The wrong question
"Which business processes can be automated with AI" has one answer: almost all of them. Technically true. Practically useless, because no decision follows from it.
The useful question is different: where does checking the result cost less than doing the work. AI does not take the responsibility for the result off you. It only changes who writes the first version.
A three-question test for every process
Take a process and answer three questions. The numbers need not be exact — the order of magnitude is what matters.
1. What does it cost to check the result, compared with doing it yourself? If checking takes as long as the work, automation gives you nothing. A short call summary is quick to check. A legal opinion is not: to confirm it is right, you effectively have to write it again.
2. What does one mistake cost, compared with an hour of a person's time? A misrouted email reaches the right team in a minute. A wrong figure in a commercial proposal becomes a commitment.
3. Is the input stable? Identical documents in one format make a good candidate. Every case being unique makes a poor one, because you will get a confident answer to a question the assistant understood differently.
A process that passes all three is worth trying first. One that fails even one should run only with a person who checks everything.
The hidden cost: review
This is where most savings estimates fall apart.
A hypothetical example. A manager answers 300 routine requests a month at 10 minutes each — 50 hours. An assistant drafts replies in seconds, but a person still reads and edits every one, say at 4 minutes each. That is 20 hours.
The saving is 30 hours, not 50.
And that is the optimistic case. If one reply in a hundred contains an error the client reads as a promise, those 30 hours have to be weighed against the price of that one promise. Sometimes it is higher.
Hence the rule we apply in process diagnostics: if a person still checks 100% of the output, you are not saving time, you are moving it from writing to reading. The saving appears only when checking can be done by sampling.
What does not pass — and why
Calculations with consequences. Tax, payroll, contract amounts. The mistake is noticed late, and fixing it costs many times the hours saved. An assistant can prepare a draft, but a person has to reconcile the figure — and then the gain is marginal.
Wording that becomes an obligation. Contract terms, replies to clients about deadlines and guarantees. A plausible phrase you did not mean carries the same legal weight as one you wrote deliberately.
Processes without a written standard. If three employees do one task three different ways, there is nothing to automate. First agree on the right way — and often that agreement alone gives more than any tool.
Decisions with personal accountability. Hiring, dismissal, lending. The issue is not the quality of the answer but who answers for it, and that cannot be handed to a system.
What usually passes well: sorting and routing requests, draft replies to routine questions with spot checks, extracting fields from documents in one format, meeting summaries for internal use.
They share one trait: the mistake is visible immediately, and it is cheap.
Will AI replace your support team
No, and promises to the contrary are a sign you are being sold a fantasy, not a tool.
Realistically, an assistant takes over part of the routine requests — the ones where the answer is always the same and can be checked at a glance. The complex, contentious and unusual ones stay with people, and their share of the team's work only grows: the easy cases disappear, the hard ones remain.
Our position: the first process to automate should be a boring one. Not the one with the most hours, but the one where mistakes show up fastest. The exception is a company that simply has no boring processes — then you start with the most expensive one, but with a person checking everything until the numbers show they can check less.
What it costs
The subscription is the smallest part. For reference: Microsoft 365 Copilot for business costs $21 per user per month on annual billing, and only on top of a base Microsoft 365 licence bought separately. Price checked on Microsoft's page on 23 September 2026.
The real cost is made of other things:
- Review time — as the example above shows, it can eat most of what you saved.
- Integration. An assistant that cannot see your CRM and email makes people copy data by hand. Connecting non-standard systems often costs more than a year's subscription.
- Describing the process. Work nobody has written down can be neither automated nor checked.
So cost an implementation in hours, not in seat price: how many hours the process takes now, how many will go to review afterwards, and what one mistake costs. If the number still looks good after that, the process is worth a try.