Ai feature go no go

AI Feature Go/No-Go

A decision, not a discussion. The output is GO or NO-GO, and it names the one criterion doing the deciding — everything else is ranked context.

Verification gates (defined first; output is blocked until all pass)

• G1 — Pivot criterion named: the decision states the SINGLE disqualifying (for NO-GO) or qualifying (for GO) criterion it turns on. Five co-equal reasons with no named pivot fails the gate. A conditional GO states its exact condition; "it depends" is not a decision.
• G2 — Non-decisive factors marked: supporting factors are ranked below the pivot and explicitly marked non-decisive. If removing the pivot wouldn't flip the decision, the pivot is wrong — find the real one.
• G3 — No fabricated context: the decision argues from provided context only. No invented compliance rulings, competitor moves, usage benchmarks, or user demand. Arithmetic uses input numbers; anything else is a labeled estimate.

Steps

1. Fix the failure surface. What happens when the AI is wrong, who sees it, and can it be caught before harm? Error tolerance × review point is where most AI features live or die — check it first.
2. Check the remaining axes, in order: value density (does the feature remove real work — do the arithmetic on volume and FTEs), feasibility at quality bar (can current models meet the tolerance found in step 1), economics (per-unit cost vs. price/margin — route deep dives to roadmap-reality-check), and trust/adoption (will users accept AI in this moment).
3. Find the pivot. One axis almost always decides; the rest calibrate. Test it: if this criterion flipped, would the decision flip? If not, it's not the pivot.
4. Write the decision. GO / NO-GO / GO-IF (with the exact condition). Then the reversal line: what change in the world or the feature's design would flip this call — a NO-GO on unreviewed output may be a GO with human-in-the-loop, and the output must say that's a different, smaller feature.
5. Gate pass. One named pivot (G1), non-decisive factors marked (G2), every contextual claim traceable to input (G3). Fix and re-run; maximum 2 repair loops, then report the failure instead of the output.

Output format

DECISION: NO-GO
PIVOT CRITERION (disqualifying): unreviewed generative output in a regulated,
payments-adjacent support flow with stated near-zero error tolerance. One wrong
auto-sent answer about a payment is an incident, and nothing in the design catches
it before the customer does.
NON-DECISIVE FACTORS (would not flip the decision alone):
- volume: 400 tickets/mo ÷ 1.5 FTE — no capacity crisis to justify the risk [input math]
- CSAT 4.6/5 — protecting a strength, not fixing a weakness
WHAT WOULD FLIP IT: agent-reviewed drafts instead of auto-send (a different, smaller
feature: GO-worthy on the same context) · or error tolerance materially loosening
GATE CHECK: G1 pass (one pivot) · G2 pass · G3 pass

Hard rules

1. One pivot criterion, named, always. If two criteria genuinely co-decide, pick the one that fails first in deployment order and note the second as next-in-line — never present an unranked list as a decision.
2. Never hedge a decision into meaninglessness. "GO with monitoring and phased rollout" that doesn't name what would disqualify it is not a decision — it's postponed accountability.
3. Never invent context to make the call easier. Missing load-bearing context (error tolerance, volume) → ask for it or state the decision is provisional on the named missing fact.
4. The reversal line is mandatory. A decision that can't say what would flip it hasn't found its own pivot.

Limitations

• The call is a structured judgment on provided context — it cannot see org politics, roadmap opportunity cost, or strategy fit beyond what the input states.
• Feasibility-at-quality-bar reads current-generation model capability as commonly known; a borderline call deserves a technical spike, and the output says so rather than guessing.
• GO-IF conditions are design requirements, not guarantees the condition is achievable.
• One feature per call; portfolio-level sequencing across many candidate features is roadmap territory.