Every SaaS founder is running the same two-question stress test right now.
Can someone rebuild this at one one-hundredth of the cost with AI-native capabilities? Will customers ask an AI head to do the job and skip the application entirely? Is there any value to monetize sitting between the AI head and the transactional backend?
The honest answer for a lot of categories is yes, yes, and not much. Reporting is over. Most analytics is over. Horizontal workflow automation is mostly over. The argument that customers won’t build it themselves because of reliability, availability, security, and lifecycle (who patches this thing in three years?) holds up for some categories and falls apart for others.
The interesting question isn’t whether the apocalypse is real. It is. The interesting question is which axes actually predict survival.
Six of them matter: total addressable market, what the software does and what happens when it’s wrong, whether the AI has corpus and data to bootstrap from, the buyer, the user, and distribution. Those six explain most of what’s about to happen.
TAM changed before anyone noticed
Talk to venture capitalists and you’d think the only software investment worth making is writing the entire fund check to Anthropic.
There are now three stacked markets, not one. The software budget, services budget around the software (building and customizing), and the labor budget for the people doing the work the software supports. That third layer is the big one, and it’s the layer after which AI (agents) is going.
The TAM for legal software isn’t legal software. It’s lawyers. The TAM for developer tools isn’t developer tools, it’s developers. Software development tools / devops are roughly $150 billion. Software development services are roughly $550 billion. That’s short of a trillion dollars in a single category (not taking into account the users of the software that can be replaced by agents), worth the attention of any AI platform aiming at trillion-dollar market caps. The TAM for customer support software isn’t ticketing seats, it’s the support team. The TAM for sales automation isn’t CRM licenses, it’s the people working the pipeline, e.g. SDRs to start.
When AI can replace the people and the software in one shot, the whole stack collapses and everyone in the middle becomes collateral damage. The SaaS apocalypse stops being about software vendors at that point and starts being about knowledge-work labor categories.
There’s a second-order effect worth naming. A large TAM used to be almost purely good news. It meant the market was worth chasing. Big TAM now means exposure, too. The bigger the labor pool, the more documented the work, the more obvious the target. Big TAM is opportunity and risk in the same line item. After all, the AI companies need massive TAMs to go after to drive their own valuations and IPOs (coming to a market near you).
What happens when the software is wrong?
That axis matters most because it gets at the customer’s real tolerance for risk. The frame is simple: what happens when the software gets it wrong?
If the answer is a bad summary or a stale dashboard or a misfiled meeting note, the category is exposed. The cost of failure is low, so the cost of trying something cheaper is also low. If the answer is that money moves to the wrong place, inventory ships twice, a warehouse rejects an order, a customer’s package vanishes, or a company fails an audit, the category lives in a different universe of buyer behavior.
That single question sorts most of the software.
Analytics and reporting are probably dead. Why does anyone need a reporting application if an agent can answer the actual question and produce a custom report scoped to the decision in front of them? Dashboards on top of dashboards, with analysts hired to explain why the dashboards disagree, was a bad equilibrium before AI showed up. AI just makes the bill come due faster. The data warehouse survives, the stuff on top doesn’t.
Productivity apps are the most exposed category. Project management, scheduling, note-taking, internal wikis, and lightweight workflow tools are all thin wrappers around work the AI already understands. A small team can rebuild the core of a project management tool in a weekend (whether they can sell it is a different question). The technical moat is gone.
Systems of record are stickier. CRM, HR, financials, support ticketing, anything that holds history. The AI can replicate the schema. It can’t replicate ten years of customer notes, deal history, audit trail, or permissions. Customers don’t swap systems of record on a Tuesday because the cost of migration is real and the cost of losing institutional memory is worse. Systems of record also sit in a strange position with AI agents: every agent needs somewhere to read from and write to. An AI SDR has to log notes somewhere. An AI support agent has to update a ticket. An AI finance agent has to write into a governed system. The application interface may die. The underlying record probably doesn’t, at least not soon.
Transactional systems are the stickiest. Payments, core banking, tax, orders, settlement, inventory, fulfillment. Anywhere money or its equivalent moves, the cost of being wrong is denominated in dollars per error, and customers behave accordingly. A hallucinated shipping address becomes a chargeback. A double shipment is money out the door and a customer who never comes back. Customers want vendors with SLAs, contracts, and someone to sue when it breaks. People will vibe-code a departmental time-tracking app, but they get much more careful when the system in question moves money or equivalent.
Development tools split into two halves. The way software gets built is going through the most consequential change since cloud (some would argue since the move off the mainframe). The front end of the development process is exposed. The back end, monitoring and security, seems a much safer place. One look at DataDog’s stock and you can see the market is already pricing this in.
Infrastructure has more time. Datastores, data warehouses, application monitoring, the underlying compute and storage layer. People will take longer to roll their own database with replication, backup, disaster recovery, and the rest. A lot longer.
Some categories aren’t really SaaS at all. They’re networks wearing SaaS interfaces. Telecom is a good example: connecting to every carrier on earth is the kind of complexity AI doesn’t want anywhere near its training pipeline. Think about Twilio and then imagine all those legacy telecoms that it can connect an application to. Good luck replicating that with AI. How about Stripe and payments?
