Uipath Automation Discovery

Automation Discovery

Investigate how employees actually work, then identify and prioritize internal automation opportunities backed by real behavioral evidence. Produces a UiPath-ready backlog with recommended implementation paths.

When to Use This Skill

• User asks to discover automation opportunities across their organization — before any specific automation project exists
• User wants to find manual work to automate and build a UiPath implementation backlog
• User asks "what should we automate?" while working with UiPath tools
• User wants an internal automation audit to feed into UiPath Automation Hub or a UiPath pipeline
• User asks to estimate / size / cost discovered opportunities — pack-hours, delivery effort, complexity bands, contingency
• User explicitly invokes /uipath-automation-discovery

Critical Rules

1. Authorization and privacy first. Confirm the requester is authorized to analyze the selected systems and employee data. Avoid private channels, DMs, and special-category HR data (payroll, performance reviews) unless explicitly approved. Pseudonymize SPOFs by default (e.g., "Sales Ops Lead A"); use real names only when explicitly authorized. Maintain consistent pseudonyms across the entire report — assign each individual a stable label on first mention and reuse it throughout. Ask about jurisdiction constraints (GDPR, works council, internal policy); apply the stricter rule when uncertain.
2. Never assume — always ask first. Complete the full intake (Phase 0) before mining. You need company context, tool access, org structure, privacy scope, and scope agreement.
3. Verify access before mining. Test each data source with a minimal read-only operation. If access fails, note it and move on — don't block discovery.
4. Evidence over opinion. Every opportunity (Tiers 1-3) must cite a specific source, quantitative metric, and affected role or team. No unsupported claims. If a source yields fewer than 5 signals, mark all findings from that source as low-confidence. Do not promote low-confidence findings above Tier 3, except per Rule 5.
5. Replication is always Tier 1. A proven model backed by a working automation that could replicate elsewhere is the highest-value finding — this overrides Rule 4's Tier 3 cap. If the replicable model's source has fewer than 5 signals, classify as Tier 1 with a low-confidence flag until corroborated by a second source. Always lead with replicable models. Never skip the replicable-model search (Phase 2C).
6. Never invent pack-hours or complexity thresholds (Phase 4.5). When estimating build effort, the band→hours numbers and matrix thresholds come from the user-supplied Core RPA / Agentic complexity matrices and Pack-Hours catalogue. Do NOT recall or fabricate them — if they are not supplied, STOP and ask. Adjustment-factor and contingency percentages are [CALIBRATE] defaults the user confirms against actuals. Fabricated numbers recreate the estimation error this phase exists to prevent.

Workflow Overview

Phase 0: INTAKE    → Gather context, verify access, agree on scope and privacy
Phase 1: MINE      → Gather raw data from all verified sources
Phase 2: ANALYZE   → Extract patterns, SPOFs, replicable models, gaps
Phase 3: REFLECT   → Layer on business strategy for strategic gaps
Phase 4: REPORT    → Produce prioritized report with 4 tiers
Phase 4.5: ESTIMATE→ Size build effort (band → pack-hours → contingency) — on request
Phase 5: HANDOFF   → Map opportunities to UiPath implementation skills

Stop conditions: Quick scan caps at 10 findings, Standard at 25, Deep dive at 35. Max 2 retries per failed source. Phase 1 timeboxed at 3 hours for deep dives. See references/intake-guide.md §0G for details.

Phase 0: INTAKE (interactive)

Build a complete picture before mining. Ask — don't assume.

See references/intake-guide.md for detailed steps covering company context, tool inventory, access verification, org structure, output preferences, user hypotheses, scope control, and privacy authorization.

Key outputs from intake:

• Company context and department list
• Tool & system inventory with verified access
• Agreed scope (quick scan / standard / deep dive) with finding caps
• Privacy scope (pseudonymize by default, jurisdiction constraints)

Phase 1: MINE

Cast a wide net. Prioritize by signal density. Use parallel agents (Agent tool with subagent_type: general-purpose, one agent per source category).

See references/mining-guide.md for detailed per-source guidance on what to look for and how to search.

Source priority when time is limited:

1. Messaging help channels — highest signal, fastest to mine
2. Email patterns — reveals hidden recurring work
3. CRM/ERP — reveals structured process bottlenecks
4. Wiki/docs — reveals existing automation landscape
5. Issue tracker — reveals service desk patterns
6. HRIS — reveals people-process friction
7. Web research — reveals strategic gaps

Note: Web research (priority 7) feeds Phase 3 strategic analysis. Even under time pressure, do a brief web search for the company's public financials and strategy — this takes minutes and enables Tier 4 findings.

Work with whatever access is verified. Even messaging channels alone can yield 15+ opportunities. Each additional source adds depth, not changes the methodology.

Checkpoint: After Phase 1, share a raw signal summary with the user: "I found X help channels, Y existing automation projects, Z departments. Want me to go deeper on anything before I analyze?" If the user requests deeper mining, run at most 1 additional targeted pass, then proceed.

Phase 2: ANALYZE

Transform raw data into structured findings.

2A. Behavioral Patterns

Per department, answer: What's manual? What questions repeat? What approvals stall? What reports are compiled by hand? What data is swivel-chaired between systems? What handoffs break? What scheduled tasks are done by humans?

2B. Single Points of Failure

Identify roles that are sole responders. If they're out, the process stops. Pseudonymize by default — use role/team labels unless naming is authorized.

| Role/Pseudonym | System/Channel | Function | Risk |

These are the highest-urgency targets.

2C. Proven Replicable Models

The most important finding. Look for automation already working in one area that could replicate to others:

• Bot in one channel but not others
• Auto-routing in one team but manual elsewhere
• Dashboard auto-generated for one dept but compiled by hand for another

| Working Model | Where It's Missing | Addressable Volume |

Greenfield case: If no existing automations are found (nothing to replicate), Tier 1 will be empty. Promote the highest-volume Tier 2 finding to the headline slot and note that the company has no proven models to replicate yet.

2D. Department Coverage Map

| Department | Existing Automations | Key Gap |

Flag ZERO-coverage departments as biggest blind spots.

2E. Process Deep Reads

For promising existing projects, extract: pain point, manual process today, volume/frequency, ROI if documented, systems involved, dev status.

Apply low-confidence handling per Critical Rule 4.

Checkpoint: Share analysis summary with user before reflecting: "Here are the top patterns, SPOFs, and replicable models. Anything surprise you? Anything I should investigate further?" If the user requests deeper analysis, run at most 1 additional targeted pass, then proceed to Phase 3.

Phase 3: REFLECT

Identify gaps behavioral data won't reveal.

3A. Business Context

Research via web search, investor docs, or internal strategy pages: revenue, growth, strategic priorities, competitive challenges, key metrics.