I've been racking my brain over which businesses continue to survive the next few years, and which ones die. Which moats are still standing at the end of the great software repricing?
A whole set of trends converges on the same conclusion - the way SaaS has operated for twenty years stops being viable. To see why, you have to be honest about what SaaS actually was. Software was never the product. SaaS sold three things disguised as software: a usable interface to data, encoded domain expertise, and a bundle of coordination overhead a company couldn't do itself. Two simultaneous shifts destroy all three:
- The cost of writing software is collapsing toward zero. Coding agents crossed the "good enough" line a while ago, and with Claude Fable 5 now generally available, the gap between an idea and working software is a weekend in most cases.
- The users of software are becoming agents rather than humans - buying, searching, operating, integrating.
Everything whose value depended on human friction - learning curves, vendor-discovery costs, retraining costs, brand recall, seat counts - commoditises. I made a narrower version of this argument in The Trust Barrier: every barrier made of human labour falls to agent scale, because labour is the thing agents mint for free and at scale in parallel.
The attention budget that turned a city of restaurants into a shortlist of four, the expertise budget that kept a trader in one jurisdiction for an entire career: both gone the moment you hand the job to a fleet of competent agents. The one barrier I argued there survives is trust, because it's built from verification lag rather than hours, and no amount of intelligence makes the lawsuit arrive sooner. This post is my thinking on a larger scale - not just the trust moat, but every classical software moat that survives vs dies.
What survives is everything an LLM cannot regenerate: proprietary data that compounds through operation, trust in hard-to-verify work, position inside a transaction, regulatory standing, physical and capital infrastructure, and the velocity of a learning loop. SaaS as a delivery mechanism persists; SaaS as a moat category is dead imho.
The winners aren't the app companies soaking up mega rounds right now but the operators - small, agent-native firms that run outcomes end-to-end and own the data exhaust - and the infrastructure those operators and their agents stand on. And while all of that plays out, the firm itself shrinks: agents collapse every dimension of a company's operational stack at once, so the median company gets radically smaller while the number of companies explodes.
Two Motions
Two motions are compounding each other.
The first: the supply of software is trending towards infinity. Coding agents make "buildable in a weekend" table stakes, so your competitor is no longer just another SaaS company. It's your own customer, rolling their own replacement software or even just rolling out an agent and a markdown file of business rules the SMEs have put together. At SAMMY we saw this over and over when we were working in CS - we saw buyers stop evaluating our product and instead estimate how long it would take to rebuild it internally.
One of our customers, a customer-success leader, would openly talk about how he'd torn a well-known CS platform out of two different companies he'd been at. He explicitly told us his team is "very much build versus buy."
The catch he named in the same breath is the whole opportunity. His company can build because it has real developer culture and product horsepower, and (keep in mind this was early 2025) "the vast majority of tools, platforms, products out there are not like that". I think that gap has been squeezed since.
You can already watch it closing in the numbers: vertical companies writing their own software instead of buying it, while the model layer's run rate dwarfs the entire AI app layer beneath it. When your customer is also your competitor, no roadmap saves you.
Play the SaaS game unchanged and that single shift closes on you from both sides. Cheaper software means more competitors - commoditised clones shipping faster than you can differentiate - and it means fewer buyers, because the customers who used to purchase now build the slice they need themselves. The old business lived in the gap between those two: enough friction that rivals couldn't flood in, enough that customers wouldn't bother building it themselves. Remove the friction and both give way at once.
The gap closes. Your competitor is now your own customer with an agent and a markdown file of business rules — it closes on you from both sides.
The second: the demand for software goes agentic. The buyer, the user, and the integrator all become agents. An agent doesn't mind using 30 networks instead of 1, doesn't need to relearn an interface, reads every search-result page rather than the first, and benchmarks a vendor's claims directly instead of trusting its brand. Gartner projects that by 2028, 90% of B2B buying is agent-intermediated - north of $15T through agent exchanges.1
So why does this prediction land when "SaaS is dead" and "everyone becomes a free agent" failed every prior cycle? Each earlier tooling revolution - PC, internet, AWS, Shopify, Stripe - collapsed one dimension of the cost of operating a business while every other wall stayed up. The binding constraint was bundling, not intelligence or ambition. You still needed the accountant, the support team, the marketer, the designer. Agents are the first technology that compresses every dimension of the operational stack on a single S-curve - a difference in kind, not degree.
