What should a business automate first?

The queue that scores highest on three factors multiplied together: volume (hours per week), structure (how rule-describable the work is), and stability (how often the process changes). High-volume, rule-heavy, stable queues repay automation fastest; everything else waits its turn.

Should you automate rules or judgment first?

Rules. Deterministic steps go to boring automation: cheap, debuggable, reliable. Judgment steps (reading, classifying, drafting) come second, with AI, evals, and escalation paths. Automating judgment before the rules around it are solid builds an expensive layer on sand.

What should never be automated first?

Broken processes (automation calcifies them), low-volume tasks (nothing to repay the build), and decisions with high, unreviewable failure costs. If two teams disagree about how the workflow works, fix that disagreement before any tool encodes it.

What to automate first: a scoring method for real queues

The short answer

Rank every queue by three factors multiplied: volume (hours per week), structure (how rule-describable the work is), and stability (how rarely the process changes). Automate the top scorer, rules before judgment, with humans keeping the calls that matter. The traps are automating broken processes, low-volume irritations, and unreviewable decisions: all three feel urgent and repay nothing.

Every operation has a dozen candidates and budget for one. The choice usually goes to whatever annoyed a director most recently, which is how companies end up automating a monthly report while the support queue eats two salaries. A scoring method beats a mood.

The three-factor score

–Volume. Hours per week, honestly counted, including the shadow work around the task: the retyping, the chasing, the checking. Below a few hours weekly, stop scoring; there is nothing to win back.
–Structure. How much of the work follows describable rules? Score high when an experienced person could write the steps down; low when every case is a fresh negotiation. Judgment at the edges is fine, and is what AI extends automation into; judgment at the core is a hiring question, not an automation one.
–Stability. How often does the process change? A queue that mutates monthly turns any automation into a maintenance subscription. Stable beats exciting.

Multiply, do not average: a zero in any factor zeroes the queue. The multiplication is the discipline; averaging is how the director's pet report sneaks back in.

The decision path

The automate-first decision path · score the queue, then route each step to the cheapest tool that survives it

The traps that invert the score

–Automating a disagreement. If sales and ops describe the queue differently, any automation encodes one team's version and starts a war. Process first, tools second.
–Automating the annoyance. The task that irritates leadership is rarely the one eating hours. The score exists to defend against seniority-weighted anecdotes.
–Automating the unreviewable. Steps whose failures are costly and hard to catch (money out the door, commitments made) keep a human by design, whatever the volume says.
–Automating around a spreadsheet. If the queue runs on an outgrown sheet, automating its edges preserves its problems; the sheet itself is the project.

Multiply, never average: a zero in volume, structure, or stability zeroes the queue.

The first week

01List the queues. Support, intake, invoicing, scheduling, reporting: everywhere hours go into repeatable work.
02Score each on the three factors. Rough numbers with the person who runs the queue. An hour of honest counting beats a month of opinions.
03Take the winner through the path. Process agreed, volume real, rules to boring tools, judgment steps to an AI layer with evals and escalation.
04Define done before building. Hours back, error rate, exception route: written down, so the automation has a scoreboard.

The scoring session is precisely what a Workflow Audit is: thirty minutes, one queue mapped with its owner, and a verdict that is sometimes "automate nothing yet." If the winner needs no-code, use no-code, and know the graduation signs for later.