Skip to content
arthova
InsightsField notes

Which business process should you automate first?

A selection matrix for the first automation candidate: repeatable, bounded and measurable.

Put this to work

Start with the process whose inputs, boundaries and output you can describe.

The most exciting process is rarely the best first choice

Your first automation candidate should contain a recurring decision with clear inputs, bounded exceptions and a measurable output. Anthropic recommends starting with the simplest solution that fits and adding agentic complexity only when it demonstrably improves the outcome.

Three prerequisites that cannot offset one another

Working rule

Clear ownership

One person owns the decision

Name the decision, expected output and the person who handles exceptions.

Working rule

Permitted data

Inputs are available and usable

Confirm access and permitted use before starting the test.

Working rule

Verifiable output

Correct and incorrect can be distinguished

Define the expected result and stop rule before the first run.

The selection matrix

Compare each candidate using three states: documented, unclear or missing. A “missing” entry is a work item, not a reason for a more optimistic classification.

Table – scroll sideways if needed

CriterionDocumentedUnclearMissing
Repetitionvolume and period knownoccasional casesone-off case
Inputsrequired fields existsources partly knownrequired data missing
Decisionexplicit if-then rulesrules only verbaldiscretionary decision
Exception pathstop rule and owneroccasional handoffno owner
Outputexpected output and metricpartly checkablecannot be checked
Effortbaseline measured over a short periodeffort estimatedeffort unknown

For a first test, choose a candidate where inputs, decision, exception path and output are documented. If several qualify, compare volume, effort and expected reversibility using your own data. There is no universal cutoff.

Hard stops

Stop when there is no process owner, required data is missing, a wrong output could cause unacceptable harm, or nobody owns the exception. Potential upside does not override these stops.

Two cases, two starting points

Hofmann shows a bounded enquiry flow: service, deployment details, distance, surcharges and quote. The case material records 38 enquiries in 7.5 weeks, 21 of 38 outside office hours and an automated cycle time under 60 seconds. Usage and cycle time are measured; the approximately 25-minute manual baseline is estimated. It is an example of explicit rules, not a general success benchmark.

Leopardo Negro connects four administrative workflows: invoicing, inventory, production planning and customer communication. The approximately six hours per week is a modelled release based on the owner’s estimate. It is a useful example when several small workflows are connected and capacity value is kept separate from revenue.

The first test run

Start with, for example, ten permitted, preferably anonymised cases and add missing, contradictory and uncommon inputs deliberately. Ten is a starting point, not evidence of sufficient test coverage. Write the expected output before running them. Measure completeness, time to output and share of handoffs. For variable outputs, repeat important cases with the same input. Then use the rule card before automation, the capacity calculator and the acceptance checklist.

Before estimating the benefit, record current effort with the baseline measurement log.

Sources & context

The working template is Arthova’s editorial interpretation. External guidance and our project evidence are linked separately.

  1. Anthropic: Building Effective AI Agents
See the documented project
From reading to doing

Bring one specific workflow.

We will discuss what can usefully be automated, what information is missing and how to check the outcome.

Discuss a workflow

A useful next step

Acceptance before production: A checklist for AI-supported workflowsCapacity before ROI: An honest process model