Two years into mainstream AI adoption, a pattern has settled in. Some organisations have quietly rebuilt parts of their work around AI and would not go back. Many more have licences, a policy written in a hurry, a pilot nobody mentions and a vague sense of having missed something.
The difference is rarely technical. It comes down to where AI was pointed.
The work AI reliably improves
Across sectors, the dependable gains cluster in a small number of task shapes.
High-volume language work with a checkable output. Summarising documents, drafting routine correspondence, converting notes into structured records, producing first drafts that a knowledgeable person will review. The gain is real because the task is frequent, the output is verifiable and a human stays accountable for the result.
Retrieval from the organisation's own knowledge. Most organisations answer the same internal questions endlessly: what the policy says, how the process works, what was decided last time. An assistant grounded in your actual documents converts that from interruptions into seconds. The prerequisite, often overlooked, is that the underlying knowledge must exist somewhere and be roughly current.
Structured transformation between formats. Meeting audio into minutes and actions. A regulation into a compliance checklist. A spreadsheet into a readable summary. These tasks have clear inputs, clear outputs and tolerant error modes, which is exactly where current models are strongest.
Analysis a person would do if they had time. Reading every submission rather than a sample. Comparing this quarter's incident reports against last year's. Checking a draft against a style guide. AI does not do this better than your best person on their best day. It does it on the days nobody had time, which in most organisations is most days.
The work AI does not improve
Decisions that need accountability. Anything where a person must be able to explain and own the outcome: hiring, funding allocations, clinical and safety judgements, decisions about individual people's entitlements. AI can prepare material for these decisions. Moving the decision itself into a model does not remove the accountability; it just makes it harder to locate.
Work whose value is the relationship. Stakeholder engagement, difficult conversations, community consultation. An AI-drafted response to a community's concerns is often worse than no response, because people can tell, and what they learn is that you did not consider them worth a person's time.
Broken processes. This is the expensive one. Automating a process that produces the wrong thing produces the wrong thing faster. A surprising share of AI readiness work turns out to be process work in disguise: before the question "can AI do this?" comes "should anyone be doing this, and this way?"
Anywhere the error cost exceeds the review cost. The practical economics of current AI are simple: it is valuable where checking the output is cheaper than producing it, and dangerous where checking is as expensive as doing the work, because under deadline pressure the checking stops.
A test you can run on any proposal
When an AI opportunity is proposed in your organisation, four questions will sort most of them.
- Frequency. Does this task happen often enough for improvement to matter? Automating a quarterly task rarely pays for its own upkeep.
- Verifiability. Can a competent person check the output quickly? If not, who bears the error?
- Accountability. Does someone need to own this outcome personally? If so, AI assists; it does not decide.
- Process health. Would we want this process at ten times the speed? If the honest answer is no, fix the process first.
A proposal that passes all four is worth piloting. A proposal that fails two or more is usually novelty wearing a business case.
Start with the work, not the tool
The organisations that get this right run the sequence in one particular order. They map how work actually moves. They find the friction that is frequent, verifiable and unloved. They pilot narrowly, with the people who do the work, and they write down what the pilot is supposed to prove before it starts. Then, and only then, they buy things.
The sequence matters because it protects the scarcest resource in any adoption effort, which is not budget. It is the workforce's willingness to try the next thing. Every tool that arrives without a purpose spends a little more of it.
AI will keep improving. The discipline of knowing your own work well enough to point it correctly will not go out of date.