The Agentic PMF Standard

The story

Consider a B2B SaaS team that ships an AI summarization feature for sales call transcripts. NPS is high. Adoption climbs to 60% of the user base in the first quarter. Retention metrics look healthy. The team declares PMF.

Six months in, a competitor launches a similar feature. Within four weeks, the team's AI feature usage drops 30%. Customers cancel their subscriptions to that specific add-on. Retention on the core product is unaffected — users keep buying the SaaS — but the AI add-on revenue collapses. The team's PMF was a substitution PMF. Users liked the feature when it was offered. They didn't need it. The moment a competitor offered a similar feature, switching cost was zero.

This is the Agent Tax anti-pattern. Users tolerate the AI. They don't depend on it. The dashboard says PMF. The economics say "another tax line item the customer will cut at first opportunity." Traditional PMF metrics — adoption, retention, NPS — under-detect this pattern because they measure whether the feature is used, not whether the workflow has restructured around it.

A second team takes a different approach. They build an AI agent for invoice extraction at a logistics company. The agent isn't fast or flashy at launch — extraction accuracy is 71% in the first month. The team focuses on a different metric: what percentage of the invoice processing workflow now assumes the agent's output is the source of truth? In month 1, that number is 12%. In month 2, 34%. By month 4, 78%. By month 6, the workflow can't function without the agent — finance has retired the manual extraction template, the audit team has rebuilt their reconciliation process around the agent's confidence scores, and three downstream systems consume the agent's output as their primary input.

Accuracy continued improving (91% by month 6). But the more important metric was workflow restructuring. The competitor who shipped a similar agent in month 9 couldn't displace this team — because displacement now required redesigning three downstream systems, not just swapping a vendor. The team had built agent-native indispensability. That's the PMF standard for AI products.

The core idea

Traditional PMF asks: do users keep using the feature? That question worked for SaaS because the cost of switching was the cost of switching software. AI products break this question because users can keep using a feature without depending on it — the Agent Tax pattern — and they can stop using a feature without churning from the core product.

Agentic PMF asks a different question: has the workflow restructured around the AI such that returning to pre-AI is measurably costly? If the answer is yes, you have indispensability. If the answer is no, you have an Agent Tax — a feature users tolerate that competitors can poach.

The concept — visualized

The Indispensability Index

The Indispensability Index is the single defensible number for agentic PMF. It's calculated as:

Three calibration points:

The Index is harder to fake than retention or NPS. You can't survey a workflow into restructuring itself. The number reflects whether downstream systems have been rebuilt around the agent's output — which is observable in API logs, integration architectures, and process documentation. A finance team that retires the manual-extraction template, an audit team that builds reconciliation on the agent's confidence scores, three systems consuming the agent's output as primary input — these are workflow restructuring artifacts. They produce the Index naturally; they can't be faked.

The 7 Fits adapted for AI products

Marc Andreessen's traditional PMF concept ("product-market fit means being in a good market with a product that can satisfy that market") maps loosely to AI but misses the layers underneath. The 7 Fits framework is the AI-native expansion:

A team can have 6 of 7 fits and still fail. Most AI initiatives die because they have problem fit and solution fit but lack economic fit or harness fit — the AI works, users want it, the unit economics don't.

Where this hits in production

The 5–17% of AI initiatives that ship into profitable production share a measurement profile. They track the Indispensability Index from week 1, not week 26. They report all 7 Fits in their quarterly reviews. They have a defensible answer to "what would it take for users to abandon the AI?" — usually expressed as the redesign cost of downstream systems.

Klarna's mixed AI story is a 7-Fits diagnostic. The 2024 announcement claimed AI replaced 853 FTEs (problem fit ✓, solution fit ✓, economic fit appeared ✓). The 2025 reality — re-hiring humans for nuance work — surfaced that harness fit and solution fit were weaker than the headline suggested. The agent handled volume but not nuance. Volume work didn't require deep workflow restructuring (low Indispensability Index for the nuance segment). The savings were real for one segment and overstated for another. A 7-Fits review would have caught this in month 2, not 14.

The Cursor case study is a 7-Fits diagnostic of a different kind. Problem fit ✓, solution fit ✓ (developers loved it), market fit ✓ (willingness-to-pay was high), channel fit ✓ (PLG worked). Model fit and harness fit started strong and degraded as power-user usage patterns emerged — the harness wasn't built for multi-hour agentic sessions. Economic fit collapsed. The fix was harness re-engineering: context compression, semantic caching, loop pruning. Once the harness fit was restored, the economic fit followed.

The Apple Intelligence pattern is the 7-Fits aspiration. All seven fits are addressed in the architecture: Stateless AI plus Private Cloud Compute (harness fit and economic fit), the model swap from Gemini to in-house Ferret-3 (model fit), workflow integration deep in iOS (problem fit, solution fit, channel fit), monetization through device margin (market fit). The architecture treats the harness as the moat — exactly the L1-T01 lens.

The trap

The trap is celebrating adoption as PMF. Adoption is necessary but insufficient. A 60% adoption rate with a 15% Indispensability Index is the Agent Tax pattern. Users see the feature, click it occasionally, and would notice if it disappeared but wouldn't redesign their workflow. The first competitor with a similar feature and aggressive pricing takes the segment.

The fix is to instrument the Indispensability Index from week 1. Every workflow review should include the question: which downstream steps now assume the AI's output as source-of-truth? If the answer is "none" or "very few" after three months, the team is shipping toward Agent Tax, not toward indispensability.

The second trap is treating indispensability as a binary state. It's not — it's a continuum, and the Index is a continuous measurement. A team at 35% should be running experiments to push to 50%. A team at 65% should be reinforcing the workflow ties so the Index doesn't slip backward. The competitive risk doesn't disappear at 60%; it changes shape — from displacement risk to continuous-improvement risk.

The third trap is conflating user attachment with workflow attachment. Users may love a feature emotionally and abandon it functionally. The Indispensability Index measures the workflow, not the affect. A user who says "I love this feature" but whose downstream systems don't consume the output is still in Agent Tax territory.

Remember this

In practice

Step 1: Calculate your Indispensability Index for each AI initiative. Pick the target workflow. List the workflow steps. For each step, ask: does this step now consume the AI's output as source-of-truth? Calculate the percentage. If the Index is below 25% after 3 months, you're shipping toward Agent Tax. Above 60% means you've crossed into indispensability.

Step 2: Run a 7-Fits review on each initiative. Score each fit 1–5. Identify the weakest fit. Most AI initiatives are weakest on harness fit or economic fit. The weakest fit is the workstream for next quarter.

Step 3: Reframe the roadmap as outcomes-not-outputs. Stop reporting "AI shipped feature X." Start reporting "the workflow moved Indispensability Index from 32% to 47% on the target use case, driven by feature X plus integration into systems Y and Z." Outputs are activity. Outcomes are the workflow change that activity produced.

Step 4: Add the Index and 7 Fits to the quarterly review. A defensible AI portfolio review reports, for each initiative: Indispensability Index trajectory, weakest fit, plan to address weakest fit. Compare to previous quarter. Promote initiatives whose Index is climbing. Restructure or kill initiatives where the Index is flat after two quarters.

Step 5: Translate the Index into business language for leadership. No CFO knows what an Indispensability Index is. Translate: "The workflow is now 67% restructured around the AI. Switching to a competitor would require three integration rewrites and a process redesign — a 2-quarter project for the customer, which makes their renewal probability ~85% versus the ~62% baseline for AI add-ons across the industry." The Index turns abstract PMF into concrete renewal probability.

The practice — visualized

References