Deciding whether an agent is worth building does not require a pilot. It requires three numbers you already have, and arithmetic anyone in the room can check.
Two activities dominate knowledge work and both are mechanical: looking things up, and moving the result into another system.
Looking things up
| Input (replace with yours) | Example |
|---|---|
| Lookups per person per day | 20 |
| Minutes each, including finding the right source | 8 |
| Time spent searching, per person per day | 2 h 40 m |
| Per month, 22 working days | ≈ 59 h |
Not all of that is recoverable. What an agent removes is the retrieval — part numbers, stock, historical pricing, discount rules, spec cross-references — the parts with a definite answer that already exists somewhere in your own material. Judgement stays where it is.
| Now | With retrieval delegated | |
|---|---|---|
| Minutes per lookup | 8 | ≈ 2 |
| Per person per day | 2 h 40 m | 40 m |
| Recovered | ≈ 2 h/day |
That is roughly a 75% reduction on the lookup step. The denominator is the lookup, not the job: deciding, negotiating and checking are all still there.
Moving results between systems
The second activity is the one nobody counts, because it never appears as a task: switch window, copy, paste, reformat, save.
| Input (replace with yours) | Example |
|---|---|
| Results moved between systems per day | 20 |
| Minutes each | 2 |
| Per day | 40 m |
| Per month | ≈ 15 h |
When agents can call your APIs and write back directly, this step goes to approximately zero — the result lands where it belongs instead of being carried there. Fifteen hours a month, per person, producing nothing.
Why we do not pre-fill these numbers
Because "a lookup" means something different in every industry. A machine shop looks up part numbers and stock; a contractor looks up unit rates and past tenders; a distributor looks up factory discount tiers and lead times.
Different things to look up means a different compressible fraction, and that is not knowable without walking through the actual workflow. So every cell above is labelled as an assumption. Put your own three numbers in and the answer appears; if it comes out below half an hour a day, the case is weak and we will say so.
The number that is not in the table
Response time changes behaviour in a way hours do not capture. An answer that arrives in about ten seconds is one the person waits for; an answer that takes two minutes is one they stop asking for.
That is also a design parameter rather than a fixed property — in our own routing layer the same question went from 12.7–17.3 seconds to 7.9. If the workload is latency-sensitive, it can be specified up front and built for.
