Most automation projects that stall do so after the first process, and the cause is rarely the software. A demonstration runs the clean path. The real process has branches nobody counted before the contract was signed.
We wrote a buyer's guide to intelligent automation companies that covers the split between platform vendors and delivery partners. This post pulls out the engineering artefact that does most of the filtering: a written map of one process with every exception on it.
What the map contains
Pick one process and write it down end to end. For each branch, record how often it occurs and what should happen to it. A plain YAML file is enough:
process: supplier-invoice-intake
volume_per_month: 0 # count from last month
happy_path:
- read invoice
- match to purchase order
- post to ledger
exceptions:
- id: photo-of-printout
frequency: 0.0 # share of volume, from real cases
handling: extract, then queue for review
- id: changed-reference-format
frequency: 0.0 # share of volume, from real cases
handling: queue for review, update the rule
- id: amendment-after-approval
frequency: 0.0 # share of volume, from real cases
handling: stop, notify the approver
Fill the frequencies from a few weeks of real cases. Guesses defeat the purpose.
How to use it in selection
Send the same file to three suppliers and ask each for:
- [ ] hours to build the pilot
- [ ] how many exception paths their last comparable project ended up with
- [ ] a fixed price for a pilot on your own documents
- [ ] exception handling as a separate line in the quote
The answers diverge fast. A supplier who has delivered this before knows the exception count from a comparable job. A supplier who has only demoed it will talk about the model.
Two more checks engineers should own
Access first. An automation holds credentials into finance or operations systems, so ask how permissions are scoped and what the audit trail records for each action.
Then model licence cost at your year three volume. Per bot, per action and per document pricing behave very differently at ten times today's numbers.
Pay for the pilot. A free one runs on the supplier's data and success criteria, and a small fee buys a real result plus the right to walk away.
The full scoring list of 6 criteria is in the original article on dev-station.tech.

Top comments (0)