Automate first a job that happens at least weekly, is mostly reading or writing text, takes someone two or more hours a week, and produces output a person can check in under a minute before any customer sees it. For most small businesses, that means drafting replies to routine enquiries, not complaints, pricing or anything that moves money.
It's a deliberately unglamorous answer. The job you most want rid of is often the wrong first pick, because it's usually annoying for a reason: it's messy, it needs judgement, or a mistake is expensive. Below are the traits to look for, the first jobs that fit most firms, what to leave for later, and a way to rank your own shortlist in one sitting.
The aim of a first automation isn't the biggest saving. It's a quick, safe win that teaches your team how AI behaves on your real work, so the second and third projects are chosen with experience rather than hope.
Six traits of a good first AI job, with thresholds
A job doesn't need to be perfect on every line, but a first project should clear at least five of these six.
| Trait | Threshold | Why it matters for a first project |
|---|---|---|
| Frequency | At least weekly, ideally daily | You need 30 or more repetitions in a month-long trial to see whether it really works |
| Time taken | 2+ hours a week across the team | Below that, setting up and checking can swallow the saving |
| Mostly text | Reading, writing, sorting or summarising | Current AI tools are strongest here; "if X then Y" jobs are cheaper with plain rules |
| Quick to check | A person can judge the output in under a minute | If checking takes as long as doing, nothing is saved |
| Cheap mistakes | Errors are caught before a customer, or cost little to fix | You will get things wrong while learning; do it where it doesn't hurt |
| Ordinary data | No health, financial or ID details needed | Keeps your first project clear of the hardest data-protection questions |
The "mostly text" trait catches a lot of people out. Sending a reminder the day before an appointment is automation, but it isn't an AI job: a booking system or a simple rule does it perfectly for almost nothing. If you're unsure which side a task falls on, AI versus rule-based automation walks through the test.
First jobs that suit most small businesses
These five pass the traits in most firms. Each keeps a person between the AI and the customer, which is the right shape for a first project.
Drafting replies to routine enquiries
Questions about prices, opening times, availability and "do you do this?" arrive every day and have stable answers. The AI drafts from a saved price list and FAQ; a member of staff reads and sends. What stays human: anything that mentions a complaint, a refund, an injury or an unusual request. Watch-out: the drafts are only as current as the price list you gave it, so date the list and update it when prices change.
Turning rough notes into proper records
Voice notes after a job, scribbled call notes, a supervisor's end-of-day handover: AI turns them into a tidy record in the format you use. Checking is quick because the person who made the notes reads the result. Watch-out: AI will fill gaps with plausible detail, so the instruction should say "if something isn't in the notes, write 'not recorded'".
For an illustrative two-person heating and plumbing firm, an engineer's voice note from the van ("Annual service at the house on the hill, boiler fine, pressure was low so topped it up, flue terminal a bit close to the new fence, told them to keep an eye, back in twelve months") comes back as a job record like this (illustrative output):
Job type: Annual boiler service
Work done: Service completed; system pressure low, topped up
Parts used: not recorded
Issues noted: Flue terminal close to newly installed fence;
customer advised to monitor
Follow-up: Next service in 12 months
Time on site: not recordedThe two "not recorded" lines are the instruction working: without it, the first draft had invented "no parts required" and "1 hour on site". The engineer's one real edit was the flue line. "Customer advised to monitor" undersold it, so he changed it to what he'd actually do: "Recommend check against clearance guidance at next visit." That's the kind of judgement that stays with the person, and checking took him about 30 seconds.
Summarising long emails and documents into actions
Supplier price updates, changes to terms, long threads with a client. The AI pulls out what changed and what needs doing, and the reader checks the summary against the original for anything important. Watch-out: never let a summary replace reading a contract clause that matters.
A realistic miss: a café owner pastes in a three-page price-change letter from a coffee supplier and gets back "Price per kilo rises 6% on the house blend from 1 November; decaf unchanged; new oat-milk range available" (illustrative). All correct. What the summary left out was a sentence near the end: minimum order for free delivery rises from $100 to $150. For a café ordering about $120 a week, that's a new delivery charge every week, and it mattered more than the 6%. Adding "list every change to prices, minimums, delivery or payment terms, however small" to the prompt caught it on the rerun.
First drafts of repetitive writing
Product and service descriptions, social captions, replies to positive reviews. These are high-volume, low-risk and easy to judge by eye. Watch-out: the facts in them (sizes, ingredients, prices) must come from your data, not the AI's general knowledge.
Pulling details out of emails and forms into a spreadsheet
Order forms, enquiry forms and booking requests often arrive as free text. AI can extract the name, date, item and quantity into columns. Checking a row takes seconds. Watch-out: extraction goes wrong quietly, so spot-check a sample every week rather than assuming it's fine.
Here's what quiet looks like. An order email to a small bakery reads: "Hi, for the office do on Friday 14th could we get two dozen sausage rolls, a large lemon drizzle and whatever brownies you have, collecting about 11." The extracted row came back as Date: Fri 14 | Item: Sausage rolls | Qty: 2 | Item 2: Lemon drizzle (large) | Qty: 1 | Item 3: Brownies | Qty: blank | Collection: 11:00. Every field looks tidy, and the sausage rolls are 22 short. Nobody would spot it from the spreadsheet; you only catch it by reading a few rows against the original emails. The fix was one line in the prompt ("convert words like dozen into numbers: a dozen = 12") and a Monday check of ten rows.
What to leave until later, even if it's the job you hate most
These jobs may well be worth automating one day. They're poor first projects because a mistake is costly, public or hard to spot.
- Complaints and refunds. They need judgement and tone, involve money, and an AI that gets the tone wrong makes the complaint worse.
- Quotes and prices. Language models can make arithmetic slips and state them confidently; why AI is bad at maths explains what to check. A typical slip: an AI-drafted quote for a sign-maker listing three shop signs at $85 each, "total $225". The line items are right and the total is $30 short, and on a first project nobody is yet in the habit of re-adding the AI's totals.
- A fully automatic chat assistant on your website. Every error is in front of a customer, and you haven't yet learnt how the AI fails on your questions.
- Categorising bookkeeping entries without review. Errors compound silently across months.
- Screening job applicants. Bias and fairness risks need more care than a first project allows.
- Health, safety or allergy advice. A fluent wrong answer here can injure someone.
There's also a strong case for starting behind the scenes rather than in front of customers; customer-facing or back-office AI first sets out the trade-off in full.
Three businesses, three different first picks
The traits stay the same, but the winning job changes with the business. These are illustrative examples, with rough figures of the kind you'd gather in a one-week log.
A nail salon with five technicians and a front desk
The front desk handles about 60 messages a week asking about prices, gaps in the diary, whether a particular design is possible and aftercare. At around three minutes each, that's roughly three hours a week. First pick: AI drafts replies to price and availability questions from a saved price list and the week's openings; the receptionist edits and sends. Not first: automatic booking, or anything about skin reactions to gel products, which goes straight to a senior technician.
An optician with two optometrists and four support staff
The tempting job is patient recall letters, but those already run from the practice system on rules, and anything clinical involves health information. First pick: summarising supplier and lab emails (order status, frame price changes, discontinued lines) into a daily list for the practice manager, which takes her about 40 minutes a day now. It involves no patient data at all. Not first: drafting answers to patients' questions about symptoms, which is a clinical conversation and a safety issue.
A pet shop with a web shop and about 1,200 product lines
New lines sit unlisted for weeks because writing descriptions is slow. First pick: first drafts of product descriptions built from the supplier's spec sheet, checked against the pack for ingredients and feeding amounts before publishing. Not first: answering "what should I feed my puppy?" automatically. Feeding advice depends on the animal, and a confident generic answer can be wrong for a particular breed or health condition.
Rank your own shortlist in one sitting
You need a week of rough data and about an hour. Ask each person to keep a tally of the repetitive tasks they do, then fill in a sheet like this:
Task | Times/wk | Mins each | Hrs/wk | Mostly text? | Check <1 min? | Could a mistake reach a customer unchecked? | Keep?
Price/availability msgs | 60 | 3 | 3.0 | Yes | Yes | No (desk sends) | YES
Aftercare questions | 15 | 4 | 1.0 | Yes | Yes | Yes (skin reactions) | no
Instagram captions | 5 | 15 | 1.25 | Yes | Yes | No | maybe
Supplier invoice checks | 8 | 10 | 1.3 | Partly | No | No | no
Day-before reminders | 120 | 0.5 | 1.0 | No (rule) | - | - | use booking system- Knock out anything where a mistake could reach a customer unchecked, or where checking takes longer than a minute.
- Move rule-shaped jobs (the same action every time something happens) to a separate list for your booking system or a simple automation tool.
- Sort what remains by hours per week and pick the top one. The second becomes your backup if the first stalls.
- Sanity-check the winner against whether the process itself is stable. If three people do it three different ways, standardise it before adding AI; how to tell if a process is ready to automate has the checks.
