Which business processes should you automate with AI first?

Filter one: volume × repetition

Automation earns its keep per execution. A process that runs four hundred times a day — invoice intake, ticket triage, order-status queries, document classification — pays back engineering effort quickly even when each run only saves two minutes. A process that runs monthly almost never justifies a bespoke system, no matter how painful the monthly occurrence is.

The practical test: if you multiplied the time saved per run by the runs per week, would anyone notice the reclaimed hours? If the answer is not an obvious yes, keep scanning.

Filter two: the rules are describable

Ask the person who does the task to explain how they decide. If they can narrate the decision — 'if the invoice total doesn't match the PO, I check the delivery note; if it still doesn't match, I escalate' — the process is automatable today, with the language model handling the messy inputs and deterministic rules handling the decisions.

If their explanation is mostly taste and tenure — 'after a while you just know' — you want assistive AI that drafts and recommends, with the human still deciding. That distinction, automation versus assistance, is the single most common thing teams get wrong.

Filter three: the data already exists

The best first projects consume data your systems already produce: emails, PDFs, form submissions, CRM records, support transcripts. Projects that begin with 'first we need six months of new labelled data' are research, not automation — worthwhile sometimes, but not as a first bet.

Filter four: failure is cheap and visible

Every automated system will be wrong sometimes. Pick a first process where a mistake is cheap to catch and cheap to correct — a mis-routed ticket, a draft reply a human reviews — rather than one where an error ships money or medical advice. You want the failure mode to be 'a human fixed it in thirty seconds', not 'legal is on the call'.

This is also why we build review queues into early deployments: the system does the work, humans approve the low-confidence tail, and the approval data continuously sharpens the system.

Processes that usually pass all four filters

• Inbound email and document intake — classify, extract, route, file.
• First-line support triage — categorise, draft, escalate with context attached.
• Data entry between systems that don't share an API — the invisible re-keying tax.
• Quote and proposal drafting from a price book and a discovery call transcript.
• Compliance and QA checks that are rule-heavy but currently eyeball-driven.
• Report assembly — the Monday-morning deck someone builds from four dashboards.

Start with one, instrument everything

The right first project is deliberately narrow: one process, one team, measurable before-and-after. Instrument it from day one — items processed, touch rate, error rate, minutes saved — because the numbers from the first automation are what buy the mandate for the second.

If you want a second pair of eyes on your list of candidates, that is literally what we do: we turn AI ideas into production-grade systems, and the first conversation is free.