Incumbents can't simply respond, and the reason is structural rather than a failure of nerve. Rearchitecting means admitting the constrained data model - the rigid schema of objects and dropdown fields that forced business reality into forms, discarded everything that didn't fit, and generated hundreds of billions in value doing it - was wrong.2 Bolting AI onto that model makes things worse: the reasoning engine can only see the thin slice a sales rep typed into the boxes, not the forty emails where the deal actually happened. You get confident answers, blind to 80% of the signal. That gap is the startup opening.
And I agree, this sounds like doomer logic you hear on X. It might be. I've also heard "death of SaaS" thrown around so often it makes me roll my eyes. It does feel overstated because when some VC says SaaS is dead, they just swap the word "SaaS" for "AI agents" and sell the same thing: same CAC, same churn, new logo.3 I think this sentiment is right, and I also think it leaves the thesis untouched. Software still arrives as a metered service you log into or call; the subscription survives. What dies is the defensibility that justified the multiple.
The Moat Ledger
The verdict on every classical software moat. If you're building something right now, read the first table as an obituary and the second as a to-do list.
Moats That Die
| Dead moat | Killed by | Canonical example |
|---|---|---|
| Learned interfaces ("we made it usable") | Natural language; agents have no learning curve | Years of Salesforce-admin training replaced by chat |
| Encoded workflows / domain logic | Domain logic is now a markdown file a non-engineer writes | The consultant-configured workflow, regenerated in an afternoon |
| Public-data access & processing | LLMs parse catalogs, laws, filings natively | TurboTax: the tax code is public; the AI reads it for you |
| Switching costs (human retraining) | Context lives in files the customer owns; plain-English operation | Cursor → Claude Code → Codex, switched within an hour. Highest PMF category in AI, least moat - empirically |
| Distribution / brand recall | Agents do 20 searches × 20 pages and file a memo per vendor | DocuSign: a literal verb, still unmoated among identical clones |
| Branding on verifiable work | The agent just runs the benchmark itself | Test Turbopuffer's QPS claim directly vs Pinecone; ads irrelevant |
| Network effects built on listing/search friction | Agents multi-home effortlessly on both sides | Etsy: agent maintains listings on Amazon, Etsy, Shopify at once |
| Economies of scale in software | Marginal cost of software → 0 | Jira: agents doing the work build their own coordination system for pennies |
| Bundling | "The agent IS the bundle" | No reason to buy the all-in-one suite |
| Talent scarcity | Experts encode their own methodology directly | Competition goes from 3 incumbents to 300 startups |
| Seat-based pricing | Value no longer proportional to human headcount | Most vendors projected to refactor to consumption/outcomes by 2028 |
Every row in that table was really a tax on human friction. The clone economy is the tell: when a developer clean-room copies a shipping AI coding tool in a day, the lesson is that the feature was never the moat. If a competent team can rebuild your core over a weekend, you had a feature, not a company. You can replicate the feature in an afternoon and still not replicate the distribution, the trust, or the data.
Moats That Survive
| Living moat | Why it holds | Condition |
|---|---|---|
| Proprietary data + self-improving loop | As models commoditise, scarcity shifts from model to data. The full cycle - watch real customers interact, auto-improve the product live - is the strongest version | Only if governed: discoverable, access-controlled, auditable. Data alone isn't the moat; the loop is |
| Trust on hard-to-verify work | The verification loop stays expensive even when software is free. You can't run an end-to-end "did we get hacked" eval | CrowdStrike survives; vector-DB benchmarks don't. Ask: can an agent verify this output cheaply? If yes, trust commoditises |
| Transaction embedding | Payments, settlement, disputes, fraud, refunds, compliance - agents route through whoever handles these | Strengthens when transaction data feeds back into the governed data loop |
| Network effects built on trust (not friction) | Reviews attached to identity, dispute resolution at volume | Airbnb holds (reputation is the asset); Craigslist doesn't |
| Regulatory standing | GDPR/CCPA/FDA are structural; new bar: every AI decision traceable through lineage to source | The caveat: a moat that is only bureaucracy gets valued low once everyone can see that's all it is |
| Capital & physical infrastructure | AI didn't make electricity, GPUs, warehouses, or trucks cheaper | Amazon, data centers, logistics - untouched |
| System of record → system of intelligence | The record doesn't die; it becomes the agent's governed contextual memory | Must own the memory layer, not rent it |
| Velocity of the learning loop | The one moat you build rather than own: how fast model improvements + failures + edge cases become better product | "The advantage is not the product snapshot you have today" - whoever's loop compounds fastest wins |
What's left standing: trust, payment, atoms, proprietary loops, and speed. Everything scarce is either physical, relational, or kinetic.
