Why is return on AI so hard to pin down?
Because most firms try to measure the wrong thing. They ask what the AI licences returned, which is a question about a tool spread thinly across many people and many tasks. There is no before, no after and no single job to point at. The answer is a shrug or an anecdote.
Measure a process instead. A single job, such as producing the monthly client report or assembling a fee proposal, has a clear start and end, a known frequency and people who can tell you how long it takes. Change it, and you can measure the difference. Everything below assumes you are measuring one process at a time.
What should you record before the change?
The baseline is the part firms skip, and it is the part that makes everything else credible. Record these for the chosen process, over a few normal weeks if you can:
- Time per instance. How long one report, proposal or update takes, from starting to gather inputs to sending. Include review time, not just drafting.
- Volume. How many instances happen in a week or a month.
- Who does it. The seniority of the people involved, and roughly how the time splits between them.
- Turnaround. Elapsed days from trigger to delivery, which is often what the client notices.
- Rework. How often a draft comes back from review with corrections, and how much work those corrections take.
A short diary kept by a few people, checked against timesheets and file dates, is enough. The method for turning weekly hours into an annual figure is set out in calculating unbillable hours.
What should you measure afterwards?
The same five things, taken the same way, so the comparison is fair. Add two more that only exist after the change:
- Adoption. How many of the period's instances went through the new process, and how many were still done the old way. A process used for some jobs and not others delivers only part of its value.
- Review findings. What the reviewer corrects in the new drafts. Falling corrections show the process is learning the firm's standards; recurring corrections show where it needs adjusting.
How do you turn saved hours into value?
This is where honest measurement differs from optimistic measurement. Saved hours are only worth money when they go somewhere. Work out where, and value each route differently:
- Billed. Freed hours spent on chargeable work are worth the charge-out rate you actually realise, not the headline rate.
- New work won. Senior time moved to proposals and client development shows up in work won, though it takes longer to see.
- A hire avoided. If the team absorbs more work without adding a person, the value is the cost of the hire that did not happen. Automation or hiring looks at that choice directly.
- Overtime and pressure reduced. Evenings no longer spent on reporting do not appear on the profit and loss, but they affect retention, and losing an experienced fee earner is expensive.
The capacity calculator helps put a figure on freed time. Be conservative: count only the routes you can see evidence for.
How do you work out the return itself?
Add up the value of the redeployed hours over a year. Set against it what the change cost: the fee for the rebuild, any running costs, and the internal time spent on mapping, testing and training. With a fixed fee, the cost side is known in advance, which makes the calculation far cleaner than with open-ended spend. How fees are set is covered in how AI consultancy is priced.
Because a rebuilt process keeps running, the value recurs every year while most of the cost falls once. That is why the first year understates the return. It is also why not acting has a cost of its own, explored in doing nothing or acting now.
Which figures flatter and should be avoided?
- Logins and prompts. Activity is not outcome.
- Staff sentiment surveys on their own. Useful context, but people can like a tool that saves them nothing.
- Theoretical time saved. Multiplying a demonstration's speed by every task in the firm produces a large figure that never appears in practice.
- Every improvement credited to AI. If the firm also changed a template or a team, some of the gain belongs there.
- Headline charge-out rates. Use what you realise.
When should you report it, and to whom?
Take readings at three points: the baseline, the end of the first few weeks live, and after the new way has settled. Report to whoever approved the spend, in the same terms as the baseline: hours per instance, volume, turnaround, rework and where the freed time went. A short, like-for-like comparison is more persuasive than any projection, and it is the strongest case for choosing the next process. To start with an estimate of where your hours go now, take the audit.