What has to be agreed before a process is automated

An automated process does not improve the data it runs on. It applies the same rules faster, more often and with fewer people looking at the result. If two systems disagree about which customer record is current, a person doing the work by hand can see both and stop. A rule cannot. It applies whatever tie-break it was given, and that choice carries into every invoice, report and notification downstream until somebody reconciles it by hand.

Manual effort absorbs data problems. Automation exposes them, at volume.

Different totals are a definitions problem

When finance and operations report different revenue figures for the same month, a new reporting tool will produce a third figure rather than reconcile the two. Ask each team what it counted, then compare the answers: the event that triggers recognition (invoice raised, job completed, payment received), whether intercompany work is included or netted off, how a cancelled and rebooked job is counted, and which system is treated as the source.

Both figures can be correct against their own definition. The disagreement survives because neither definition is written down. Writing it is the work: what counts, from which system, at which point in the process, excluding what, and who decides when the rule changes. That definition is the deliverable. The tool sits downstream of it.

The same applies to the records themselves. A supplier entered three times under three spellings appears as three suppliers in every report that groups spend by supplier. Correcting it in the report rather than the ledger means correcting it again next month.

Four things to settle before the build

None of the four is technical.

• Ownership. Every field the automation reads has one named person accountable for its accuracy. Not a team. A person. If nobody will accept that, the field is not ready to be a trigger.
• Definitions. Each measure has a written rule, agreed by the people who use it to make decisions rather than by whoever built the report.
• Exception handling. Rules cover the ordinary case. Decide in advance what happens to the record that does not fit: who is notified, where it queues, how long before it escalates. An automation with no exception route will either stop or apply a default nobody chose.
• The audit trail. Every automated step should record what it did, to which record, under which version of the rule, and when. Without that, the first time an output is challenged, nobody can answer.

Where automation genuinely pays

The strongest candidates are narrow, high volume and dull: re-keying between two systems with no interface, matching invoices to purchase orders, routing scanned documents by type, assembling the same report every Monday. They have stable rules, a single correct answer and a manual cost somebody can measure.

Weak candidates are the reverse: low volume, contested rules, judgement in the middle. Automating a monthly task that takes two hours saves twenty-four hours a year, set against something that has to be maintained, retested and re-pointed every time a source system changes.

Ask any supplier proposing automation three questions. Which fields does it read. Who owns them. What does it do with a record that fails validation. An answer about the platform rather than the data means the sequence is wrong.