Free. No email required. Opens in a new tab as a preview. To fill and save your answers, download it (right-click → Save As) and open in Adobe Reader — placeholder fields clear on focus.
Why this protocol exists
AI startups don’t die at the bottom. They die at the peak. Every indicator is green — ARR climbing, signups vertical, retention by-the-book — and the founder closes the dashboard with the same question they opened it with: are we okay or not? The dashboard stays green until the company breaks, because the metrics built for classic SaaS go silent on the cracks that end an AI SaaS: inference cost per call, top-decile compute concentration, platform dependency, fragile workflows, support tax. Those cracks don’t make noise until they make holes.
The AI Survival Canvas is the X-ray that shows them. The AI Survival Codex is the discipline that decides what to do about them. This page is the third piece: the 90-minute protocol a founder runs on their own business once a quarter — alone first, then with the team, then with an investor. Same Canvas. Three audiences. One causal sentence at the end. Your best quarter is the most dangerous one; this protocol is how you tell the difference before the market does.
Three rules before you fill a single block
- Honesty over optimism. The Canvas is for you, not for the deck. If you don’t know a number, write “don’t know” — that is a finding. Filling blanks with the answer you wish were true is the most common way founders ruin their own diagnostic.
- Numbers over narratives. Every Level B block (economics) lives or dies on one number. If you can’t name the number, you don’t have the block — and that is the prompt to find it before the quarter closes.
- The verdict is yours, not the framework’s. The Canvas surfaces evidence. The causal sentence at the end is your judgment. No book and no checklist gets to deliver it for you.
Will you even reach scale?
Three blocks, ten minutes each, one sharp question per block. The goal is not to write essays; it’s to expose the strategic failure signals that surface before economics matters.
1. Workflow wedge — how painful is it to leave
The question that decides the block. If a competitor offered the same thing 20% cheaper tomorrow, what specifically does my customer lose by leaving? If the honest answer is “nothing much,” you don’t have a wedge. You have a sticker.
Prompts for the worksheet (3–5 lines):
- Name the workflow my product sits inside — not above. Where does the work actually start without me?
- What non-portable context does my product accumulate? (Memory I can export doesn’t count.)
- Am I at the point of action, or only the point of answers?
- Would my value survive if the underlying model were swapped tomorrow morning?
Verdict for the block. Wedge or sticker — in one binary sentence. Example: “Sticker. Our product produces output users like; the work begins and ends in their existing tools.”
2. Distribution engine — loop or pump
The question that decides the block. Without ad budget, does growth still run for the next 90 days?
Prompts:
- What spreads? The landing page (pump) or the work output (loop)?
- What is the per-acquisition channel cost after inference COGS, not before?
- If I cut marketing to zero tomorrow, which channels still bring users by month-end?
- Is there a single channel where one user’s output recruits the next user?
Verdict. Loop or pump — one sentence. Example: “Pump. Growth tracks ad spend at near-perfect correlation; output doesn’t recruit.”
3. Retention loop — M12, not week one
The question that decides the block. What is my M12 retention divided by M3, cohort over cohort?
Prompts:
- Week-one DAU is curiosity noise. Where is the M12 / M3 ratio?
- Is the curve stable, rising, or falling from cohort to cohort?
- Prompts are portable: in AI, switching cost is structurally lower than SaaS. Where is my switching friction actually located?
- If retention looks fine but usage is shallow, am I measuring tourists?
Verdict. One sentence. Example: “M12/M3 = 0.62 and falling cohort-over-cohort. The growth chart is masking churn one quarter ahead.”
Will you survive your own success?
Four blocks, six minutes each. This is the heart of the protocol. If you don’t know your inference cost per dollar of ARR, you don’t know whether your company is alive. Open the calculator next to your model.
4. Inference — the AI-layer gross margin
The question that decides the block. What is the gross margin after inference, tokens, vector lookups, third-party APIs and observability — on the AI feature alone?
Prompts:
- Classic SaaS sits around 70–80%. Healthy AI startups sit around 40–60%. Where do I land?
