Automate first the processes that happen at least weekly, arrive as text (emails, PDFs, forms, call notes), follow rules you can write down, and get checked by a person before anything leaves the business. Inbox sorting, meeting actions, enquiry logging, call summaries and a weekly summary email usually top the list; decisions about money, credit or staff come last.
To make that concrete, the 20 processes below are ranked by a simple score out of 25, and each gets a worked example from a real type of business, the tools involved, a rough running cost and the way it goes wrong. Your own ranking will differ: a job you do twice a year drops right down, and a job that eats three hours every morning jumps up. Score your own list the same way before you build anything.
The five factors behind each score
| Factor | Scores 1 when... | Scores 5 when... |
|---|---|---|
| Frequency | A few times a year | Many times a day |
| Minutes saved per run | Under a minute | Ten minutes or more |
| Text-heavy input | Clean numbers already in a system | Emails, PDFs, photos, speech |
| Safe to get wrong | An error costs money or trust immediately | An error is cheap and caught quickly |
| Easy to review | Checking needs an expert and takes as long as doing | Anyone can check it in seconds |
The "text-heavy" factor matters because it's where AI earns its place. If the input is already clean data, a plain rule-based automation does the job more cheaply and predictably. The broader question of where to begin is covered in what a small business should automate first with AI; this list is the catalogue of candidates.
Start here: scores of 21 to 23
1. Sorting a shared inbox into sales, support and accounts (23)
A new email arrives in the general inbox; an AI step reads it and assigns one of five labels (new enquiry, customer query, invoice or payment, supplier, junk) plus an "urgent" flag; the automation applies the label and posts urgent ones to the team chat. For an illustrative courier firm receiving about 40 emails a day, the controller's 45 minutes of morning skimming shrank to ten minutes checking an "unsure" folder.
In Zapier the trigger is free and each action step is a task, so roughly two tasks per email (an AI step using your own model key counts as one; Zapier's built-in AI step uses one, three or five depending on the model tier), about 1,800 tasks a month: the 2,000-task Professional tier at $49 a month billed annually. Watch-out: never let the automation delete or archive anything, and send damage claims or complaints to a person whatever the label. AI email triage for shared inboxes covers label design.
2. Meeting notes into action lists (22)
After a recorded meeting, a recap tool produces a transcript and summary; an AI step pulls out each action with an owner and due date; the automation creates tasks in your task app. An illustrative packaging supplier's weekly 45-minute production meeting used to produce actions on a whiteboard photo. The extraction came back like this: "Reorder 3mm double-wall board, purchasing lead, by Wednesday; confirm night-shift cover for the 14th, production manager, by Friday; NO OWNER: fix the label printer jam."
That last line is the value: unowned actions become visible. Tool notes: Teams' intelligent recap needs Teams Premium or a Microsoft 365 Copilot licence; Google Meet's "Take notes for me" needs Business Standard or above; Otter.ai Business is $19.99 a user a month billed annually. Watch-out: pause recording for anything confidential, such as a staff issue.
3. Enquiries logged in the CRM with a summary (22)
A web form or enquiry email arrives; AI extracts name, company, what they need, volumes and deadline; the automation creates a contact and deal in the CRM with a three-line summary as a note. An illustrative testing laboratory gets about 120 enquiries a month ("need 40 water samples tested for metals, results by the 20th"). At three tasks per enquiry that's about 360 Zapier tasks a month, inside the 750-task Professional tier at $19.99 a month billed annually, and HubSpot's free CRM holds the records.
Before, enquiries lived in one person's inbox and a third were never logged. After, every enquiry has a record the same hour. Watch-out: duplicates. Search for the email address before creating a contact, and let a person merge near-matches rather than the automation.
4. Call and voicemail summaries into the customer record (21)
If your phone system produces call transcripts, an automation can send each one to an AI step that writes a short summary, extracts any order or request, and logs both against the customer with a callback task where needed. For an illustrative wholesaler taking 25 phone orders a day, a summary might read: "Customer TR-0412 wants 200 x 40mm stainless clamps and 50 blue caps (code not given), delivery Tuesday. Callback: confirm cap code."
The AI cost is trivial: on OpenAI's gpt-5.6-luna ($0.20 per million input tokens, $1.20 per million output) a few hundred calls a month costs well under a dollar. Watch-out: transcription mishears product codes and numbers, so treat extracted codes as suggestions and read them back to the customer on the call. Tell callers that calls are recorded.
