick way to find gaps before an internal review. Watch: it's a first pass, not a sign-off; a person still reviews the file. An illustrative result for one note: "Needs: recorded (fleet of six vans, two drivers under 25). Options considered: two quotes mentioned; a third insurer declined, reason not recorded. Reason for recommendation: missing; the note says 'went with Option A' but not why." That last line is the gap an internal reviewer would have raised, found in seconds.
Comparing documents. Last year's policy wording against this year's, an old contract against a new one, a supplier's terms against your standard terms: AI lists the differences with section references, and a person checks each flagged change against both documents.
In an illustrative renewal for a client's workshop, the comparison flagged two changes: the escape-of-water excess rising from $1,000 to $2,500 in section 4, and the unoccupied-premises condition shortening from 60 days to 30 in section 7. Both were real. The broker's own read found a third that the list missed: the definitions section now said "the premises" excluded detached outbuildings, which quietly narrowed several other sections without changing a word of them. Where the client keeps stock in a yard store, that's the change that matters most. Tell the AI to compare the definitions first, and still read that section yourself.
HR and people
Job adverts and interview questions. From a short role description, AI drafts an advert, a list of interview questions tied to the skills that matter and a simple scoring sheet. Ask it to remove jargon and anything that might put off good applicants unnecessarily.
A before and after from the brokerage's account handler advert, with the real job details pasted in: "We're looking for a rockstar self-starter to join our fast-paced, dynamic team" became "You'll look after renewals and changes for about 150 commercial clients, with a senior broker to ask about anything unusual. Hours are 9 to 5.30, with one late shift a fortnight." The second version tells applicants what the job actually is, which is what the good ones are looking for.
Onboarding plans. A two-week plan for a new account handler, with who they shadow, what they read and what they should be able to do by day ten, drafted in minutes and adjusted by their manager. Give the AI your real systems and team names, or it will invent a generic induction full of steps you don't have. The plan is also a useful check on your own process: if the draft has nothing sensible to say about day three, neither does your current onboarding.
Answering staff questions from the handbook. "How much holiday can I carry over?" answered from your own handbook rather than someone's memory. An AI helpdesk for staff HR questions covers the setup. The real test is what it does when the handbook is silent. Asked "Can I carry over holiday if my leave was cancelled in a busy renewal month?", a good illustrative answer is: "The handbook allows up to five days to be carried over with your manager's approval (section 4.3). It doesn't cover leave cancelled by the business, so please ask [HR contact]." A badly set up helpdesk invents a policy to fill the gap.
Watch in HR especially: AI that screens or ranks candidates is classed as high-risk in Annex III of the EU AI Act, with those obligations now applying to stand-alone systems from 2 December 2027. If you recruit in the EU, keep a person making every shortlisting decision, and test any screening tool for bias whatever your location. Staff data is sensitive; keep it in business plans only.
The owner's desk
Monthly commentary on the numbers. Paste the month's figures and last month's commentary; ask for a one-page draft explaining what changed and what needs a decision. Check every figure it quotes. It turns a blank page into an editing job. An illustrative opening line: "Commission income was $4,300 below last month, mainly because two large fleet renewals moved into next month; new-business enquiries rose from 31 to 38." Check both parts: the figures against the pack, and the "because", which AI will supply confidently if your notes don't.
Preparing for decisions. Before a pricing change, a new hire or a new office lease, ask AI to argue the case for and against, list what you'd need to believe for each, and suggest the questions to put to your accountant. Say the owner is weighing a second account handler against more automation in client service. A prompt that asks "What would have to be true for the hire to pay back within a year, and what would have to be true for automation to?" produces a list of assumptions the owner can check against real figures. The AI doesn't make the decision; it makes the owner's thinking visible.
Your own inbox. Owners are often the biggest email bottleneck. Built-in AI in Outlook or Gmail can summarise long threads and draft replies in your voice for you to approve. Take a 23-message thread with the landlord and a fit-out contractor about refurbishing the brokerage's office. An illustrative summary reads: "Agreed: work starts on the 14th. Open: landlord to confirm weekend access; who pays for the new door-entry system." The owner then glances at the last three messages, a habit worth keeping with any long thread, and finds the landlord confirmed weekend access in message 21. With that one correction the summary is right, and replying takes two minutes instead of re-reading 23 emails.
The first use case to try in each department
| Department | Best first use case | Tool route | Setup time | What to measure |
|---|---|---|---|---|
| Sales | Recommendation letters from quote comparisons | Chat assistant with a saved prompt | 1 to 2 hours | Minutes per letter |
| Marketing | One guide into many formats | Chat assistant | 1 hour | Pieces published per month |
| Client service | Email sorting and routing | Built-in email AI or an automation platform | 1 to 3 days | First-reply time, misrouted emails |
| Finance | Coding and reconciliation | Your accounting software's AI | A few hours | Items needing a person each month |
| Operations | Procedures from transcripts | Note-taker plus chat assistant | 2 hours per procedure | Procedures written and used |
| HR | Adverts and onboarding plans | Chat assistant | 1 hour | Time from vacancy to advert |
| Owner | Monthly commentary | Chat assistant | 1 hour | Time to produce the pack |
Don't start everywhere at once. Pick the one department where the first use case scores best on hours saved and ease of checking, get it working for a month, then move to the next. The shared pieces, such as the usage rules, the prompt library and the approved tool, get built once and reused. What a small business should automate first has a scoring method if two departments look equally promising.
Further reads
- AI Quick Wins: 12 Things a Small Business Can Set Up This Week — Twelve things you can set up this week, across the business.
- AI Customer Service for Small Businesses: What to Automate First — Go deeper on the client service use cases.
- AI for HR in Small Businesses: 12 Tasks You Can Hand Over Safely — Twelve HR tasks you can hand over safely.
- Where AI Saves Time in a Small Insurance Brokerage — Where the hours actually go in a brokerage.
- AI Invoice Processing: Stop Typing Supplier Bills by Hand — The finance use case of stopping hand-typed supplier bills.
- How to Map Your Customer Journey and Find Where AI Helps — Find use cases by following a customer instead of an org chart.
- AI Use Case Template: Score Every Idea on One Page — A one-page template with scoring anchors, knock-out questions and a worked veterinary example for ranking AI ideas before you spend anything.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: HubSpot Breeze and credits pricing pages, Otter.ai pricing page, QuickBooks Online AI agents overview, Xero JAX product announcements, Microsoft 365 and Google Workspace plan pages, EU AI Act Annex III timetable as amended by the Digital Omnibus on AI.