How to prioritize workflows for automation

Evidence

Make every automation candidate comparable before ranking it

Use one concise candidate brief. Name the triggering event, unit of work, intended outcome, person affected, accountable owner and boundaries. Record volume as a distribution across periods and segments. Measure active handling, elapsed delay, first-pass completion, corrections, abandonment and consequential errors. Describe the normal path and the largest exception families. List source and destination systems, identity or permission boundaries, current policy and known planned changes. A sentence such as automate invoice processing is too broad to compare; a defined decision and outcome can be evaluated.

Separate observations from assumptions. A timestamped event log is stronger than an estimated average from a workshop. An operator sample can still be useful if its size and bias are explicit. Give each baseline a confidence level and range. State whether value would reduce cost, increase capacity, shorten wait, improve quality, prevent loss or strengthen a control. Do not combine these into money unless the conversion is defensible. A candidate may deserve discovery because the outcome matters even when its present evidence is weak, but the uncertainty should reduce commitment rather than disappear.

Decision model

Keep hard gates separate from weighted portfolio tradeoffs

Apply gates before scoring. A workflow may be unsuitable while decision authority is unresolved, required data cannot be used, an unsafe failure cannot be controlled, policy is contradictory or no operating owner will accept the result. The right outcome can be stop, policy work, data remediation or process redesign. A numerical benefit should not compensate for a prohibited action. Document the gate, evidence, owner and condition for reconsideration so defer does not mean disappear.

For candidates that pass, assess value, readiness, delivery effort and residual risk on explicitly defined scales. Use ranges or bands and record confidence. Weighting expresses current strategy, so publish it and test sensitivity: if a small weight change completely reverses the order, the ranking is fragile. Avoid adding correlated factors several times, such as volume, hours and labor saving. Preserve the underlying measures beside the score. Decision makers should be able to see why two candidates differ without reverse-engineering a spreadsheet formula.

• Use gates for non-negotiable authority, data, safety and control conditions.
• Define every scoring band with observable evidence.
• Record uncertainty and confidence alongside the central estimate.
• Test whether reasonable weighting changes alter the shortlist.
• Keep raw evidence visible so the score remains explainable.

Portfolio

Sequence shared dependencies, change capacity and learning value

Candidates are not independent. Several may need the same master data, identity layer, document ingestion, event model, integration or human-review capability. Map these foundations and decide whether an early investment creates reusable value. Do not charge the full foundation to every candidate or hide it outside all business cases. Also identify conflicts: two projects may change the same frontline role, policy or core system in the same quarter. A theoretically valuable portfolio can fail if it exceeds training, subject-matter expertise or operational release capacity.

Sequence for learning as well as return. A bounded candidate with representative data and measurable outcomes can validate a platform assumption before a larger dependent workflow. Conversely, an easy but atypical pilot can produce false confidence. Mark prerequisites, parallel work, exclusion windows and decision dates on the roadmap. Retain some capacity for operational fixes after launch. Portfolio optimization is not simply choosing the highest-ranked candidates; it is arranging evidence, foundations and change so later decisions become cheaper and more reliable.

Governance

Use prioritization to buy evidence, then rerank the portfolio

The first priority decision should usually authorize the next evidence stage, not full deployment. A process observation can test the baseline. A data sample can expose missing fields. A technical spike can test an uncertain interface. A manual simulation can reveal review effort. A controlled pilot can measure outcome, exception demand and adoption. For each stage, state the uncertainty, spending limit, evidence, advance threshold, redesign option and stop condition before work begins. This reduces the pressure to portray every pilot as a success.

Rerank at defined portfolio reviews and after material evidence changes. Replace forecast values with observations, keep the prior estimate and explain the difference. Remove sunk-cost weight from the decision. A candidate can fall in priority because another dependency became urgent, the process changed or the expected value did not survive exceptions. It can also rise because a shared capability now exists. Publish decisions and their evidence to process owners and affected teams. Transparent reprioritization builds more trust than a permanent list whose assumptions are quietly obsolete.

• Authorize the smallest stage that answers the decisive uncertainty.
• Set advance, redesign and stop criteria before results are known.
• Replace assumptions with observations while preserving forecast error.
• Revisit rank after dependency, policy, process or capacity changes.
• Treat stopping a weak candidate as useful portfolio learning.