Applied AI Startup Moats

PDF

Know-how and outcome data compound; owned distribution and trust hold once earned; the rest buys time, is unproven or is fading. For AI startups that build on other companies’ models.

Preliminary view, Nov 2026. Tiers are the author’s judgment.

Tier Acompounds

Domain / process know-how

How work really gets done in one field (edge cases, review steps, tests), learned across many deployments and built into the product.

e.g. Harvey·Legora·EliseAI

The test: Evidence that each new deployment takes less work than the last, and that the lead over a raw frontier model holds on the company’s own tasks when a new model ships.

Objections:

  • Starts manual and bespoke per customer.
  • If raw models learn to do the work end to end, this shrinks.
Sources: Tomasz Tunguz, Jul 2026·Harvey’s benchmark, May 2026

Why it’s in this tier: Raw models still finish under 10% of complex legal tasks end to end, so knowing how the work really gets done carries from one customer to the next. a16z calls this “perhaps the strongest moat of all in this new era.”

Harvey $15.5B, Sep 2026·AI assistant for law firmsLegora $8.8B, Oct 2026·AI legal workspace built inside law firmsEliseAI $4B, Sep 2026·AI leasing agents for property managers

Exclusive data that grows with scale

Data only you have, growing with every customer. At its best a data flywheel: the work shows which outputs were accepted, paid or won, and the product improves.

e.g. EvenUp·Decagon·OpenEvidence

The test: Accuracy or cost per task by customer cohort, and whether a rival with the same volume but none of the outcome labels could match it.

Objections:

  • Partner data (journals, hospital records) is a head start: owners rarely keep it exclusive.
  • If synthetic data gets good enough, real data matters less.
Sources: Oliver Wyman, Sep 2026·ABA Journal, May 2026

Why it’s in this tier: Only outcome data counts: which outputs were accepted, paid or won. It exists only if you do the work. Plain usage data did not protect AI scribes in 2026.

EvenUp $2B+, Oct 2025·learns from what past injury claims settled forDecagon $4.5B, Jan 2026·learns from which support chats got resolvedOpenEvidence $15B, Sep 2026·licensed journals plus doctors’ questions

Tier Bholds once earned

Owned distribution

Customers come to you directly and use you every day, so you don’t depend on anyone else’s channel.

e.g. OpenEvidence·Lovable·Replit

The test: The share of usage that comes directly, and how often users return without being prompted.

Objections:

  • Cursor had this and lost share within a year.
  • If ChatGPT or Claude becomes the front door, the audience follows.
Sources: JetBrains Research, Aug 2026·a16z, Sep 2026

Why it’s in this tier: An audience that comes to you directly and uses you daily builds on itself through habit and word of mouth, and it doesn’t depend on a host. It’s being tested now: OpenEvidence faces a free rival, ChatGPT for Clinicians, since April 2026.

OpenEvidence $15B, Sep 2026·AI search doctors open dailyLovable $13.3B, Aug 2026·app builder that people come to directlyReplit $9B, Mar 2026·app-building and hosting platform people use directly

Trust to act on the customer’s behalf

The customer lets the product act unchecked: a company’s agents in production, or a person’s logins and payments. A replacement must be re-tested.

e.g. Sierra·Instinct·Cognition

The test: Whether customers buy without a bake-off, and which audits, certifications or indemnities the last ten enterprise buyers required.

Objections:

  • Slow to earn; one public failure resets it.
  • If big platforms offer the same delegation, the lead narrows.

Why it’s in this tier: Trust matters more once agents act on their own. As a16z puts it, brand is how you signal reliability. Insurers are starting to exclude AI from standard cover, and a certification and insurance layer for agents is forming.

Sierra $15.8B, May 2026·AI customer-service agents for enterprisesInstinct $10B, Sep 2026·personal agent that books, buys and calls for youCognition $26B, May 2026·AI software engineer working in company code

Tier Cbuys time

Borrowed distribution

Sold inside an incumbent’s product or marketplace. Fast reach, but the host controls access and can replace you.

e.g. Abridge·Ambience

The test: The share of new customers that arrive through a partner that could ship a rival, and what stops that partner.

Objections:

  • Microsoft and Epic won 68% of new AI note-taking spend in 2026.
Sources: MedCity News, Feb 2026

Why it’s in this tier: A host’s product gives fast reach, but the host decides. In AI note-taking, Microsoft and Epic won 68% of new spending in 2026, and 71% of Epic health systems look to Epic first when buying AI.

Abridge $5.3B, Jun 2025·AI visit notes, sold inside EpicAmbience $1.25B, Jul 2025·AI visit notes for health systems

Licensed, liable party

The company holds the professional licence and carries the liability, so it can sell finished work. Model labs avoid this role.

e.g. Crosby·Harper

The test: Whether the licence is scarce or capped, who was second to get it, and how long it took them.

