Observe the work before choosing tools. Inventory recurring processes, sample real cases, measure volume, touch time, wait time, errors, rework and outcomes, then prioritise stable workflows with clear ownership, usable data, material consequence and manageable exceptions.
This is an AI automation and business process selection problem. Repetition is useful evidence, but the best opportunity is the process where improvement changes an operating or financial result.
Microsoft describes process mining as using system event data to understand real processes, identify improvement and automation opportunities, compare variants and monitor KPIs Source: Microsoft Learn.
Define measurable value
State whether the objective is lower cost, shorter cycle, more capacity, fewer errors, faster cash, better conversion, lower risk or improved service.
Avoid broad goals such as save time. Name the outcome, owner, baseline and financial mechanism.
Build a process inventory
List recurring workflows across sales, service, finance, operations, HR and marketing. Capture trigger, outcome, volume, owner, systems and customer impact.
Invite frontline staff to identify hidden spreadsheets, inboxes and manual reconciliations. Do not use job descriptions as the only source.
Observe representative cases
Shadow ordinary, peak, failed and exceptional cases. Record steps, decisions, handoffs, tools and workarounds.
People often omit small checks that make the process safe. Preserve privacy and avoid exposing customer data in observation notes.
Compare the procedure with actual work.
Measure volume and frequency
Calculate cases per day, week and season, plus growth and peak distribution. Separate eligible and exceptional volume.
A task taking five minutes can matter at high volume. A monthly process may still matter when it blocks large revenue or carries material risk.
Use ranges where records are incomplete.
Separate touch and wait time
Measure active employee effort, queue delay, external waiting and total cycle. Find where work actually accumulates.
Automating a short entry step will not shorten a week-long approval queue unless the constraint changes.
Map service-level and customer consequences.
Quantify quality and rework
Record incorrect fields, duplicate cases, missed deadlines, corrections, complaints and downstream incidents. Define what a correct case means.
Do not assume manual work is the quality baseline. Sample it and calculate uncertainty.
Map decisions and judgment
Identify rules, lookups, interpretation, negotiation and high-consequence judgment. State which decisions require accountable human review.
A repetitive interface can conceal varied judgment. Do not automate approvals merely because the form looks consistent.
Document appeal and override requirements.
Assess process stability
Check whether inputs, policies, roles and downstream systems are stable enough to automate. Frequent undocumented change increases maintenance cost.
Fix unclear ownership and contradictory procedures first. Version the intended process.
Prefer a bounded case over an organisation-wide workflow.
Assess data readiness
Inventory source formats, completeness, identifiers, permissions, history and ground truth. Test representative documents and messages.
Missing or inconsistent data may require form, master-data or integration work before automation. Do not let plausible model output replace authoritative records.
Define retention and access.
Assess integration readiness
Check supported interfaces, authentication, rate limits, idempotency, test environments, monitoring and vendor change. Identify the system of record.
Avoid automation that depends on brittle screen clicks when a stable integration or process change exists. Plan partial-failure recovery.
Calculate integration ownership.
Score risk and consequence
Assess privacy, security, financial, legal, safety, customer and reputation harm from false action or delay. Choose controls proportionate to consequence.
NIST's AI RMF calls for context-relevant measurement and prioritising risks by impact and likelihood Source: NIST.
Exclude unsafe autonomous outcomes from the first pilot.
Estimate addressable value
Calculate eligible volume multiplied by current cost or outcome gap, then reduce for exceptions, adoption and remaining human work. Add avoided delay and errors where defensible.
Distinguish capacity from budget savings. Include implementation, review, infrastructure and support costs.
Use low, base and high ranges.
Compare alternatives
Test whether removing a step, changing a form, clarifying policy, integrating systems or reallocating ownership solves the problem more simply.
Automation should compete with the best practical process change, not only the current broken workflow.
NIST includes viable non-AI alternatives in its manage considerations Source: NIST.
Build an opportunity scorecard
Score value, volume, stability, data, integration, exception rate, risk, ownership and time to learn. Weight factors to the organisation's priorities.
Do not let a numeric score hide a critical safety veto. Keep evidence and confidence beside each value.
Review with finance and frontline owners.
Choose the pilot boundary
Select one trigger, case type, team and outcome. Define eligible and excluded cases, manual fallback and stop conditions.
Use draft, recommendation or human-approved automation before autonomous writes when consequence warrants it.
Preserve the current process for comparison.
Instrument before automating
Capture baseline volume, time, quality, outcome and cost before release. Define event and case identifiers across systems.
Without baseline and reconciliation, a faster interface can be reported as value without changing the process.
Protect sensitive content in telemetry.
Measure realised results
Track eligible volume, adoption, automation rate, review, correction, false acceptance, cycle time, customer outcome, capacity and net financial value.
Allow downstream outcomes to mature. Compare with a control or stable baseline and record concurrent changes.
Report theoretical and realised benefits separately.
Test exception economics
Estimate how often exceptions occur, who resolves them, how long they take and what happens when they wait. A workflow with an impressive straight-through rate can still cost more if the remaining cases require rare specialist effort. Include escalation, customer communication and investigation in the operating model. Improve upstream inputs when one preventable exception class dominates.
Check change frequency
Review how often policies, forms, suppliers, products and systems changed during the last year. Identify who announces changes and how rules are updated. High change does not prohibit automation, but it requires versioned configuration, representative regression tests and maintenance capacity. Include that ongoing work in the ROI instead of treating the first release as the final cost.
My take: automate the constraint, not the keystroke
My take is that the best opportunity removes or strengthens the step limiting a valuable outcome. Automating visible typing while leaving the queue, decision or poor source data unchanged creates activity without value.
Start from the constraint.
Create a quarterly opportunity review
Review the inventory, new evidence, changed risk and realised value of live automations. Stop candidates whose assumptions no longer hold.
Feed exception and incident data into future selection. Keep a small pipeline of discovery work rather than launching many pilots at once.
Assign executive and operational owners.
Frequently asked questions
Is high volume required for automation?
No. A lower-volume process may justify automation when delay, risk or value per case is material.
Should we start with the easiest workflow?
Start with a bounded workflow that can prove meaningful value and controls, not difficulty alone.
How do we value saved time?
Identify whether time becomes lower cost, more capacity or a better outcome. Do not treat every minute as realised cash.
When is AI unnecessary?
Use rules, forms or direct integration when inputs and decisions are stable. AI should address genuine variation or interpretation.
Choose automation from measurable evidence
If your business needs a governed automation opportunity portfolio, get in touch for AI automation.
Related guides
For further background, read our ai crm integration and sales marketing ai workflows. The ai crm integration glossary explains the terminology used in those guides.