What business processes can AI actually automate?

The test is simpler than most articles make it: if a process involves a person reading something, applying rules or judgment that could be written down, and then entering or moving information somewhere else, the current generation of AI can probably automate most of it. That covers a remarkable share of the administrative work inside a typical business, and it is a much wider net than the chatbots and email drafting most people still associate with AI.

The clearest way to answer this question is with real examples, so here are four from our own client work, spanning very different industries with an identical underlying pattern.

Four real examples

A building compliance consultancy had qualified assessors spending 90 minutes per job reading building plans and entering specifications into their job system. We built a platform where AI reads the plans, including the diagrams rather than just the text, pre-populates around 40 fields, shows a confidence score against each one, and flags anything uncertain for human review. Each job now takes three minutes, and the assessors spend their time on assessment rather than typing.

A regional vocational college had an admin team manually shepherding every enrolment from a web form into their CRM and then into their student management system. We connected the three so an enrolment flows end to end automatically. They scaled from 200 to 2,000 students without adding a single administrative role.

A construction company tracked contractor compliance by hand: licences checked manually, induction videos chased by email, paperwork filed by a person. Their onboarding workflow now runs itself, checking documents against requirements, following up on gaps, and notifying the hiring manager only when a contractor is ready or something needs a judgment call.

A management consulting firm had a methodology that lived in the partners' heads and could only be delivered in person. We helped them build a guided application that walks clients through the process and produces the diagnostic outputs the consultants used to write manually.

The pattern underneath

Notice what these have in common. None of them is a chatbot. Every one of them is a workflow, a sequence of reading, deciding, updating and notifying that previously required a person as the connective tissue between systems. That is the category to look for in your own business, and the practical way to find candidates is to ask where a person regularly touches information that already exists somewhere and moves it somewhere else.

What AI should not automate

Just as useful is knowing what AI should not automate. Anything requiring genuine relationship judgment, negotiation, or accountability for a consequential decision should keep a human in charge, with AI preparing the groundwork. The best implementations are explicit about this boundary: the system does the volume, flags its own uncertainty, and hands the exceptions to a person.

Questions we get asked

How do I work out which process to start with? Pick the one that costs the most or produces the most errors, and make sure it is bounded enough to finish. Mapping where the time actually goes in your business is the most useful first step.

Does our data need to be in good shape first? Less than you would think. Messy inputs are precisely what AI handles better than the rule-based automation tools that came before it.

Can it work with old or unusual software? Usually yes, though older systems add integration effort, which is one of the main things that moves a project up the price range.