Task-Level Decomposition & Automation Fit

Purpose

Prevents the most common mistake in identifying AI opportunities: assessing an entire role or process at once ("could AI handle customer service?") instead of breaking it down into individual tasks, each assessed separately. AI generally doesn't replace whole job titles or entire complex processes — it replaces or assists specific cognitive and routine TASKS within a process. This skill is the input base for the ai-opportunity-portfolio skill: it produces a prioritizable raw list of classified tasks, from which the portfolio then selects and scores the best ones.

Anchored in research

• A research report supplied by the user, "Methods, Frameworks, and Competencies for Identifying AI Opportunities and Capacity in Business" (2026) — task-level atomization, the Dual Decomposition model, data-driven techniques for process/task analysis.
• Brynjolfsson, E. & Mitchell, T. (2017/2019) — the Suitable for Machine Learning (SML) index. The original model assesses tasks against nine criteria; this report explicitly names four of them (see Method point 2) — the other five are not itemized in this report's source, don't invent them, use only these four until the owner or a more precise source fills in the rest.
• Nuvepro (2026) — The Dual Decomposition Framework: People Path + Process Path.

Method

1. Choose the scope of the review. One role/team (People Path) or one value-chain/process stage (Process Path) — don't try to do both at once on the first pass. If the Process Path view is missing entirely so far, first do a rough value-chain mapping (see ../../../business-design-frameworks/skills/value-chain-mapping/SKILL.md) before this deeper task-level decomposition.
2. Decomposition line for a People Path review: Organization → Department → Role → Tasks. Decomposition line for a Process Path review: Value chain → Workflow → Tasks. Both lines converge at the task level — this is intentional: the same task often shows up from both perspectives and is worth cross-checking.
3. Gather the task list in a data-driven way, not just from interviews. Interviews systematically underestimate the amount of routine work (people forget/downplay repetitive small tasks). Use where applicable:
   • Process Mining (e.g. Celonis, UiPath Process Mining) — extracting the process's real-world executions from system logs (ERP, CRM): long lead times, rework loops, manual data-transfer steps between systems.
   • Task Mining — tracking at the level of user screens/actions: points where an expert copies information from one system to another or looks up information across multiple documents at once.
   • Cognitive friction analysis — where the employee's mental load is highest (e.g. analyzing a long document vs. the final decision based on it) — these are the points where Augment-type support is often more valuable than Automate. If none of these tools are in use, do the same analysis in a lighter form: have the employee keep a log for one day of every switch from one system to another and every point where they feel uncertainty or load.
4. Classify every task against four SML criteria (1–5 or yes/no):
   • Input and output clarity — does the task have a clearly definable digital input and output?
   • Cognitive nature — is the task based on pattern recognition, language translation, summarization, or data classification (natural fit for AI), or on physical presence, negotiation, or ethical judgment (not)?
   • Error tolerance — can the process tolerate a non-deterministic, probability-based result (e.g. a draft, a proposal), or does it require 100% deterministic accuracy (e.g. drug dosing, statutory reporting)? A quick illustrative split, useful for calibrating this judgment: billing logic and access/permissions decisions are deterministic — no error margin is acceptable, and the routing in ../../../human-ai-collaboration-design/skills/hitl-maturity-and-confidence-routing/SKILL.md should treat them accordingly; content summarization and creative ideation are probabilistic — they benefit from linguistic flexibility and contextual synthesis, and forcing deterministic precision onto them wastes the model's actual strength.
   • Time scale and response time — does the task require a split-second reaction (real-time) or deep, long-term deliberation?
5. Classify every task into one of three categories:
   • Automate — AI/an agent performs the task independently without human intervention. Typically: routine, high volume, deterministically verifiable.
   • Augment — AI acts as a human assistant/co-agent (human-in-the-loop). Typically: complex decision-making, creative drafting, expert background research, context retrieval.
   • Human-Only — stays entirely with a human. Typically: strategic judgment, high-stakes negotiation, physical presence, ethical decision-making.
6. Produce a structured task list with three columns: task / SML assessment in brief / classification (Automate/Augment/ Human-Only) + justification. This is the input to the ../ai-opportunity-portfolio/SKILL.md skill, where Automate and Augment tasks are grouped into larger opportunities and scored.
7. Watch for two systematic mistakes: (a) don't classify an entire role as "Automate" at once just because some of its tasks are — most roles are a mix of task types; (b) don't classify a task as Human-Only just because it's complex — complexity by itself doesn't rule out Augment-level AI support, it just raises the error-tolerance requirement.

What this skill does NOT do

• Doesn't assess a task's/opportunity's business value or feasibility more broadly — that's ../ai-opportunity-portfolio/SKILL.md's job. This skill only answers "does this task suit AI at all, and at what level," not "is it worth doing."
• Doesn't replace ../../../business-design-frameworks/skills/value-chain-mapping/SKILL.md for function-/process-level mapping — this skill goes one level deeper, into individual tasks within functions.
• Doesn't do technical feasibility assessment (model choice, architecture) — that's ../ai-use-case-feasibility-and-poc-scoping/SKILL.md's job at a later stage.
• Doesn't include the full nine-criterion SML model — only the four criteria the source material itemized. Don't present the other five criteria as established without a more precise source.

Refinement notes

Areas to keep deepening with real practice:

• your own rules of thumb about which task types most often surprise you (assumed Human-Only but turns out Augment-eligible, or vice versa)
• concrete examples of using process mining / task mining tools in your own engagements (into ../../references/)
• the SML model's remaining five criteria, if you find them in a more precise primary source (Brynjolfsson & Mitchell 2017/2019)

Once this section is filled in and validated in practice, update the maturity field in skills_index.json to draft, validated, or canonical (see ../../../meta/maturity_levels.md). Don't add new fields to the frontmatter — name and description are the only ones allowed (see ../../../meta/frontmatter_schema.md).

Continue from here

• Before this (if the value chain isn't mapped yet): ../../../business-design-frameworks/skills/value-chain-mapping/SKILL.md
• Next in this pack: ../ai-opportunity-portfolio/SKILL.md — groups and scores the Automate/Augment tasks this skill produces into opportunities.
• A parallel, top-down approach (a faster first-pass mapping before a detailed process description): ../ai-capability-pattern-matching/SKILL.md
• A ready-made skill chain for this situation: see ../../../playbooks/
• This pack's shared guardrails: ../../CLAUDE.md

References

• ../../references/ — the pack's shared background material
• ../../../human-ai-collaboration-design/references/hitl-partnership-heuristics-research.md — grounding for the deterministic/probabilistic worked example added to the error-tolerance criterion
• ../../CLAUDE.md — the pack's shared guardrails