5. A weekly summary email built from your own numbers (21)
Every Monday at 7am, the automation pulls last week's rows from the sales sheet and shipment tracker, asks an AI step for a 150-word summary with the three figures that moved most, and emails it to the owner. An illustrative import-export business's version: "Revenue 38,400, down 12% on the four-week average; two containers delayed at the port, both for the same customer; overdue receivables up to 21,000, mostly two accounts."
It runs four or five times a month, so running costs are negligible; Make's free plan (1,000 credits a month) covers it, while Zapier's free plan won't because it only allows two-step Zaps. Watch-out: models love to explain why numbers moved, and they guess. Instruct it to report changes and flag anything over 20%, and leave the explanations to you.
Second wave: scores of 17 to 20
6. Emailed purchase orders into the order system (20)
Trade customers email orders as typed lists, spreadsheets, PDFs or photos. An AI step extracts the lines as structured fields; a lookup maps customers' descriptions to your product codes; complete orders become drafts in the order system, and anything uncertain goes to a person with the reason. For an illustrative wholesaler with 600 emailed orders a month and six minutes' keying each, a draft that takes 90 seconds to check would save 45 hours a month if every order flowed through; with a quarter still going to a person, the saving is nearer 34 hours.
It scores lower than the inbox examples only because an error ships the wrong goods. Watch-out: the mapping table ("blue caps like last time" means SKU CP-220-BL for this customer) is the real work, and it needs an owner who updates it weekly.
7. Supplier invoice capture and coding (20)
Invoices arrive by email; extraction reads supplier, date, amounts and line items; the automation matches them to purchase orders and proposes an expense account, and the bookkeeper approves in the accounting software. An illustrative spare-parts manufacturer handling 250 supplier invoices a month cut data entry from about 15 hours to 4 hours of checking.
Dedicated capture tools and most accounting packages do this natively, so check what you have before building anything. Watch-out: never let an automation accept changed bank details on an invoice; that's the classic payment-fraud route, and it needs a phone call to a known number. Matching supplier invoices to purchase orders automatically covers the matching rules.
8. Payment reminders that change tone by stage (19)
A daily check finds invoices 3, 14 and 30 days overdue; an AI step drafts a reminder that mentions the customer's PO number and what was delivered; stages one and two send automatically, stage three waits for approval. An illustrative packaging supplier with around 30 overdue invoices a month moved from "whenever someone remembers" to consistent chasing, and the stage-one note ("Just a reminder that invoice 4471 for your PO 5520, delivered 2 September, fell due on Friday") drew replies within days.
Watch-out: exclude disputed invoices. If a customer is mid-complaint about a delivery, an automatic chaser is the fastest way to lose them. Keep a "disputed" flag in the accounting software and have the automation check it first.
9. Draft replies to repetitive customer questions (19)
For the questions you answer twenty times a day, such as "where is my parcel?", "can you redeliver tomorrow?" and "can I change the address?", an AI step can look up the tracking status and draft a reply for staff to send with one click. An illustrative courier firm answered about 70 such emails a day, at three minutes each; drafted replies took under a minute to check.
Dedicated help-desk AI can go further and answer on its own: Intercom's Fin charges $0.99 per resolved outcome, and HubSpot's Customer Agent uses about $0.50 of credits per resolved conversation on any Professional or Enterprise hub (usually Service Hub). Watch-out: address changes are a fraud and theft risk. Drafts can acknowledge the request, but a person verifies the requester before any change.
10. Review requests and reply drafts (18)
Five days after an order is delivered, the automation sends a short review request; when a review comes in, an AI step drafts a reply for someone to post. For an illustrative spare-parts manufacturer selling replacement parts online, a draft reply to "part fitted perfectly, delivery a day late" reads: "Thank you, glad the bearing housing was a good fit. Sorry it arrived a day later than promised; we've raised the late collection with our carrier." That's specific enough to sound human.
Watch-out: never offer anything in exchange for a review, and write replies to negative reviews yourself, using the draft only as a starting point. Platforms' rules on review requests vary, so read them before you switch on automated requests.
11. Receipt and expense capture (18)
Staff photograph a receipt; extraction reads date, supplier, amount and tax; the automation files it, adds a row to the expense sheet and routes it for approval. For an illustrative courier firm whose 20 drivers each submit about 12 fuel receipts a month, the office had been typing 240 receipts from a shoebox. With the photos arriving by an upload form that asks for the vehicle registration, typing dropped to checking the few the model couldn't read.
Watch-out: blurry and duplicate photos. Ask the AI step to return "unreadable" rather than guess, and have the automation reject any receipt with the same date, amount and supplier as one already filed.