The sharpest pushback on this comes from the people who claim model moats are dead, and data moats are mostly a mirage - that the only real moats are closed-loop execution speed, distribution, and trust. I mostly agree, and it's why every surviving row above carries the same condition. Data at rest is inert. The moat is never the data; it's the loop that turns the data into a better product faster than a competitor can copy the output.
What Stops Existing
The ledger names the mechanisms; this is the casualty list.
-
The thin vertical SaaS app. Its value was interface + encoded workflow + integrations - all three now free. It dies from three directions at once: horizontal agents absorb its use-case, hundreds of startups clone it, and its own customers build it in-house. On the order of $1T of software value already marked down. The tarpit list grows: thin LLM wrappers, "AI PwC", AI hedge funds, GEO/AI-SEO me-toos - and the quiet failure mode behind them all, drifting into a dev shop without noticing.
-
The AI-feature bundle. "SaaS + AI features" is the sunk-cost trap wearing a new shirt. AI on an impoverished data model is worse than no AI.
-
Seat-priced knowledge-work software. The seat is a proxy for a human, and the human count is shrinking. Aaron Levie put the pricing problem cleanly on the MAD Podcast: one customer might do all its work with a single agent, another might split the same work across a thousand, so pricing per human seat becomes nearly impossible.4 The migration to consumption and outcomes is already happening - Salesforce charging per resolved action, HubSpot halving its per-resolution price,5 SAP moving to consumption billing.
What nobody has cleanly solved is who absorbs the variance: when a hard task burns 10x the tokens, seat pricing protects the buyer, consumption makes them flinch at the invoice, and outcome pricing means you eat the bad month. The pricing model is unsettled (the death of the seat is not).
-
Coordination software for humans. Ticketing, project management, internal wikis as products - agents doing the work coordinate themselves. (Jira is the named casualty here...)
-
Brand-as-distribution. Being the verb no longer wins when the buyer reads everything.
-
The Organization Man - and with him, the large-headcount firm as the default container for work. Not extinct, but the corporation stops growing as a share of economic activity. Entry-level white-collar hiring contracts first. The median new company trends toward 1–3 people. The 60-year era of mass employment by giant firms was the anomaly; small operators were the historical norm, and we revert. Directionally confirmed already: record US business formations, record revenue-per-employee, no verified solo unicorn yet.
-
The pure app-layer startup as a durable asset class. Petersson's inverted U: value rises as engineering effort creates capability, then falls as the next model release makes that engineering obsolete.6 The soft landing is acqui-sale to a lab that wants the user base or the data - more on this fight in The Disagreement, below.
Seven Shapes
Seven durable company shapes, strongest moat first.
1. The Vertical Operator
Jason Shuman frames the shift in one line: vertical SaaS competed for IT budgets, vertical AI competes for labour budgets.7 That's a different and far larger pool - you aren't fighting for a line item in the software budget, you're replacing a cost centre. Todd Saunders describes the same arc from inside the category: vertical SaaS 1.0 digitised a workflow and charged per seat, and the moat was simply being the only vendor willing to sell to funeral homes or HVAC contractors; the operator graduates that system of record into a system of intelligence.8
The operator runs the operation and sells the outcome. Waymo doesn't help you find a ride - it operates the ride. The next Airbnb doesn't list homes - it operates them. One early-stage operator I've heard of went $0→$400k ARR in 2 months running short-term rentals end-to-end, with its first customer live within weeks and zero human involvement in the operation.
The playbook has three moves: capabilities not features (fixed agent-harness infrastructure plus a dynamic per-customer logic layer the agents grow on top), forward-deployment (build inside customer ops, abstract every edge case back into the core), and choose customers who make the product smarter.
The inverse of that last one matters just as much: first customers who only want a chatbot don't make the product smarter - it's the wrong environment to learn in. The operator ends up owning the ops-data exhaust, the service-trust relationship, and the transaction. This is where "services > SaaS" cashes out - services-margin business, software-margin headcount.
