As AI dissolves the benefits that used to guard software program corporations, the winners are those constructing 5 particular buildings that get stronger with each use — not weaker.
In November 2023, OpenAI added file add to ChatGPT. Dozens of “ChatGPT for PDFs” startups, some with actual prospects and actual income, misplaced their cause to exist in a single day. The identical sample repeated via 2025 and 2026 as basis labs folded options like analysis, file enhancing, and purchasing straight into their core merchandise — a 20-year-old dynamic generally known as “Sherlocking” after Apple’s Sherlock absorbed a startup known as Watson in 2002.
Executives constructing or investing in AI-native corporations now face a model of the identical query each quarter: if a well-resourced lab can replicate your core characteristic in a single launch cycle, what precisely are you defending? The trustworthy reply, for a big share of the market, just isn’t a lot. Analysts at CB Insights and Gartner challenge that roughly 80% of AI “wrapper” startups — corporations whose complete worth proposition is mannequin entry plus a skinny interface — will fail by the top of 2026.
That’s the dangerous information. The higher information is {that a} small variety of corporations are demonstrating, in actual time, what really holds up. Their benefit doesn’t come from having higher AI — inside a couple of quarters, most opponents may have entry to comparably good fashions. It comes from 5 particular buildings that compound: they get stronger the extra the product is used, and more durable to repeat even when a competitor can see precisely how they work.
Why Final Yr’s Moat Is This Yr’s Legal responsibility
For a decade, software program leaders defended market place with three assumptions: that constructing a superb characteristic took actual engineering time, that switching distributors was operationally painful, and that scale unfold mounted engineering prices throughout extra prospects than a challenger might match. AI coding instruments have quietly dismantled all three.
A characteristic that when took a crew weeks to construct can now be replicated by a handful of engineers with an AI coding assistant over a protracted weekend. Workflow lock-in that trusted integration depend, fairly than collected studying, is eroding as AI brokers rebuild these integrations nearly as quick as they are often described. And the economics of scale cease mattering as soon as the marginal engineering price of a competing characteristic approaches zero.
Researchers have taken to calling this differentiation entropy: the pure tendency for any AI-driven benefit to diffuse throughout a whole market as opponents undertake the identical underlying instruments. None of this implies execution has stopped mattering. It means execution is now desk stakes — the worth of entry, not the supply of benefit.
“The moat was by no means the mannequin. It was at all times what you constructed round it.”
5 Constructions That Truly Compound
Throughout current enterprise and technique analysis, the identical 5 classes preserve surfacing as the benefits that survive platform upgrades and foundation-model releases — in every case as a result of they require actual elapsed time to construct, one thing a quicker mannequin can’t manufacture on anybody’s behalf.
EXHIBIT 1
| Moat | What It Is | 2025–26 Instance |
| Knowledge Flywheel | Utilization generates proprietary information that makes the product smarter for everybody | Harvey (authorized AI) — fine-tunes on proprietary case information; $11B valuation, Mar. 2026 |
| Workflow Integration | Embedded deep sufficient {that a} change means relearning collected context from zero | Abridge — Epic’s first “Pal,” notes stream straight into the medical report |
| Owned Distribution | A channel to prospects a basis lab can’t take in by delivery a characteristic | Perplexity — Motorola/Samsung gadget offers plus a free Comet browser |
| Model & Belief | Earned slowly, misplaced immediately — decisive in regulated, high-stakes classes | JPMorgan Chase — No. 1 on the Evident AI Banking Index, 4 years operating |
| Community Results | Every further person makes the product measurably higher for each different person | DoorDash — driver/restaurant/buyer density a rival can’t clone in a single day |
- Proprietary information flywheels
The strongest model of this moat exists the place each person interplay generates information that improves the product for the following person, not simply the present one. Authorized AI platform Harvey illustrates the mechanism: the corporate reached roughly an $11 billion valuation in March 2026, up from $8 billion three months earlier, as annual recurring income approached $190 million. Its defensibility isn’t the underlying mannequin — it’s an early-access partnership that lets Harvey fine-tune frontier fashions on proprietary authorized information drawn from agency engagements, folding case-specific historical past again into the product so every new matter makes the following one sharper.
