How to Find the Best AI Automation Opportunities

Teams waste months when they begin with a tool shortlist. A better sequence is to observe where work piles up, where handoffs stall, where rework is common and where people spend time searching for information that already exists in a system. AI automation creates value when those frictions involve language, documents, classification or other variable inputs that fixed rules handle poorly. If the bottleneck is missing data, unclear ownership or a broken form, fix that first. The AI automation pillar explains the operating model; this article focuses on finding the right candidates before you design or build.

Treat opportunity discovery as a short discovery programme with an owner, a timebox and a decision. Interview operators, watch live cases and pull a sample of tickets, emails, documents or queue exports. Ask what a correct completion looks like, which exceptions occur, which systems are touched and what happens when confidence is low. Without that evidence, scoring becomes theatre and the roadmap becomes a list of preferences.

List candidate workflows by trigger, volume, cycle time, active handling time, error or rework rate, systems involved, data sensitivity and business owner. Include support triage, order amendments, document intake, product data cleanup, lead qualification, reporting preparation and internal research packs. Exclude one-off projects and work that only one specialist understands unless that specialist is leaving and the knowledge risk is material.

• Trigger: what starts the work and how often it arrives.
• Outcome: what correct completion means in business terms.
• Inputs: structured fields, free text, documents, images or external lookups.
• Actions: reads, drafts, writes, messages, approvals and handoffs.
• Exceptions: missing data, policy edge cases, fraud, escalation and customer pushback.
• Baseline: volume, time, rework, waiting and known cost drivers.

Keep the inventory at the level of an owned workflow, not a department slogan. “Improve operations with AI” is not a candidate. “Classify inbound wholesale order emails, extract line items and create a draft sales order for review” is. The second statement can be scoped, estimated, evaluated and owned.

Score each candidate from zero to two on value, AI fit, data readiness, integration feasibility, exception design and risk. Value covers queue pressure, rework cost, customer impact and capacity released. AI fit covers whether variation is linguistic or document-heavy rather than purely deterministic. Data readiness covers access to timely, authorised context. Integration feasibility covers typed APIs or reliable exports. Exception design covers whether uncertain cases can route to an owned queue. Risk covers irreversibility, privacy, compliance and brand exposure.

Do not average away a zero on ownership, risk or data authority. A high total with no owner is still a poor first project. Prefer a medium-value workflow with clean boundaries over a high-value workflow with irreversible actions and weak evidence. Where agents may be involved later, compare against AI agents explained so you do not force an agent pattern onto a single model call inside a deterministic flow.

Hard filters include no named business owner, no measurable baseline, no access path to required data, no exception queue capacity and actions that cannot be reversed or compensated. Also filter out workflows whose main problem is policy ambiguity. Automating confusion produces consistent mistakes at higher speed. Clarify the policy, then revisit automation.

1. Reject “automate everything the team hates” without volume and error evidence.
2. Reject copying a competitor’s use case without matching systems and risk profile.
3. Reject open-ended research agents as the first production write-path.
4. Reject projects whose success metric is demos completed rather than correctly completed cases.
5. Reject any design that relies on the model to enforce its own permissions.

Another anti-pattern is ranking opportunities by model novelty. Classification of inbound requests may look ordinary and still release more capacity than a generative customer-facing agent. Ordinary can be the correct commercial answer. The goal is dependable throughput under control, not a showcase.

Group survivors into now, next and later. Now candidates are bounded, measurable, reversible and owned. Next candidates need data cleanup, an integration adapter or stronger review design. Later candidates may be valuable but depend on governance, identity or knowledge foundations that do not yet exist. Publish the rejected list with reasons so the same ideas do not reappear every quarter without new evidence.

For each now candidate, write a one-page brief: outcome definition, baseline metrics, in-scope actions, out-of-scope actions, data classes, systems, review points, kill switch and decision date. That brief becomes the intake for design. It also stops scope from expanding into adjacent processes during build. If the brief cannot be written, discovery is incomplete.

Review the roadmap quarterly against actual queue pressure, incident history and integration maturity. Promote or demote candidates based on evidence. Link opportunity selection to ROI discipline early; how to calculate AI automation ROI should inform which now projects clear a commercial bar. When a selected workflow must replace manual steps safely, use replacing manual workflows with AI as the implementation companion. For agent-shaped candidates, keep AI agents for business in view so permissions and evaluation are designed before tools are connected.