2. The Trust Layer
The missing layer: how does a buyer-agent discover, trust, sign up for, and pay a vendor with no human in the loop? Rails are shipping (ACP/ChatGPT Instant Checkout, Google AP2's Human-Not-Present mandates, Visa Trusted Agent Protocol, Mastercard Agent Pay, x402) but everything above the rails - reputation, verification, agent identity and authorisation, dispute resolution - is greenfield. Picks-and-shovels for the entire agent economy; the "$15T through agent exchanges" toll road.
The consensus among people building on these rails is that the rails are the easy part. The stack so far: x402 for execution, AP2 for authorization, ACP for commerce - and the thing still missing is enforcement: who can spend, how much, under what policy, with what audit trail. Put another way, the harness is the failure point, not the rail.
Dispute resolution is its own greenfield - projects like Genlayer are building the judge layer for when two agents disagree on whether the terms were met, and identity standards like ERC-8004 are racing to fill the reputation gap. The rail moves the money; nobody yet owns the layer that decides whether the money should have moved.
3. The Data Flywheel
A product whose usage itself generates data no competitor can buy - especially live, multi-party interaction data - and which auto-improves from it. X is the cleanest example: every conversation on the platform is training data for Grok that no rival can license. Usage compounds; features don't.
4. Agent Governance
This is Saker's prescription for the enterprise side:2 composable data architecture, unified catalogs, fine-grained agent access control, automated lineage, open formats. It sells because roughly 70% of enterprises blame data governance for failed AI scaling, and "zero-trust for agents" becomes mandatory. If the enterprise governs its own data, the enterprise becomes the platform - this category arms them.
5. Atoms and Licenses
Where the cost was never software: energy, compute, logistics, biotech, licensed operations. AI is internal leverage; the moat is atoms, capex, or the license. Physical means safe.
6. Small-Operator Tooling
If US businesses go from 33M toward 60–80M, mostly 1–3 person agent-native firms, someone sells them incorporation, tax, compliance, banking, insurance, and agent-fleet management. Stripe Atlas / Clerky were the prototype; the agent-era versions are 100× bigger. The regulatory burden on small operators is the product opportunity, not just a risk.
7. Software as Substrate
Petersson's survivor clause: agents could rebuild any tool from scratch, but compute costs make existing platforms the cheaper substrate. Software that survives here is API-first, priced on usage, with the deepest nonstandard edge-case coverage in its domain - the dirty integration work nobody wants to redo. The weakest moat on this list - it survives on cost-efficiency, not defensibility - but it's something.
The subtle version of this moat is positional. When agents orchestrate work through tool calls, there's value in being the canonical authority in the agent's calling graph: the default, trusted node it routes to for a given job. Boring infrastructure that everything calls wins again, simply by being the obvious thing to call. The same pull shows up on the human side, for as long as humans stay in the loop: developers won't leave Jira to learn another tool, so whoever lets them query it from wherever they already sit owns the transition - right up until the ledger's verdict on coordination software catches up with it.
The Winning Firm
Across all seven: tiny, agent-native, forward-deployed, outcome-priced. A handful of humans doing taste, trust, and edge-case judgment; agents doing everything else. Value-per-employee ran $1M (1990s) → $5M (2010s) → $50M+ (2020s) and keeps steepening. The minimum viable team for $10M ARR trends toward one.
The Disagreement
The bear case is vertical AI apps are doomed - horizontal agents from the labs absorb every simple vertical, batch by batch, with each model release. Only defensible-resource verticals survive, and the customer's relationship migrates to the horizontal agent.
The bull case is the short-term-rental operator example: vertical AI is compounding right now - but only when it's an operator, not an app.
Both are right, about different objects:
Vertical software dies. Vertical operators win.
We'll see the death of the vertical app - the workflow wrapper, the thin layer. The operator demonstrates the survivor: the company that owns the operation has nothing for a horizontal agent to absorb, because its moat was never the software.
The evidence this year supports the split verdict: verticals that go deep into operations mint revenue (Harvey crossed $300M ARR this spring,9 Sierra past $150M on pure outcome pricing5) while the labs visibly absorb shallow use-cases (ChatGPT apps, agent mode, Workspace Agents). The inverted U is real - for apps. Operators aren't on it.
Each frontier release flips simple verticals to horizontal. The app collapses; the operator keeps compounding — its moat was never the software, so there’s nothing for a horizontal agent to absorb.