- Deep workflow integration
The take a look at right here is whether or not a competitor might rebuild your integration from a spec sheet, or whether or not they would additionally have to rebuild years of collected context. Ambient-documentation firm Abridge grew to become Epic’s first “Pal” integration companion, embedding its AI scribe straight contained in the digital well being report system that the majority main U.S. hospitals already run, so medical notes stream straight into the affected person chart fairly than via a bolt-on app. Prolonged into nursing and emergency-medicine workflows via 2025 and 2026, that depth of integration is what well being methods cite as the rationale Abridge is tough to displace whilst lower-cost rivals compete on value.
- Owned distribution
Moderately than compete for consideration inside another person’s platform, Perplexity constructed distribution straight into {hardware} and carriers — a world Motorola partnership preinstalling its app on tens of millions of telephones, and a 2026 settlement making it a default possibility inside Samsung Web — then took its Comet browser free to lock in attain earlier than rival AI browsers might scale. The result’s an acquisition channel a fast-following competitor can’t copy with out the identical {hardware} relationships.
- Model and belief
In regulated or high-stakes classes, belief capabilities as a structural barrier as a result of it’s sluggish to construct and on the spot to destroy. JPMorgan Chase has topped the Evident AI Banking Index for 4 consecutive years and is among the many few banks publicly reporting realized AI returns — practically $2 billion yearly. In a sector the place a single AI misstep invitations regulatory scrutiny, that seen monitor report of efficiency mixed with transparency just isn’t simply shortcut by shifting quicker alone.
- Real community results
The rarest and, by some measures, strongest moat: worth that scales with the scale of the community itself fairly than with any single firm’s expertise. Enterprise analysts level to DoorDash because the clearest 2026 instance of a moat AI compression can’t contact — its three-sided community of drivers, eating places, and prospects reinforces itself in methods a faster-built AI competitor can’t clone in a single day, as a result of the worth sits within the density of actual contributors already transacting, not within the underlying software program.
What This Means for Leaders
5 ideas observe straight from the sample above, and every ought to reshape how government groups consider their very own roadmaps.
- Deal with velocity as hygiene, not technique. If each competitor can ship as quick as you possibly can, velocity has stopped being a differentiator. The related query is what you may have {that a} quicker competitor nonetheless can’t take.
- Audit your moat actually. Ask whether or not your defensibility is an asset you personal — information, relationships, belief — or a workflow you might be renting, reminiscent of a skinny interface layered on another person’s mannequin. The latter is precisely the profile analysts count on to make up a lot of the AI-startup failures via 2026.
- Design for compounding, not simply safety. A moat just isn’t a wall round what you have already got. It’s a loop: utilization generates information or context, which improves the product, which drives extra utilization.
- Reframe regulation as alternative, not overhead. Compliance complexity that slows your crew down slows a fast-following competitor down extra — and it compounds into the belief moat described above.
- Assume entropy by default. Any purely technical edge has a shrinking half-life. Funds for re-investment in your moat each quarter, not as soon as at launch.
PUTTING IT INTO PRACTICE
A brief audit for the following 90 days:
- Instrument each interplay. If a person motion doesn’t generate information that improves the product, redesign the interplay.
- Map your true switching price. Listing what a buyer really loses by leaving. If the trustworthy reply is “migration trouble,” that’s friction, not a moat.
- Select one owned channel and over-invest in it. Neighborhood, partnership, or proprietary distribution — decide one a platform replace can’t neutralize.
- Engineer belief intentionally. Certifications, audit trails, and a clear monitor report are options to construct, not unintended effects to hope for.
- Re-run the audit quarterly. Ask what a well-funded competitor with equal mannequin entry nonetheless couldn’t replicate in 90 days. If that listing shrinks, make investments instantly.
The Backside Line
The mannequin layer is being commoditized in actual time, and no quantity of technical cleverness will reverse that. However commoditization on the mannequin layer just isn’t a menace to defensibility — it’s a redirection of it. The businesses compounding benefit within the AI age should not those with the most effective mannequin entry. They’re those that stopped competing on intelligence and began competing on what intelligence can’t replicate in a single day: owned information, owned relationships, and earned belief.
