The Core Thesis
What's the real prize in Vertical AI — and where does it hide?
The most valuable data in vertical AI gets created constantly and captured almost never. Every time a human overturns a denied prior auth, settles a freight dispute, or swaps an equipment order, the reasoning behind the call evaporates the moment they move to the next case.
Agents can already follow rules, pull data, and generate documents. What they can't do is handle exceptions — the case that doesn't fit the template, where a human intervenes based on market, customer, and counterparty context. Those calls don't live in an ERP, CRM, or WMS. They pile up in the exception queue and vanish as fast as humans clear them.
El-Ayat's claim: long-term enterprise value accrues to whoever captures why the call was made, whether it worked, and how that reasoning should change the next case.
The early breakouts all sit on a queue
Different products on the surface — identical underneath. Each sits inside a repeated, high-consequence workflow that leaves behind a reusable decision trace with every resolution.
- Tennr — owns referrals (healthcare intake)
- EvenUp — owns demand packages (legal / personal injury)
- Abridge — owns clinical documentation
- Fieldguide — audit evidence & review notes
- SmarterDx — clinical documentation & revenue integrity
The Two-Axis Screen
The one genuinely reusable tool in the essay. Tap a quadrant.
The two axes, defined
- Exception density — how often do non-standard, judgment-requiring cases recur in this workflow?
- Cost of failure / delay — what's the damage when a case is missed, mishandled, or resolved too late? Shows up as unrealized revenue, idle labor, margin leakage, compliance exposure, or stuck cash flow.
"Glue functions" — where to hunt
The best wedges sit on work that lives across fragmented systems — spreadsheets + email + phone + portals + people's heads. Care coordination, claims processing, freight audit, permit coordination, procurement exceptions.
The profile in one line: judgment calls someone is already paid to clear + high cost of delay + a recurring queue.
Two ways into the queue
At the exception itself (Tennr, EvenUp) — resolution is the product. The trace comes free with every case. Risk: you're in the messiest, most integration-heavy part of the workflow.
Upstream at the authoring event (Abridge) — own the high-context moment first; exceptions surface downstream. Risk: you generate context but never reach the queue, leaving you "just a productivity tool." Only works if the authored object necessarily flows downstream (clinical note → coding → billing).
Score a Wedge
Drag the two sliders for any workflow or target you're evaluating.
The Juice
Verified-useful signal only. Noise and unproven claims stripped out.
1. The two-axis screen
Exception density × cost of failure/delay. The only real wedge is high × high. Useful as deal-DD scaffolding and market mapping. (Not original in deep structure — Lindy-traceable to Christensen + the vertical-SaaS-moat tradition — but the AI operationalization is crisp.)
2. Delay is failure
"Cost of failure" isn't just errors — it's unrealized revenue, idle labor, margin leakage, compliance exposure, and stuck cash flow. A workflow can technically complete and still have failed.
3. "Glue functions" as a hunting heuristic
Fragmented context across systems + judgment-dependent resolution + a recurring queue = strong wedge. A concrete way to spot where the value hides.
4. Two entry points
Enter at the exception (resolution is the product) or upstream at the authoring event (own context, exceptions surface downstream). Upstream only works if the authored object necessarily flows into a downstream queue.
5. The decision-trace mechanism
The trace (which policy applied, which precedent mattered, who approved the override, what the customer accepted, how long, what happened after) is genuinely first-party and lives nowhere structured today. Two compounding loops: across-customer (more variants) and within-customer (the firm's own decision logic). Interesting mechanism — but treat as a hypothesis to validate, not a proven moat.
What got cut
- The moat conclusion — asserted, not demonstrated. No theory of why the trace data is non-replicable; exceptions can be commoditized once a model gets good enough to eliminate them.
- The five winners as "validation" — selection artifacts from a VC's deal pipeline. No graveyard of companies that attacked the same queues and failed.
- The KFF prior-auth + 35%-3PL stats — verify independently before citing. The 35% figure has no clear source chain and could be pitch-deck math.
The Filter
6-lens source credibility check. Verdict: 🟡 Extract Selectively.
Asymmetric credibility (3-of-4)
| Signal | Result | Why |
|---|---|---|
| Did they do it? | ✗ | VC investor with pattern exposure, not an operator inside these queues |
| Survived time? | ~ | Underlying insight is Lindy; the AI application is ~18 months old |
| Bleed if wrong? | ✗ | Lagged, diversified VC consequences; Substack has zero accountability |
| Actionable? | ✓ | The two-axis test is specific enough to apply |
Negative filter — the flags
Never operated in the domain
VC analyzing systems he hasn't run. Sophisticated observer, not a builder.
Survivorship bias
Five winners, zero graveyard. What separated Tennr from companies that hit the same queues and failed? Without that, the thesis is confirmed only by the winners it selects.
Borderline unfalsifiable
Almost any vertical AI company can be post-hoc rationalized as "winning the queue" or "not there yet." The clean chart looks rigorous; both axes are subjective ex-ante.
Repackager (partial)
The two-axis test is his genuine formulation, but the core insight (exceptions = sticky software) is repackaged from decades of workflow-software thinking.
Engagement-optimized (partial)
Named examples + clean two-axis chart = built for shareability in VC/tech circles. Pressures toward clean narratives over messy reality.
Source of the source
Legit upstream lineage (Christensen, vertical-SaaS moat tradition, KFF as an institution). But the downstream citation chain is likely circular — other VCs cite this essay, which cites companies that are themselves the data source, looping back to the ecosystem validating its own thesis.
For Accord
How this maps to an AI-powered audit roll-up.
Audit IS an exception-queue business
- Fieldguide (audit evidence / review notes / findings) and SmarterDx (coding gaps, documentation mismatches) are the two on his list closest to your world. Evidence requests, review exceptions, audit trails — that's the queue.
- Use the screen as deal DD scaffolding: for any AI target, ask where does the human exception queue live, how dense is it, what's the cost of a wrong/late call?
- Necessary, not sufficient. A target that doesn't touch a queue is probably weak — but touching one doesn't make it strong. Don't let the framework launder mediocre targets into "defensible."
- The danger: absorbed uncritically, "they're in the exception queue!" becomes a false signal of defensibility. Treat it as a screen, validate the moat claim with operator-level evidence.