Objections:

  • Not scarce: a funded rival gets a licence in 6 to 24 months.
  • Regulators can change the rules, in either direction.
Sources: Baker Botts, Oct 2026·Akerman, Nov 2026

Why it’s in this tier: Model labs stop at tools for licensed professionals, so a company that holds the licence and carries the liability can sell finished work. It isn’t higher because licences aren’t scarce: at least six licensed AI law firms appeared in about 18 months.

Crosby $400M, Mar 2026·law firm that uses AI to review contractsHarper raised $47M, Feb 2026·AI-run licensed insurance brokerage

Counter-positioning to incumbents

A pricing or business model incumbents can’t copy without cutting their own revenue.

e.g. Decagon·WithCoverage

The test: What it would cost the incumbent to match the model, and when it would accept that cost.

Objections:

  • Buys time: Zendesk and Salesforce now also charge per resolution.
Sources: Zendesk help centre

Why it’s in this tier: It’s how most AI startups get in. As Sequoia puts it, for incumbents “selling the work means cutting their own customers out of doing it.” It only buys time: Zendesk added outcome pricing in 2024, and Salesforce agreed to buy Fin (formerly Intercom) for about $3.6B and closed the deal in 2026.

Decagon $4.5B, Jan 2026 vs Zendesk·per resolution, not per seatWithCoverage raised $42M, Jan 2026 vs brokers·flat fee, not commission

Emergingunproven

Network effects

Each new participant makes the product more valuable to the others, as in a marketplace.

e.g. Harvey·Tennr

The test: Evidence that a new customer makes the product better for existing customers.

Objections:

  • Promising; the evidence so far is company-reported.

Why it’s in this tier: The hardest moat to attack once it’s real, but no applied AI startup has an independently verified one yet. Harvey, Legora and Tennr are building toward it with shared client workspaces and referral networks.

Not the same as scale (two separate powers in the 7 Powers framework a16z applies): network effects make the product worth more to each customer as more join, while scale makes each customer cheaper to serve. See a16z, Mar 2026.

Harvey $15.5B, Sep 2026·law firms and clients in shared workspacesTennr $605M, Jun 2025·links referring doctors with receiving providers

Multi-product platform

Covers more of a customer’s work on one shared layer that learns how they operate. Each added product is easier to sell, harder to remove.

e.g. Glean·Ramp·Bunkerhill Health

The test: Net revenue retention and products per customer by cohort, and what a customer would lose by moving its data to a rival.

Objections:

  • Unproven: no AI startup publishes retention or attach rates.
  • Only context found nowhere else counts; models pick up most of the rest in days.

Why it’s in this tier: Each added product is cheaper to build and sell on one shared layer, and harder to remove: one health system went from 1 to more than 20 Bunkerhill agents, per Sequoia. It’s still emerging because lock-in is weaker than it looks: developers moved from Cursor to Claude Code within months.

Glean $7.2B, Jun 2025·AI search and agents across a company’s appsRamp $60B, Oct 2026·most customers use 2+ of its finance productsBunkerhill Health raised $55M, Jul 2026·1 to 20+ agents at one hospital

Owned hardware

A device the company controls sits where the data is created.

e.g. Plaud·Heidi Health

The test: What happens to subscriptions when a phone or glasses maker ships the same capture.

Objections:

  • Phone and glasses makers can absorb the device.
Sources: TechCrunch, Dec 2025

Why it’s in this tier: A device sits where the data is created, and real revenue exists: Plaud says it has $100M ARR, and Heidi Health launched its own clinical microphone in March 2026.

Plaud $100M+ ARR, 2026·AI voice recorderHeidi Health $900M, Sep 2026·AI scribe with its own microphone

Fadingcounted as moats until recently

Thin wrappers and UX

Labs ship the same features quickly.

Why it faded: When ChatGPT added PDF upload in 2023, it threatened PDF-chat apps. Cal AI, a breakout consumer wrapper, sold to MyFitnessPal in March 2026.

A smarter custom model

The next general model catches up.

Why it faded: GPT-4 beat BloombergGPT on almost all the finance tasks tested in 2023.

Lock-in from stored records alone

Agents now do the migration.

Why it faded: a16z: “Switching costs are perhaps the one moat that really is going to change.” AI also “lowers the cost of recreating the first 80% of a system of record,” a16z adds.

Static or plain usage data

Can be bought, generated or bypassed.

Why it faded: Oliver Wyman on Chegg: “Chegg’s data didn’t disappear. Its moat did.” AI-native scribes also lost share in 2026 despite their usage data.