12. Bank transaction categorisation suggestions (17)
Accounting packages already suggest categories from bank rules, and many now use AI to propose more. The automation gain here is usually switching those features on properly and reviewing them weekly rather than building anything. An illustrative import-export business with 600 bank lines a month found the suggestions right for most routine lines; the misses were transfers between its own currency accounts, which were being proposed as sales income.
Watch-out: some AI features can't be switched off individually (Intuit says this of QuickBooks Online's), so review what they post rather than assuming they're off. Keep your accountant's categorisation rules written down and check a sample of AI-categorised lines each month.
13. Quote first drafts from an enquiry and a price calculator (17)
An enquiry arrives with dimensions, quantities and specification; an AI step extracts the spec into fields; a spreadsheet calculator (plain formulas, no AI) prices it; the AI then drafts the covering email; an estimator approves. For an illustrative packaging supplier quoting custom boxes, the estimator's 25 minutes per quote dropped to about 8, and quotes went out the same day.
The design rule is non-negotiable: AI extracts and writes, formulas price. A model asked to "estimate the price" will produce a confident number with no basis. Watch-out: incomplete enquiries. The extraction should list missing details (board grade, print colours) so the draft asks for them instead of assuming.
Worth building once the basics run: scores of 13 to 16
14. Goods-in checks against purchase orders (16)
Staff photograph the delivery note at goods-in; extraction reads supplier, PO number, codes and quantities; the automation compares them with the open purchase order and flags differences before stock is booked. For an illustrative spare-parts manufacturer receiving 35 deliveries a week, the check caught short deliveries that used to surface only at month-end stock counts. The details, including when stock can update without a person, are in whether AI can read delivery notes and update stock automatically.
Watch-out: a delivery note says what the supplier claims it sent. The count on the bench beats the paperwork every time.
15. Supplier quote comparison tables (16)
Four suppliers reply to a request for quotation: one PDF, one spreadsheet, two emails. An AI step reads all four into one table: unit price, currency, delivery terms, lead time, minimum order and payment terms, plus a column for anything missing. For an illustrative import-export business sourcing valve bodies, a two-hour job became twenty minutes of checking.
Watch-out: quotes on different delivery terms aren't comparable on unit price. A price "ex works" excludes freight that a delivered price includes, so the comparison needs your own freight and duty assumptions added as formulas. Tell the model not to convert currencies unless you give it the rate.
16. Shipment document consistency checks (15)
Before documents go to the freight forwarder, an AI step compares the commercial invoice, packing list and purchase order: do quantities, weights, product descriptions, commodity codes and consignee details match across all three? For an illustrative import-export business shipping 30 consignments a month, one mismatched carton count can hold a shipment for days, and the check takes the model seconds.
Watch-out: consistency isn't correctness. The AI can tell you the commodity code is the same on every document; it can't confirm it's the right code for the goods. Classification stays with whoever is qualified to make it.
17. A staff assistant that answers from your procedures (15)
Load your SOPs and method sheets into a shared ChatGPT or Claude Project or a Gemini Gem (becoming a skill from November 2026), and staff can ask "what's the holding time for chilled samples?" and get the answer with the SOP number. For an illustrative laboratory with 60 procedures, new technicians stopped interrupting the senior analyst for routine lookups. (Don't build this as a custom GPT: OpenAI is retiring them.)
It scores lower because a wrong answer about a safety step matters, and old versions cause most wrong answers. Keep only current versions in the Project, and instruct the assistant to send any safety question to the supervisor as well.
18. Tagging complaints and returns for monthly patterns (15)
Each returns note or complaint gets an AI-assigned tag from a fixed list (damaged in transit, wrong item picked, customer ordered wrong item, quality fault, other), and a monthly pivot shows the pattern. An illustrative wholesaler's first month of 200 tagged returns showed wrong-item picks concentrated on two pairs of products with near-identical codes, fixed by moving them apart on the shelves.
Watch-out: tags drift if the list isn't fixed. Give the model the exact list, allow "other" with a one-line reason, and read the "other" pile monthly to see whether the list needs a new tag.
19. First drafts of report cover notes (14)
When results are authorised, an AI step drafts a plain-English cover note for the client: which results sit inside the limits on their method sheet, which don't, and what the client asked to be told. For an illustrative laboratory issuing 350 reports a month, the analyst's ten minutes per cover note fell to about three minutes of review.
