---
name: scale-align
description: S.C.A.L.E. step 3, Align. Decides whether the friction needs AI at all, names the capability chain and OAR level, makes the build, buy, or partner call, scores up to three vendors against a real account record, and writes guardrails with a cancel trigger. Use when the user says "align the capability", "step 3", "do we need AI for this", "which vendor", "build vs buy", "guardrails", or after step 2.
---

# scale-align

Write `SCALE/3-align-capability.md`. Right capability to the right problem, graded by evidence.
## Preflight (every run)
1. Find the `SCALE/` folder (Desktop or Documents), or `GTM-AI-OS/artifacts/scale/` if the user runs the GTM AI OS. If neither exists, run `scale-start` first; it creates the folder and asks for access.
2. Read `SCALE/progress.md`: mode (quick or deep), AI experience level, which steps are done, the one-line context.
3. Talk short and direct. Say what each answer is for. Beginner: one question at a time, plain words. Advanced: batch questions, skip explanations.
4. Quick mode: ask only the questions marked (Q). Deep mode: ask all. Accept "skip" and "don't know"; record them as gaps.
5. If the previous step's artifact is missing, say so and offer: do that step now, or answer three questions and continue with a draft marked fast path.
6. When the user names a system (CRM, calls, email, spreadsheet), request access to that one or take an upload. Never assume a product.
7. On finish: write the artifact, update `progress.md`, say the next step in one sentence, and ask one open question if any remain.


## Part A: The gate
Ask in order; stop at the first yes:
1. Is the rule known and the input structured? → Automation. No AI.
2. Two systems not sharing a field? → Integration. No AI.
3. Nobody documented the process, or people don't follow it? → Fix the process first.
4. Reading unstructured text, judging, summarizing, deciding under ambiguity? → AI candidate.
Then: would AI cost more than the friction costs? If the verdict is anything but AI candidate, write it, skip to Part D with "no vendor needed," and say that's a good result.

## Part B: Capability
1. Name the chain in order: classification, summarization, generation, extraction, search and retrieval, orchestration. Patterns in `references/capability-map.md`.
2. OAR level: Optimize (time), Amplify (revenue-linked), Reinvent (changes the model). Lowest that solves it.
3. Human in the loop: pothole (more AI), road (mixed), bridge (more human). Outbound to buyers is a road with brand risk.
4. Deep: the 4 Ps (Possibilities, Probability, Payoff, Perspiration) 1 to 5, and readiness (data, integration, governance, feedback loop). Any 1 or 2 becomes a prerequisite.
5. The sentence: "If we deploy <chain> at <level> on <friction>, we expect <metric> from <baseline> to <target> in <window>."

## Part C: Build, buy, or partner
1. (Q) Build when AI changes what or how you sell, the context is yours, and you have the people. Buy when speed, budget, talent constrain and the capability is commodity. Partner when you need it faster than you can hire and it doesn't touch your account records. Total cost of ownership either way. A pilot is almost always Buy.
2. (Q) Shortlist of three, max. Score with `references/criteria.md`: Features, Fit, Security, Scale, pricing, integration. Weight first.
3. The demo test: bring a real account record (from the GTM AI OS if they run it; otherwise a minimal one from `references/account-stub.md`). Better, faster, or more trustworthy, or the vendor fails.

## Part D: Guardrails
1. (Q) What never ships without a human. Defaults: anything built on unverified facts, forecast changes, customer commitments, irreversible writes.
2. Who can change the record; what data never leaves.
3. (Q) The reversal trigger: "In 90 days, what result means cancel?" Owner and date.

## Write
`3-align-capability.md` from `references/decision-template.md`. Quality gate: no vendor recommended that wasn't tested against a record; no shortlist over three; no decision without a cancel trigger.

## Hand off
Mark the step done in `progress.md` and set `resume at`. Fast mode: say "Step done. Moving to Launch." and continue to the next step without waiting. In-depth mode: say "Step done. Say **next** when you're ready for Launch."