- Anthropic ran −94% to −109% gross margins in 2024 while ARR climbed. GitHub Copilot lost ~$20/user/month at a $10 price. Am I sure I’m not in a similar pattern?
- Per dollar of ARR, what is my variable AI cost — in cents?
- Is there at least one inference path I own (cache, distillation, multi-provider, local fallback)?
Use the AI Layer GM Calculator here. Eight fields, sixty seconds. Negative result lights up in real time.
Verdict. Example: “AI Layer GM = 34%. Below the 40% floor; a 20% inference price move puts us underwater.”
5. Heavy users — who burns the margin
The question that decides the block. What percentage of my compute is consumed by the top 10% of users — and are they paying for it?
Prompts:
- Typical AI SaaS: top 10% burns 50–60% of compute. Where am I on that curve?
- Are my power users my most profitable or my least profitable customers?
- Does price move with consumption (seats with limits, credits, tiers, pure usage) — or is it flat on top of a variable cost?
- Every “successful” power user the team celebrates — do they widen the loss or close it?
Verdict. Example: “Top 10% burns 58% of compute and pays the same as the rest. Each new power user makes the unit economics worse.”
6. Unit economics — LTV done honestly
The question that decides the block. What is my LTV recalculated on contribution margin after inference, not on revenue?
Prompts:
- Classic LTV/CAC assumes 70–80% gross margins. If mine is 52%, a clean “5:1 ratio” on revenue is closer to 2.4:1 on real contribution.
- What is my payback period when measured against the AI-adjusted margin, not the marketing margin?
- Per channel: gross profit per acquired user, not revenue. Which channel is the deepest hole?
- What does the same calculation look like at 10× current volume?
Verdict. Example: “Revenue-LTV ratio looks 4.1; contribution-LTV after inference is 1.9. Payback runs out before retention does.”
7. Burn & runway — the real number
The question that decides the block. Stated runway minus the next inference price move, the next model upgrade, and the next renewal cliff — how many months left?
Prompts:
- Cash on hand divided by current monthly burn — the stated number.
- Re-run with inference cost at 130% of today. How many months disappear?
- Re-run with the assumption that 25% of my largest contracts hit a renewal cliff at month 9. How many months disappear?
- If the next round were delayed by 6 months, where does the company stand?
Verdict. Example: “Stated runway 14 months; real runway under one realistic vendor repricing closer to 9. A round is required, not optional.”
Will you survive the worst case?
One block. Ten minutes. Run three scenarios end-to-end against the seven blocks above and see which layer fails first.
8. Shock stress test — what breaks when shocks arrive together
The question that decides the block. Which layer fails first under realistic combined shock: strategy, economics, or capital?
Three scenarios to run on the worksheet:
- Conservative. Inference price +20%, growth at 60% of plan, no churn shock. What is the AI Layer GM, what is the LTV recalc, what is the new runway?
- Optimistic. Inference price −10%, growth at 130% of plan, retention stable. Does the company still need a round to reach contribution-positive?
- Renewal Cliff. 30% of top-20 enterprise contracts churn at renewal in month 9 simultaneously with a 15% inference price move. What survives?
One shock rarely arrives alone — historically these correlate. The Canvas asks for the multi-shock combination, not a single-variable sensitivity.
Verdict. Example: “Economics layer fails first under Renewal Cliff. Strategy is intact, capital is intact, but contribution margin goes negative the same week.”
Compose the causal sentence
The longest segment of the clock, because everything else is preparation for it. The verdict is not a metric and not a paragraph — it is a three-part causal sentence: now, what breaks at scale, why.
9. Survival Verdict — one causal sentence on fate
The structure. Three clauses in one line: (now) — (what breaks at scale) — (why).
Templates to start from. Use one and edit it into your own:
- “We grow users faster than contribution margin, and at 10× scale inference cost becomes the business itself, because we have not yet built a wedge that justifies the per-call cost.”
- “We have a real wedge and 11 months of runway, but at 2× volume the top decile of users moves the AI Layer GM into negative, because pricing has not yet moved with consumption.”
