engineering notes
Automating business processes with AI: what actually works
Which business processes AI actually automates — classification, routing, drafts, status reconciliation — and where it doesn't pay off.
In brief for executives. AI automates not “business in general” but specific repeated decisions with a measurable output: classification, routing, draft preparation, status reconciliation. The effect arrives concentrated — in a few functions, not evenly across the company. So the right first step is not “adopt AI” but pick a process where repetition and measurability are highest; that is where automation pays off fastest.
“Let’s automate processes with AI” is a phrasing without an object. Which processes exactly, with what effect, where it won’t pay off — these are the real questions. Let’s go through what actually gets automated and what doesn’t.
A repeated decision gets automated, not “a process in general”.
Hypothesis: repeated decisions get automated, not “processes in general”
A good candidate for AI automation is an operation that repeats often, has a measurable output, and tolerates a controlled error. A bad one is rare, creative, or one where any error is impermissible and there’s no room for a human check. The difference is not in “AI complexity” but in the nature of the operation itself.
Impact comes not «a little everywhere» but concentrated in a few functions with repeated operations. That is the hint of which process to automate first.
Value concentrates in a few functions with repeated operations — that is the hint of where to start, not “deploy everywhere”.
Problem: people try to automate everything or the wrong thing
The typical mistake is to choose an impressive, not a payback-bearing, task: the most visible is automated instead of the most repeated. The second — trying to automate a whole process including rare exceptions and creative steps where AI gives no reliable result. The third — measuring the effect “across the company” rather than by a concrete operation, so it can be neither confirmed nor contested.
Why the usual approaches don’t work
“Make an assistant and see” doesn’t work: there’s no success criterion, so success becomes the launch itself.
“Automate the whole process” doesn’t work: a stateless scenario breaks on exceptions, and exceptions are exactly the hard half of a real flow.
“AI for AI’s sake” doesn’t work arithmetically: automating a rare or unstable operation doesn’t pay off, because there’s nothing to save on.
Engineering model: what actually gets automated
Reliably automated are classes of operations with a common nature:
- Classification and routing — requests, documents, tickets sorted by type and directed by rules.
- Extraction and reconciliation — data gathered from different systems into one consistent shape, statuses consolidated.
- Draft preparation — typical replies, documents, summaries prepared for a human to check, not instead of them.
- Monitoring and escalation — the process watches channels, notices significant signals, creates tasks and escalates by role and SLA.
Common to all four: repetition, a measurable output, and a place for human control on expensive decisions. This is not “AI replaces an employee” but removing the repetitive part of the work while keeping control.
Impact comes not from «AI in general» but in specific functions with repeated operations. Adoption pays off where there is a measurable process.
Practical takeaway for business
Choose a process by repetition and measurability, not by visibility. Name the metric before the start: it was this much time/cost — it became this much. If a metric cannot be named, the process is chosen wrong.
Automate a part, not all of it. Remove the repeated decisions, leave the expensive and rare ones to a human. The goal is to scale operations without a linear headcount increase, not to “remove people”: the second framing usually doesn’t pay off as promised.
Count the effect by operation. “This operation now costs this instead of that” is an argument for the budget; “we adopted AI” is not.
Apply this to your processes — .
Open questions
Where automation lowers decision quality, not just cost, depends on the process and is checked on the pilot. How many processes to take in parallel after the first success is a question of manageability, and the answer is usually “fewer than you’d like”. How to count the effect when the gain is removed coordination, not an explicit salary saving, is a model, not an industry standard.
Name a process where people repeat the same decisions day after day — and its payback can be estimated before the start. — we’ll work out what gets automated first and how to measure the effect.