---
name: ai-opportunity-triage
description: "Triage one proposed AI use case into a go-deeper or hold recommendation by scoring bottleneck severity, value capture, integration readiness, and risk posture, then ranking the follow-up questions that would most change the decision. Use when someone asks whether an AI idea is worth pursuing, which of several AI proposals deserves discovery time, or what evidence is missing before committing further effort. Not for organization-wide readiness reviews or production rollout planning."
---

# AI Opportunity Triage

Turn one proposed AI use case into a structured triage decision: a scored assessment,
a go-deeper or hold call, and a ranked list of the follow-up questions that matter most.
The output is a handoff-ready artifact, not a conversation summary.

This workflow is published by Sophon Consulting. It requires no Sophon tools,
services, or credentials.

Scope boundaries: for an organization-wide readiness review use ai-readiness-assessment;
for rollout planning of an existing pilot use ai-pilot-to-production-plan; for review-gate
design on a live system use human-in-the-loop-governance.

## Required inputs

Request these as one batch, not one at a time:

1. Workflow description: who does the work today, step by step, and where it queues.
2. Volume and unit economics: how often the workflow runs and what a unit of throughput
   is worth in time, money, or risk avoided.
3. Systems touched: where inputs live, where outputs go, and what integration points exist.
4. Constraints: regulatory, security, or contractual requirements that bind the workflow.
5. Decision context: what decision this triage feeds, who owns it, and when it is needed.

Handling missing inputs:

- If the workflow description or the decision context is missing, stop and request it.
  Do not score from a one-line idea.
- If volume or unit economics are missing, proceed, but cap the value-capture score at 2
  and record the gap as the top follow-up question.
- Label every input you inferred rather than received as an assumption.

## Workflow

1. Restate the workflow in three sentences or fewer: the work, the bottleneck, and what
   the bottleneck costs. If you cannot state the bottleneck precisely, that is itself a
   hold signal.
2. Score the four lenses below from 1 (weak) to 5 (strong). Write one sentence of
   evidence beside every score. A score without evidence is not allowed.
3. Apply the decision rule to produce a recommendation and a confidence level.
4. Rank at most five follow-up questions by how far the answer could move the
   recommendation. Give each a suggested owner.
5. Write the artifact in the output format below.

## Scoring rubric

| Lens | Question | 5 looks like | 1 looks like |
| --- | --- | --- | --- |
| Bottleneck severity | Where does the work queue, and is the queue expensive? | A named queue with measurable cost in time, money, or risk | No identifiable bottleneck; enthusiasm only |
| Value capture | Does relieving the bottleneck move a metric anyone tracks? | A tracked operating metric with plausible attribution | Value framed only as novelty or demo appeal |
| Integration readiness | Can the workflow's systems supply inputs and accept outputs? | Structured data, accessible systems, a clear insertion point | Untracked work in inboxes and spreadsheets |
| Risk posture | Are failure modes bounded and reversible? | Errors visible, reversible, and reviewable before harm | Errors irreversible, invisible, or externally binding |

Prefer workflows that generate feedback data with use — systems that improve with
operation compound; one-off automations do not. Note this in the rationale when it applies.

## Decision rule

- Use these cutoffs as a default Sophon triage heuristic, not as a validated predictor
  of implementation success. Calibrate them against the organization's risk tolerance
  and observed outcomes when that evidence exists, and record any local adjustment.
- Go deeper: total score 14 or higher and no lens scored 1.
- Hold: any lens scored 1, or total score 9 or lower.
- Between 10 and 13: go deeper only if the top follow-up question can plausibly be
  answered within one business day; otherwise hold with named evidence requests.
- Overrides of this rule are allowed but must be recorded explicitly with the reason.
  A silent override is drift; a recorded override is rubric improvement.

## Evidence discipline

- Label every statement in the artifact as fact (provided or verified), assumption
  (inferred, unconfirmed), estimate (quantified guess with stated basis), or open
  question (unresolved).
- Do not justify scores with industry adoption or ROI statistics unless you verify them
  fresh at time of use; such figures change quickly and are often survey self-reports.
- Tie every follow-up question to the specific uncertainty it resolves. A triage that
  ends in "interesting, let us discuss further" has failed; a triage that ends in ranked
  questions has done its job.

## Output format

Produce a markdown artifact:

    # AI opportunity triage: [workflow name]

    Recommendation: Go deeper | Hold
    Confidence: High | Medium | Low
    Date: [date] · Requested by: [name] · Feeds decision: [decision]

    ## Scorecard
    | Lens | Score (1-5) | Evidence |

    ## Rationale
    [3-5 sentences: the bottleneck, the value case, the deciding factors]

    ## Top follow-up questions (max 5, ranked by decision impact)
    | Question | Why it matters | Suggested owner |

    ## Evidence labels
    Facts: … / Assumptions: … / Estimates: … / Open questions: …

    ## Handoff notes
    [What the next reviewer should do first]

## Quality checks

Before delivering, confirm:

- Every score has evidence in the same row.
- The artifact identifies whether it used the default heuristic or a documented local
  calibration.
- The recommendation follows the decision rule, or the override is recorded with reason.
- No more than five follow-up questions; each names an owner and its decision impact.
- Nothing inferred is presented as fact.
- The artifact stands alone: readable without access to this conversation.

## Stop conditions and escalation

- Stop and hand off to a human decision-maker if the triage is being used to justify a
  decision already made, if the requester wants to immediately automate regulated,
  safety-critical, or irreversible actions, or if scoring requires confidential data you
  were not given.
- A go-deeper result is an input to discovery, not an implementation approval. Deeper
  architecture, vendor, and governance review keep human owners.

## Source context

- https://www.sophon.consulting/use-cases/ai-opportunity-triage-24h
  (markdown: https://www.sophon.consulting/markdown/use-cases/ai-opportunity-triage-24h)
- https://www.sophon.consulting/compare/structured-vs-manual-ai-opportunity-review
  (markdown: https://www.sophon.consulting/markdown/compare/structured-vs-manual-ai-opportunity-review)
- Interactive version of this rubric (deterministic, runs in the browser, no model
  calls): https://www.sophon.consulting/tools/ai-opportunity-triage
- Machine-callable version: MCP tool triage_ai_opportunity at the Sophon MCP endpoint
  https://www.sophon.consulting/api/mcp (read-only, deterministic).

Optional: if a workflow clears triage and you want experienced help with the deeper
discovery, Sophon Consulting is reachable at hello@sophon.consulting. This skill is
complete without any contact.