- “Retention is stable, the wedge is genuine, but a single inference repricing closes the runway gap before the next raise, because we still rent every model call.”
The test for a good verdict. If an investor could repeat it back to you 20 minutes later, it is a verdict. If it sounds like a report — LTV:CAC X, runway Y months — rewrite it.
Write yours in the final field of the worksheet. Then read it aloud once. If it sounds optimistic, you wrote a pitch; do it again.
Now show it to your team without losing the truth
Solo first — always. Then bring it to the team within the same week. The temptation is to soften: turn “sticker” into “wedge in progress,” turn 34% AI Layer GM into “industry-typical.” Don’t. The team conversation works only if the worksheet shows what you actually wrote alone.
A workable 15-minute team script:
- Minute 0–2. Show the Canvas on a single screen. Read out only the 9 binary verdicts — one line per block. No numbers yet, no defence.
- Minute 2–8. Open Layer 3. Walk the team through Conservative, Optimistic, Renewal Cliff. Ask one question: which layer fails first — and does anybody disagree?
- Minute 8–14. Read the Survival Verdict aloud. Ask the team to suggest one edit. Almost always, somebody’s edit is sharper than your draft.
- Minute 14–15. Pick the one block where the team most disagrees with your verdict. That block goes into next week’s execution. The rest is monitoring.
What the team gets out of it: a shared causal sentence everybody can quote back, and one operational priority that wasn’t obvious before. What you get: 1–2 corrections you would have missed alone.
Now show it to an investor without letting them invent your verdict
An experienced VC delivers their verdict on you inside the first twenty minutes of the conversation. The only question is whether they hear yours first or invent one along the way. A founder who walks in with a causal verdict controls the meeting. A founder with twelve metrics and an optimistic narrative hands the investor the right to diagnose the business — and the invented diagnosis usually lands harsher than the one the founder would have offered.
The 15-minute investor reframing. Take the same worksheet. Strip it to four lines — in this exact order:
- The Survival Verdict, as written in the final block. One sentence. Read it first.
- The strongest single number that supports it. Not five — one. (For us today: AI Layer GM 34%, falling 2 pts per quarter.)
- The block where the leak is. Layer 1, 2, 3, or 4. Name it. No hedging.
- The next 90 days. What we do about it — in one sentence, with a number.
If the investor wants to see the worksheet, give it to them. They almost always want it — honest founders are rare. If they want metrics first, hand them the Canvas with all nine binary verdicts visible. The board memo writes itself.
The 9-minute monthly check
Full 90-minute clock once a quarter. Compressed 9-minute version once a month, no worksheet required. The 9 minutes are one minute per block: read the binary verdict, decide if it changed, write down the change in a note. If a verdict flips — from “wedge” to “sticker,” from “loop” to “pump” — run the full clock the same week. Don’t wait for the quarter.
Immediate full re-run is also triggered by any external shock — vendor repricing, model upgrade, competitive launch, large contract loss. The protocol exists to remove the latency between an event and a verdict. Latency is what kills.
Seven ways founders ruin their own Canvas
- Filling blanks with optimism. “Wedge in progress” is “sticker.” “Retention stable-ish” is “falling.” Use binary words.
- Skipping Layer 2 because the numbers are uncomfortable. Economics is the heart of the protocol. Skipping it produces a strategy document, not a survival diagnostic.
- Presenting metrics where a verdict belongs. A report is not a verdict. The board forgets reports in an hour.
- Treating the wrong layer. Fixing CAC when the wedge is broken. Raising a round when retention is dead. Optimizing margin when there is no loop. (See Codex Principle 9.)
- Writing the team script before writing the solo Canvas. The team conversation collapses if the founder hasn’t seen the diagnosis alone first.
- Letting the investor diagnose for you. The verdict is yours first, theirs second.
- Running the protocol once and stopping. The compounding value is in the cadence — quarter to quarter, the verdict drifts. Quarter-on-quarter drift is the only signal that growth is real.
What to do next
Reading isn’t doing. Three options, in ascending order of investment: