Twelve business processes a small team can hand to automation, each with its input, its output and the way it tends to go wrong. Two of them don't need AI at all.
Twelve is a realistic list, and no business needs all of it. Each process below comes with its input, its output and, most useful, how it breaks. With about eleven weeks left in 2026, one or two of these can be running before January, since a single, well-scoped automation takes two to four weeks to build. The worked examples are illustrative, set in tools like Gmail, HubSpot, Shopify and Stripe.
One opinion first: if a process involves no reading or writing, it probably doesn't need AI. Two of the twelve are like that.
Inbox and paperwork
1. Shared inbox triage. Input: everything that lands in the shared info or support inbox. Output: a label, a priority and a drafted reply that a person approves or edits. Failure: a polite email that hides a real complaint ("just checking you got my third message") gets filed as routine. Refunds, cancellations and anything mentioning a chargeback always go to a person. Running it is cheap. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output, so as an illustrative sum, 1,000 emails a month at about 1,500 tokens in and 200 out costs roughly 25 cents.
2. Supplier invoice entry. Input: PDF invoices arriving by email. Output: a draft bill in the accounting system with supplier, date, line items and tax filled in. Failure: a misread total, or the same invoice entered twice because the supplier sent it again as a reminder. Two checks catch most of it: the lines must add up to the printed total, and the supplier and invoice number must not already exist.
3. Purchase orders sent by email. Input: a PO as a PDF, or an email saying "same as last month but double the blue ones". Output: an order in Shopify or your order system. Failure: units. "10" can mean ten items or ten cases of twelve, and a model will pick one with total confidence. Flag any line where the quantity is far outside what that customer normally orders.
Sales
4. Lead qualification. Input: a web form or an enquiry email. Output: a CRM record with the fields filled in, a score and an owner. Failure: a long, friendly message outscores a short one from a buyer with money set aside. Let the model pull out the facts (service, timeline, budget) and let a rule you wrote do the scoring. You can read a rule. You can't read a model's hunch.
5. First reply and booking. Input: a lead that passed step 4. Output: a reply within minutes that answers their question and offers times from your own calendar. Failure: treating an existing customer as a new lead, or offering 9am in the wrong time zone. Check the CRM first. If most enquiries arrive by phone, see what AI can and can't handle on a business phone line.
Getting paid
6. Invoice chasing. Input: overdue invoices in the accounting system. Output: reminders on a schedule, polite first and firmer later. In the UK the law sets the numbers. With no agreed payment date, a business payment is late 30 days after the customer gets the invoice or the work is delivered, whichever is later. Statutory interest is 8% plus the Bank of England base rate unless your contract sets a different rate, and the fixed sum you can add for recovery costs is £40 on a debt up to £999.99, £70 up to £9,999.99 and £100 above that. Failure: chasing an invoice the customer has disputed, or one they paid yesterday. Check for a dispute note and recent payments before each send.
7. Failed card payments. This is the first of the two that need no AI. If you bill through Stripe, Smart Retries already chooses when to retry, and Stripe's recommended default is 8 tries within 2 weeks. It won't retry hard declines, such as a card reported stolen, so the piece worth building is the email asking for a new card, sent when the invoice.payment_failed webhook fires. Failure: telling someone their card failed an hour after they updated it. Check the invoice status first.
Operations
8. Client onboarding. Input: a signed contract or a paid first invoice. Output: accounts created, a welcome email, a document request and a kickoff call booked. Failure: chasing a document the client already sent in another thread. Search their email history before any reminder, and stop the sequence once a person replies.
9. Inventory sync. The second one that needs no AI. Input: stock changes from the till or warehouse. Output: the same stock level on every channel. Failure: the last unit sells twice in the gap before the update. Shopify's GraphQL Admin API allows 100 points a second on the Standard plan and 1,000 on Shopify Plus, so a big catalogue synced by polling can fall behind on a busy day. Use webhooks, and hold back a small buffer on fast sellers.
10. The weekly report. Input: sales, bookings or support numbers from a database or spreadsheet. Output: a short email saying what moved. Failure: a confident, wrong number. Models are poor at arithmetic over a spreadsheet, so have SQL calculate every figure and let the model write the sentences around them.
Customers
11. First replies to support questions. Input: a chat message or support email. Output: an answer from your own help pages, or a handover to a person. Failure: an answer that sounds right and isn't. In February 2024 a British Columbia tribunal ordered Air Canada to pay a customer C$812.02 after its website chatbot described a bereavement refund the airline's policy didn't allow. Air Canada argued, in effect, that the chatbot was responsible for its own actions. The tribunal called that a remarkable submission and held the airline responsible for everything on its website. So answer only from approved documents, link the policy page in the reply, and send anything involving money to a person.
12. Review requests. Input: a finished job. Output: one review request. Failure: asking only the customers who seemed happy. Google's contribution policy bans selectively soliciting positive reviews, and in the US the FTC rule on fake reviews, announced in August 2024, bans incentives that depend on the review being positive or negative. Ask everyone and offer nothing in return.
Why most of these fail the same way
Most of these failures are one mistake: the system acted on something a person would have questioned. A quantity ten times the usual. A reminder about a disputed bill. A refund the policy never offered. A card reminder to a customer who fixed it an hour ago.
So decide in advance which outputs a person approves, and keep that queue small enough that people read it. When we build AI automation for business processes, sensitive steps sit behind a human approval for exactly this reason, and it's the part we'd least want anyone to skip.
The twelve on one page
| Process | Where it breaks |
|---|---|
| 1. Inbox triage | A polite complaint looks routine |
| 2. Invoice entry | A total is misread or entered twice |
| 3. Purchase orders | Items and cases get mixed up |
| 4. Lead qualification | Enthusiasm outscores budget |
| 5. First reply | An existing customer is treated as new |
| 6. Invoice chasing | A disputed or paid invoice is chased |
| 7. Failed payments (no AI) | The card was already updated |
| 8. Onboarding | A sent document is chased |
| 9. Inventory sync (no AI) | The last unit sells twice |
| 10. Weekly report | The model does the arithmetic |
| 11. Support replies | It describes a policy you don't have |
| 12. Review requests | Only happy customers are asked |
Which one to start with
Pick the one with the most volume and the cheapest mistake. Invoice entry and inbox triage tend to win: every output is a draft until someone approves it, so an error costs a minute, not a customer. Support replies and invoice chasing speak to customers directly, so leave them until the first one has earned some trust.
Then count before anything gets built. For one week, note how many items come in for the process you picked and roughly how long each takes. To use illustrative numbers, 20 invoices at three minutes each is an hour a week, which won't repay a build. At 400 invoices it's twenty hours, and the case makes itself.
Four questions before you pick one
Which business processes are easiest to automate with AI?
Ones where the input is text or a document and the output can wait as a draft: inbox triage, invoice entry, emailed purchase orders and lead qualification.
Do all of these need AI?
No. Inventory sync and card payment retries are plain integrations. Without reading or writing in the process, a language model only adds cost and a new way to fail.
How much does it cost to run an AI automation each month?
The model is usually the smallest line. As an illustrative sum, 1,000 emails a month on Claude Haiku 5.5 at its listed prices comes to about 25 cents. The build and the testing cost far more.
How long until the first automation is live?
For one well-scoped process, two to four weeks is typical. The first week goes on mapping the process and counting how often it runs.
