A routine task automation matrix turns a vague backlog into a decision sheet. It does not rank departments, vendors, or attractive demonstrations. It compares one observable task per row and asks whether repeating the task creates enough opportunity to test, whether an error can cause material harm, and whether the action can be undone cleanly. The result is not a forecast of savings. It is a defensible choice between a bounded pilot, preparation work, and continued manual handling.
Build one row per task, not per department
Start after the workflow boundary is known. If the trigger, owner, authoritative record, and accepted output still need to be mapped, use the broader guide to choosing the first operational workflow. This worksheet assumes that boundary exists. Its job is narrower: compare specific repeated actions inside known work without reopening the entire process-design question.
Give every row the same fields: task name, trigger, input source, system of record, current action, output, recipient, exception owner, and acceptance evidence. A row called “handle documents” is too broad. “Extract proposed invoice fields for review” is usable because the source, action, reviewer, and evidence can be named. Keep separate rows for extraction, approval, posting, and notification when they have different permissions or consequences.
Then assign categorical bands. Frequency is occasional, repeated, or continuous enough to justify a maintained route. Risk is contained, review-required, or high-consequence based on what a wrong action could change. Reversibility is easy, recoverable with a documented manual step, or difficult because an external message, irreversible transaction, or lost record cannot simply be recalled. Add two readiness checks beside those scores: inputs are stable enough to validate, and acceptance evidence is available to a named owner.
Apply blocking gates before comparing rows. Hold a task when the source of truth is unknown, the system would need broader access than the task requires, no one owns exceptions, or completion cannot be verified. Keep a high-consequence and difficult-to-reverse action manual unless a review occurs before execution. Label a row pilot-ready only when it repeats, its inputs can be checked, risk is contained or intercepted before action, and rollback is credible. Mixed rows receive prepare first, with the missing condition written as work rather than hidden inside an average score.
Use the same worksheet on different candidates
A status-summary row may be repeated and easy to reverse while it remains a draft. Its input stability depends on whether approved project records are current. The system can assemble supported statements and mark missing updates, but a reviewer accepts the summary before distribution. If a sentence cannot resolve to a current record, the row stops instead of turning an old comment into new progress. The accepted artifact is the reviewed summary plus its source references and correction record.
Document intake has a different profile. Proposed field extraction can be frequent, contained, and reversible when the original file stays authoritative and no posting occurs automatically. Missing required fields, conflicting values, an unsupported document type, or a failed destination write are stop conditions. The owner sees the source, proposed values, validation results, and intended destination. Approval of that packet is evidence; a polished extraction alone is not.
Request triage often scores well on frequency but poorly on input consistency. The safe row may classify the request and prepare a route, while identity, entitlement, priority, and consequential updates remain outside its permissions. Unknown categories and conflicting records go to a named queue with the original request and retrieved policy. This row can become pilot-ready even when automatic fulfillment must stay manual. The matrix prevents the useful low-authority step from inheriting the risk of the whole request lifecycle.
Turn the selected row into a controlled test
Write a small operating contract for the winning row: permitted source, fields read, fields written, review point, acceptance evidence, stop conditions, correction owner, and rollback method. Grant only the access required for that row. The three tests before integrating AI help verify that the problem, evaluation method, and control boundary are real before the system touches a working record.
Measure the same observations before and during the pilot. Separate active handling, waiting, review, rework, and exception recovery. Count work only after the owner accepts the result. A rejected draft is review effort, not time saved, and a queue waiting on another team is not work completed by the automation. When summaries or metrics are generated, use the checks for agent-generated metrics so unsupported figures do not enter the decision record.
Re-score after corrections. A row can move from prepare first to pilot-ready when inputs become stable or a review gate is added. It can move back when a source changes, exception volume grows, or rollback no longer covers the action. Expand one permission or task variant at a time, rerun the acceptance set, and keep the manual queue available until the new boundary is proven.
Frequently Asked Questions
Use one task per row and record its trigger, authoritative input, action, output, recipient, exception owner, acceptance evidence, frequency, risk, and reversibility.
No. Keep frequency, risk, reversibility, and readiness gates visible separately so a high-frequency task cannot hide severe consequences or a missing rollback route.
The task repeats, inputs can be validated, permissions are narrow, errors are contained or reviewed before action, completion is verifiable, and rollback is practical.
Keep it manual when authority is unclear, acceptance cannot be verified, errors have high consequences, or the action is difficult to reverse without a reliable pre-execution review.
If you want to score candidate tasks, define their controls, and build a reviewable pilot around your current systems, discuss AI automation with AI4SALE.