If you want a more formal assessment, seven tests to run before committing to a first AI project goes further on cost, ownership and risk.
When the best first move doesn't involve AI
Sometimes the log shows that the time isn't going where you thought. Three patterns point away from AI for now:
- The work is waiting, not doing. If a job takes a week because approvals sit in someone's inbox, AI won't shorten it. A clearer approval rule will.
- Everyone does it differently. Automating an inconsistent process locks in the inconsistency. Write down one agreed way first.
- The task is pure routine. Moving data from A to B, sending a fixed message on a fixed trigger: rules do this reliably, and your booking or accounting software may well have the feature built in.
None of these are failures. Clearing them often frees as much time as the AI project would have.
How you'll know the first pick was right
Before you start, write down how long the job takes now and how often it goes wrong. After four weeks of running it with AI, a good first pick shows four signs:
- Time per item has dropped noticeably, even after counting the minutes spent checking drafts.
- The share of drafts that need heavy rewriting is falling week by week as you improve the instructions.
- Nothing wrong has reached a customer, and the person checking can name the kinds of error the AI makes.
- The staff who use it would object if you took it away.
For the nail salon above, the four weeks looked like this (illustrative):
| Week | Messages drafted | Heavy rewrites | Average minutes per message, including checking | Errors caught before sending |
|---|---|---|---|---|
| 1 | 58 | 17 (29%) | 2.4 | 4 (all an old gel-removal price) |
| 2 | 61 | 9 (15%) | 1.8 | 1 |
| 3 | 55 | 5 (9%) | 1.4 | 1 |
| 4 | 63 | 4 (6%) | 1.3 | 0 |
Against a baseline of about three minutes a message, that's close to two hours a week back by week 4, with the rewrite rate falling every week. The week-1 errors were all one out-of-date price in the saved list, which is a typical first-week finding: most early mistakes trace back to the information you gave the tool.
If the time saving is there but staff quietly stopped using it, find out why before choosing the next job. It's usually a clumsy step, such as copying text between two windows, that a small change to the setup can remove.
The rule in one sentence
Pick the most time-consuming text job that a person can check in a minute before it reaches anyone outside the business, run it with that person checking every output for a month, and only then choose the second job. Starting small and safe feels slow, but it's the quickest route to AI that the whole team trusts.
Further reads
- AI Quick Wins: 12 Things a Small Business Can Set Up This Week — Smaller set-this-week ideas if you want a warm-up first.
- How to Run Your First AI Pilot Project in a Small Business — How to trial the job you picked over four to six weeks.
- How to Set a Baseline Before You Introduce AI — Measure the job now so you can prove the change later.
- How to Use AI on Your Own Admin First, Then Roll It Out — Why the owner's own inbox is often the safest testing ground.
- AI Workflow Automation Examples: 20 Processes to Automate First — Twenty worked processes if your shortlist is still empty.
- Is It Worth Automating a Task You Only Do Once a Week? — Checks whether a low-frequency job clears the bar.
- How to Get Started With AI in Your Small Business: First 7 Steps — Seven steps that take about a month, from a five-day time log to a keep-or-drop decision, with a yoga studio's numbers at each stage.
- How to Write a One-Page AI Strategy for Your Business — The seven boxes a one-page AI strategy needs, a filled-in garden centre example, and five tests that show whether your page will guide real decisions.
- 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.
- Do I Need AI in My Business? A Decision Guide for Owners — A decision guide for owners: the three situations where AI is worth it, when waiting is sensible, a ten-question score and a 30-day test.
- AI Use Cases by Department for Small Businesses (With Examples) — 21 practical AI use cases across seven departments, each shown in a small insurance brokerage, with how to start and what to watch.
- AI Adoption Stages: Where Is Your Business Now, and What's Next? — Place your business on a five-stage AI adoption scale with a ten-question check, then see the one move that gets you to the next stage.
- Automating a Broken Process: Why It Backfires and What to Fix — Why automating a broken process backfires, the five-pass fix to do first, and how to split the cleaned-up process between rules, AI and people.
- When Not to Use AI in Your Business: 9 Tasks to Keep Human — Nine tasks a person should always decide, what AI can safely prepare for each, and a three-question score for any task that isn't on the list.
- Generative AI vs Traditional AI: Which Does Each Task Need? — How generative and traditional AI differ in what they need, cost and get wrong, four questions that sort any task, and a print shop's six tasks sorted.
- Which Event Planning Tasks Should You Hand to AI First? — Twelve event planning tasks scored for AI, the five to hand over first with worked examples, and the ones to keep even though AI will attempt them.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.