EXECUTIVE SUMMARY - Enterprises approach automation like cartographers: map every process, score every opportunity, produce a beautiful heatmap - and ship nothing for two quarters. The mapping isn't wrong; the sequencing is. In my automation engagements, early ROI concentrates overwhelmingly in three repeatable patterns - Document-to-Decision, Triage-and-Route, and Reconcile-and-Alert - that recur across finance, operations, and support regardless of industry. This report gives you the Pattern Map, the four-question filter for choosing your first target, and the 90-Day Ladder from candidate to measured production value. Start with one pattern, one process, one number. Compounding does the rest.
Why ocean-boiling fails
Comprehensive process-mining programs fail not analytically but politically: they consume the honeymoon period - the window in which sponsors will fund without proof - on analysis instead of evidence. By the time the heatmap ships, the appetite that commissioned it has moved on. The alternative is not recklessness; it is pattern recognition: most early wins are structurally identical, so you can skip discovering what a decade of engagements already discovered.
The three patterns
Pattern 1 - Document-to-Decision. Unstructured input (invoice, application, contract, email) → extracted fields → validation → decision or hand-off. The oldest automation pattern, transformed by LLMs' tolerance for layout variety that broke template-based OCR for years. Human-in-the-loop at the confidence threshold: high-confidence extractions flow through; the ambiguous fraction routes to a person, and that fraction is your quality dial. Where it lives: AP, claims intake, KYC, contract metadata.
Pattern 2 - Triage-and-Route. High-volume incoming items (tickets, leads, emails, alerts) → classify → prioritize → route to the right queue with context attached. The value isn't answering - it's that the right human sees the right item first with the summary already written. This is the lowest-risk pattern in the map, because a mis-route costs minutes, not money, and the fallback is simply the status quo queue. Where it lives: support, sales development, IT service desks, compliance review.
Pattern 3 - Reconcile-and-Alert. Two or more systems that should agree (bank vs. ledger, CRM vs. billing, inventory vs. orders) → compare on schedule → surface exceptions with an explanation draft. Humans stop finding discrepancies and only resolve them. Quiet, unglamorous, and reliably the fastest payback of the three, because the baseline cost - analyst hours spent hunting - is fully measurable before you start. Where it lives: month-end close, order management, data quality operations.
The 90-Day Ladder
Weeks 1–2: Baseline. Measure the current process - volume, cycle time, error rate, hours - before touching anything. The baseline is what converts your launch into a business result; skipping it is the single most common way automation value becomes unprovable.
Weeks 3–6: Shadow. Automation runs in parallel; humans still decide. You are collecting agreement-rate data, not saving time yet. Shadow mode is what earns the process owner's trust - they watch it be right before it's in charge.
Weeks 7–10: Assisted. The automation acts; humans review by exception, tuning the confidence threshold weekly. Cycle-time gains begin here.
Weeks 11–13: Production + number. Full flow with exception queue, monitoring, and a one-page before/after: baseline vs. current on the same four metrics. That page - not the demo - is what funds process number two.
The compounding move
After the first win, resist the temptation to hunt a bigger process. Instead, run the same pattern on the adjacent process - second document type, next queue, another reconciliation pair. Pattern reuse cuts delivery time roughly in half each repetition because the architecture, exception handling, and trust conversation are already built. Three same-pattern wins beat one hero project on every metric that matters, including the political ones.
FAQ
Where do AI agents fit? Agents are an execution choice inside these patterns, not a fourth pattern. Adopt agentic orchestration when a pattern needs multi-step tool use - see MCP and Agentic Workflows - not as a starting ambition.
RPA or LLM-based? Wrong axis. Choose the pattern first; then use deterministic automation for deterministic steps and LLMs where inputs vary. Most production builds are hybrids.
What if our top pain point fits none of the three? Then it's probably a workflow-redesign problem wearing an automation costume - automating a broken process just makes the brokenness faster. Fix the process first; the patterns will still be here.
Action checklist
- [ ] List candidate processes; tag each with its pattern (or "none")
- [ ] Run the four-question filter; pick exactly one first target
- [ ] Capture the two-week baseline before any build
- [ ] Ship shadow mode by week 6; publish the agreement rate internally
- [ ] Go assisted at your measured confidence threshold, not a guessed one
- [ ] Publish the one-page before/after; propose the same pattern's neighbor
Related: The Four Gates · ROI Modeling for AI
Choosing your first target now? I run this as a two-week selection-and-baseline sprint, ending with the funded 90-day plan. → Book a consultation
Working through this in your organization?
I advise enterprise teams on exactly these problems. Start with a free 30-minute consultation.
Discuss Your Challenge