It ranks low because the cost of an error is high: a note saying "all within limits" when one result isn't is serious. Every draft is reviewed by a qualified person against the report, and the model is instructed to quote figures exactly and never interpret beyond the limits given.
20. A credit check pack for new trade accounts (13)
When a new trade account application arrives, the automation gathers the credit agency report, trade reference replies and any filed accounts, and an AI step summarises them into one page: years trading, payment behaviour, red flags, references received. For an illustrative wholesaler opening 15 accounts a month, the credit controller's reading time halved, and every limit decision stayed with the credit controller.
It sits last on purpose. If the applicant is a sole trader, you're assessing an individual's creditworthiness, which the EU AI Act lists as a high-risk use if you trade in the EU, and data-protection law restricts solely automated decisions about people. The AI summarises; a person decides and records why.
Picking your first three from the list
Here's how an illustrative packaging supplier with 16 staff used the scores. It shortlisted five processes that applied to it, then re-scored them for its own volumes:
| Process | Its own volume | Re-scored | Decision |
|---|---|---|---|
| Inbox sorting | 15 emails a day | 17 | Later: volume too low to matter |
| Meeting actions | 3 meetings a week | 21 | First |
| Quote first drafts | 60 quotes a month, 25 minutes each | 20 | First |
| Payment reminders | 30 overdue a month | 19 | First |
| Complaint tagging | 20 a month | 12 | Later |
The quote drafts alone saved about 17 hours a month (60 quotes × 17 minutes), and faster quotes were the owner's real goal. Notice that inbox sorting, top of the general list, dropped for this business because it didn't get many emails. Scores are a starting point; your volumes decide. Before building, map each chosen process so the automation handles its exceptions, and use adding AI steps to Zapier if that's where you'll build.
What to leave off the first list
- Decisions about people or money without review: hiring, credit limits, pricing, refunds above a threshold.
- Customer-facing replies that send themselves before you've run drafts past a person for at least a month.
- Anything done less than weekly, unless each run takes hours; the build and upkeep outweigh the saving.
- Processes that work only because one person remembers something. Write it down and fix it first.
- Anything where you can't say how you'd know it went wrong. If there's no check, there's no automation yet.
Pick two or three processes from the top of your own list, run them for a month with a person reviewing every output, and measure the minutes saved against the time spent fixing. The ones that pass become permanent; the ones that don't teach you what to fix before the next round.
Further reads
- How to Turn Meeting Notes Into Tasks Automatically With AI — A full build of example 2, from recap to task list.
- AI Call Summaries: Log Every Phone Enquiry in Your CRM — Example 4 in detail, with phone system options.
- How to Get a Weekly Business Summary Emailed to You by AI — Set up example 5 step by step.
- How to Chase Late Payments With AI Reminders That Sound Human — Reminder wording for example 8 that sounds human.
- How to Compare Supplier Quotes Side by Side With AI — The method behind example 15, with a comparison template.
- How to Spot Patterns in Complaints, Returns, and Defects With AI — Turn example 18's tags into decisions.
- How to Calculate the ROI of an AI Automation Before You Build It — Put numbers on your top three before you build.
- How to Add Human Approval Steps to AI Automations — Add the review step most of these examples rely on.
- AI Quick Wins: 12 Things a Small Business Can Set Up This Week — Twelve small AI jobs sorted by setup time, each with a copyable prompt, a real business example and the catch to watch for.
- What Finance Tasks Can AI Automate in a Small Business? — Fifteen finance jobs AI can take over in a small business, each with an example, a first step and the check that stops it going wrong.
- 8 Worked AI Automations for a Service Business, With Costs — One illustrative testing laboratory builds eight automations, from enquiry logging to a Monday summary, with task counts, AI costs, setup time and what broke.
- How Much Does It Cost to Automate One Workflow With AI? — Every cost line in one AI-automated workflow, with a returns-email example costed on Zapier and Make, three budget scenarios and where bills creep up.
- How to Audit Your Workflows for AI Opportunities Yourself — A six-stage AI opportunity audit you can run yourself in about eight hours: inventory, timing, sorting, scoring, desk tests and a ranked register.
- What Can AI Realistically Do for a Small Business? 15 Tasks — Fifteen jobs AI can realistically take on in a small business, rated by how much you can hand over, with tools, costs and the catch for each.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Zapier pricing and task-counting help pages; Make pricing; HubSpot, Intercom, Otter.ai and Microsoft and Google Workspace plan pages as summarised in the facts sheet; EU AI Act Annex III; facts sheet for API token prices. Scores, volumes and timings are illustrative.