No talk of transformation or productivity. Five concrete tasks you can hand to AI in a business system this week, three where a human should have the last word, and one rule that holds it all together.
Most writing about AI in companies ends at the sentence that it “boosts productivity”. That's hard to act on at nine o'clock on a Monday. Here's a more specific list: five tasks AI in a business system does reliably today, and three where it's sensible to let a human decide.
Five things you can hand over today
1. Pulling the details out of an incoming email
An enquiry arrives as ordinary prose: “we'd need twelve units by the end of September, please invoice to…”. AI extracts the item, quantity, date and company identifier and creates a record from them.
Why it works: identifying details inside a sentence is precisely what language models are good at. And you can check the result at a glance.
2. Drafting a quote or summary from data you already have
Not from nothing - from a specific record. It takes the line items on a job and the history of correspondence with the client, and generates the quote text or a covering email in your tone of voice.
Why it works: writing from a template and a source is a safer task than inventing content.
3. Triaging incoming requests
Tickets, enquiries, complaints. AI reads the content and assigns a category, a priority and an owner according to rules you set.
Why it works: triage has clear categories and a misfiling is easy to correct - the cost of an error is low.
4. Summarising a long thread into three sentences
Fifty messages in a client conversation that a colleague now has to step into. Instead of an hour of reading, they get a summary of what was agreed and what's outstanding.
Why it works: the original is still there. A summary is a shortcut, not a replacement.
5. Watching a deadline and reminding you
Not a calendar alert, but a rule that looks at data: a job is finished and has no invoice, a payment is due in three days and hasn't arrived, a service visit is due a year on.
Why it works: it's deterministic. Either the condition holds or it doesn't.
Five tasks for today. Three where a human decides.
The difference between the two columns isn't what AI can technically manage - it's who carries the consequences when it gets it wrong.
Three things to leave to a human
1. Decisions without review
AI that approves a discount, sends an invoice or replies to a client with nobody looking. It isn't that it can't manage - it's that when it gets it wrong, you find out from the client.
The safe version: AI prepares, a human confirms with one click. Same time saved, error caught.
2. Legal and tax output
Contracts, liability assessments, tax questions. AI will write you the text and it will sound convincing. Responsibility for it sits with your company.
The safe version: AI as preparation of material for a lawyer, not as a substitute for one.
3. Working with personal data without rules
Names, contacts, health information, salaries. Before any of that goes into any AI tool, it has to be clear where the data goes and who holds it. That's covered in detail in You upload company data into AI. Where does it actually go?.
One rule that holds it together
AI proposes, a human confirms. For anything that leaves the company or moves money, keep the last step with a person. That single rule addresses most of the worries people have about deploying AI - and costs almost none of the time saved.
How to start so it doesn't stay an experiment
The most common reason AI “didn't work out” in a company isn't the technology. It's that ten things were tried at once and none of them finished.
A pattern that works:
- Pick one activity from the first list. Ideally one somebody does daily and finds tedious.
- Let AI propose, not decide. For the first fortnight, have a person check every output.
- Count how often intervention was needed. Once it's below one case in ten, some steps can be let through.
- Only then add a second activity.
What AI agents can genuinely do in companies today, and where they stop, is covered in AI agents in business. A concrete walkthrough of an email becoming a job is in AI reads the email and creates the job.
Common questions about deploying AI in a company
Do we need someone technical for this?
For the items on the first list, no. They're settings in the system where your data already lives. A technical person becomes useful when connecting outside services.
How do we tell when AI is getting something wrong?
By having a person confirm every output for the first few weeks and counting how often they had to step in. Without that measurement, “it seems to work well” is only an impression.
Will it replace a role?
In small and mid-sized companies, so far mostly not. What actually happens is that admin shrinks - retyping, triaging, hunting - and time appears for work there was never room for.