How can AI improve productivity?

Where the time actually comes from

Independent research and our own client measurements point to the same clusters. Producing a first draft of any written artefact — a letter, report, proposal, specification or policy — is where the largest single saving sits, because starting is the expensive part. Summarisation is second: meetings, long email threads, research packs and lengthy contracts. Third is retrieval, finding the right document or the right paragraph inside it. Fourth is transformation: turning notes into a structured document, a document into a presentation, or unstructured text into a table.

None of these tasks is the core of a professional's job, and all of them consume a substantial share of the working week. That is exactly why AI's productivity effect is real but bounded. It does not make a solicitor advise faster or an engineer design faster; it removes the administrative envelope around expert work.

Measuring it honestly

Why some teams gain hours and others gain nothing

The difference is almost never the tool. It is whether the individual has connected the tool to a specific, frequent task they own. People who learn five prompts for the work in front of them build habit within a fortnight. People given a general demonstration try it twice, get a mediocre result because their prompt was a one-line question, and quietly stop.

Three behaviours characterise high performers: they supply context rather than asking bare questions, pasting the source material, the audience and the required format; they iterate rather than accepting the first output; and they know which tasks to hand over and which to keep. Teaching those three behaviours against real work is the entire substance of effective enablement, and it is why our Department-Specific AI Productivity Pack is built around a department's genuine documents rather than generic exercises.

Turning saved time into value

An hour saved is not automatically an hour earned. Leadership has to decide in advance what the released capacity is for: more billable work, faster turnaround as a service differentiator, absorbing growth without recruitment, clearing a backlog, or giving overloaded teams a sustainable workload. Without that decision, saved time disperses invisibly and the investment looks like it produced nothing.

In fee-earning organisations the cleanest measure is billable or productive hours per head. In operations, cycle time and throughput per FTE. In support functions, cost per transaction and backlog age. Pick the measure before the pilot and hold to it.

A realistic timeline

Expect a dip before the gain. In weeks one and two people are slower as they learn. By week four, trained users in document-heavy roles typically report between three and six hours a week returned. By the end of a quarter the pattern is stable enough to forecast. Beyond that, further gains come less from individual assistance and more from automating whole processes, which is where business process automation and agents take over.

Sources and further reading

• Experimental evidence on the productivity effects of generative AI

  Science / MIT (Noy & Zhang)

• Generative AI at Work

  National Bureau of Economic Research

• Microsoft 365 Copilot adoption guidance

  Microsoft Learn