Here is the life cycle of a typical enterprise AI pilot. Enthusiasm secures a small budget. A use case is chosen — usually the one with the most excited sponsor rather than the most expensive problem. Three months of work produce something that genuinely helps. Then the CFO asks the only question that matters: what did it save us?

And the room goes quiet. Because nobody measured what the process cost before the pilot began. There is no before, so there can be no after. The team offers anecdotes — "the team feels faster," "quality seems better" — and anecdotes do not survive a budget committee. The pilot is politely archived. Six months later, someone declares that "AI didn't work here."

The pilot didn't fail at the technology. It failed at accounting.

The cheapest habit in AI

The fix costs almost nothing, which is what makes its rarity so strange. Before any AI initiative starts, capture the baseline: how many hours the process consumes, at what loaded cost, with what error rate and cycle time. One week of honest measurement — timesheets, ticket counts, a stopwatch if necessary — is usually enough.

This is not sophisticated analytics. It is the same discipline behind any audit: establish the opening balance before you evaluate the movement. A chartered accountant would never certify a change in position without an opening position. Yet enterprises routinely fund AI projects with no opening position at all.

What a good baseline contains

  • Volume: transactions, documents or requests handled per period.
  • Effort: person-hours consumed, converted to loaded cost.
  • Quality: error rate, rework rate or exception rate — whichever the process owner already argues about.
  • Cycle time: elapsed time from trigger to completion, because speed often carries value the cost line misses.

Crucially, the baseline must be agreed with the process owner and finance before the pilot starts. A baseline reconstructed afterwards will always be suspected of flattery — usually correctly.

The second-order effect

Something else happens when baselines become mandatory: use-case selection improves automatically. When every proposal must state the measured cost of the problem it solves, the charming-but-trivial ideas fall away and the boring, expensive problems rise. The reconciliation that consumes 300 hours a month beats the flashy chatbot nobody costed. Baseline discipline doesn't just prove value after the fact — it points the programme at value before it.

So before your next pilot: one rule, adopted today, enforced without exception. No AI initiative begins without a number it intends to move. It is the least glamorous sentence in AI strategy, and in our experience, the single best predictor of whether a programme is still alive in two years.