One thing to point out in the soft-landing story. I said the graceful exit for a vertical app is an acqui-sale to a lab that wants the users or the data. So far this year another pattern has shown as well - legacy vertical-SaaS incumbents buy the AI startup, keep their own distribution, and bolt the model on. It doesn't rescue the impoverished-data-model problem - but it means the acquirer isn't always a frontier lab, and distribution still counts for something on the way down I guess.
The second-order agreement everyone shares but nobody headlines: the model layer captures the most value of all, with more than 80% of AI valuation sitting in a handful of firms. Everything in this thesis is about winning the layer above the models and below the labs' attention.
What To Watch
- 2026–27: vertical→horizontal adoption flips in simple verticals with each frontier release; seat pricing visibly refactors; US business applications keep setting records - already on pace at 2.55M applications Jan–May 2026, up 17% year on year.10
- Mid-2026: sell-side and consulting frameworks start openly sorting software names into survive/die buckets.
- By 2028: agent-intermediated B2B buying closes on Gartner's 90% call; Human-Not-Present payments move from spec to volume; first verified one-person $100M+ ARR company; >50% of new internet businesses are 1–3 person teams.
- Kill signals for the thesis: agentic autonomy stalls at human-approves-checkout for years (trust layer never matures); regulation pins agents to enterprise-only deployment; model progress plateaus hard enough that encoded-workflow value returns.
The Five Laws
- Software was a tax on human friction. Agents remove the friction, so they remove the tax. Any business that was the tax dies.
- The moat moved from the code to the loop. Proprietary data that compounds through operation, converted into product faster than anyone else - velocity is the only moat you can build from zero.
- Sell the outcome, not the tool. Operate the ride; don't build the app that hails it. Whoever owns the outcome owns the data, the trust, and the transaction.
- Everything scarce is physical, relational, or kinetic - atoms, trust on unverifiable work, position in the money flow, regulatory standing, speed. Everything informational is free.
- The firm shrinks; the economy fragments; the infrastructure consolidates. Build the tiny operator, or build what millions of tiny operators and their agents stand on. The worst position is the middle: a big-headcount company selling human-friction software to a shrinking pool of big-headcount customers.
The one-line test stands: when interfaces, workflows, and public data are free, what do I still own that a customer can't get anywhere else? One sentence, or pass.
References
Footnotes
-
Gartner: "AI agents will command $15 trillion in B2B purchases by 2028", unveiled at Gartner IT Symposium/Xpo 2025, via Digital Commerce 360, November 2025. ↩
-
Rob Saker: "AI is Eating Enterprise SaaS", Medium, February 2026. The enterprise-data half of the argument: constrained SaaS data models, why bolted-on AI amplifies the problem, and governance as the surviving sell. ↩ ↩2
-
"The death of SaaS has been overstated" - replace the word "SaaS" with "AI agents" and VCs are selling the same B2B software. ~ Jeff Morris Jr., X ↩
-
Aaron Levie on The MAD Podcast with Matt Turck: "State of Enterprise AI 2026: Tokenmaxxing, Rise of Headless, and AI-Proofing Your Job", May 2026 - on headless software and what replaces per-seat pricing. ↩
-
Salesforce Ben: "Huge Agentforce Pricing Shift: Salesforce Introduces Pay-Per-Resolution"; SaaStr: "HubSpot Switching AI Pricing From Per Use to Per Resolution" - HubSpot's Breeze Customer Agent went from $1 per conversation to $0.50 per resolved conversation; the same piece puts Sierra at $150M+ ARR on pure outcome pricing. ↩ ↩2
-
Lukas Petersson: "AI Founder's Bitter Lesson", lukaspetersson.com, January 2025. The inverted-U argument: value from engineering effort rises, then the next model release makes that engineering obsolete. ↩
-
"Vertical SaaS competed for IT budgets. Vertical AI competes for labor budgets." ~ Jason Shuman, X ↩
-
"The moat was being the only vendor willing to sell to funeral homes or HVAC contractors" - vertical SaaS 1.0 as a system of record, graduating into a system of intelligence. ~ Todd Saunders, X ↩
-
Sacra: "Harvey revenue, valuation & funding" - Harvey hit ~$300M ARR in May 2026, up from $195M at the end of 2025. ↩
-
US Census Bureau: "Business Formation Statistics", monthly data releases, 2026. ↩