No system is fully immune from five developers in San Francisco deciding to replicate it. The bigger question is whether what they replicate is the same thing or a rethink of the entire solution from top to bottom (where solution equals software plus people plus process). The interesting AI-native companies are collapsing software and services categories.
Corpus is destiny, until it isn’t
If the work is documented in books, contracts, RFCs, training materials, and a few thousand blog posts, the AI is most of the way there. Legal is the textbook case. The playbooks are written down to the last brief. The cases are in the public domain. CRM has forty years of methodology, sales books, marketing automation playbooks, and customer support scripts already in the training corpus. AI consumes and repurposes all of it.
Data accessibility is the second half of the same question. If the inputs sit on the open web or behind APIs that respond like the documentation says they will, the AI has what it needs. If the data sits in physical systems, regulated environments, or formats that resist scraping, the AI has to be paired with something else. Most legal cases in the US are online (so much data for the AI to learn from). Logistics network data isn’t (the logistics companies barely have it themselves).
When the corpus is thin, the AI has to be taught. That changes the cost structure, the go-to-market, and the product architecture. Industrial workflows, fulfillment, manufacturing, and anything that lives in tribal knowledge inside a few hundred operators’ heads doesn’t bootstrap from public data. The knowledge exists, however, it just doesn’t exist in a form the model can ingest.
Customer-specific history by itself isn’t a durable moat either. The data can be extracted, transformed, summarized, and replanted. The stronger moat is knowing what to do with the data when reality refuses to follow the documentation. That part the public corpus doesn’t teach.
The buyer sets the clock speed
Software people talk about technology adoption as if it’s a meritocracy. It isn’t. The buyer sets the speed of the market.
A Silicon Valley startup swaps tooling on a Tuesday afternoon. A state government procurement office on a five-year cycle isn’t ripping out its case management system this quarter. A regulated bank, a healthcare network, a retailer heading into peak season, none of them are moving fast.
Selling a competent SaaS solution to state and local government is a reasonably safe business for the next few years. Selling productivity tools to Y Combinator startups is a knife fight that’s already happening. Slow buyers aren’t permanently safe.
The user matters as much as the buyer
Developers are ruthless. They evaluate tools weekly, they read changelogs, they switch on the strength of a Hacker News thread. If a category serves developers, the AI-native competitor gets distribution fast because developers will try anything that looks better.
Business users want ease of use. Ease increasingly means typing into Claude or ChatGPT and getting an answer with no app in between. The app that survives is the one sitting behind the AI head, doing the work the AI can’t do alone – no wonder Mark Benioff said Salesforce is going “headless”. Apps that try to keep the human in their own interface are racing against a UX that’s already won – the interface is the chat box!
A lot of SaaS vendors are confused on this point. They think the application is the product. In most categories the application was just the interface to the product, and the product was the work getting done. AI attacks the work first, then the interface, then eventually the underlying system if it can. Software only fully disappears when its coordination function disappears too, which is a much higher bar than replacing a screen.
Distribution is the most underappreciated axis
AI makes products dramatically easier to build. It doesn’t make customers easier to acquire. A two-person team can now build a credible demo of almost anything, and so can everyone else. When product creation gets cheaper, distribution gets more valuable.
The race that’s actually happening isn’t legacy SaaS versus AI-native startups. The race is incumbents-with-AI versus startups-with-AI versus customers building internal tools with AI. The incumbent comes with the buyer’s email address, the procurement relationship, the security approval, the signed MSA, the admin console, the workflow history, the integration footprint, and the renewal date. None of that is glamorous. All of it matters.
Salesforce is past $40 billion in revenue and announcing AI products faster than most of the startups attacking it (of course, Salesforce has to actually deliver on those promises). Their distribution will ensure that they get those products into the hands of the customers at the customers pace. Even better, distribution lets the incumbent chase the services and labor side of the TAM directly. Salesforce going after the AI SDR market is the cleanest example.
A product can be technically replaceable and commercially durable at the same time, and the gap can last years.
What’s not defensible
Three properties in combination predict death. TAM that includes labor. Deep corpus the AI has already absorbed. Accessible data. Legal hits all three. Big chunks of horizontal SaaS hit all three. So does most of analytics. A category with all three is finished. A category with one or two might survive, depending on the other axes.
By the three-condition test, CRM should be a graveyard: giant TAM, deep corpus, mountains of accessible data. Systems of record get a stay of execution. Two things explain the gap. The first is distribution. The second is that systems of record are the source of truth the agents themselves depend on. The AI SDR writes its notes somewhere. The AI account manager reads deal history before doing anything useful. Agents make systems of record more valuable in the short run, not less, because the volume of writes goes up and the cost of an inconsistent source of truth goes up with it.
The longer-term question is whether AI memory replaces the database underneath. That’s a five-to-ten year question, not a two-year question, and it depends on whether enterprises trust the agent to be both the worker and the system of record at the same time. Bet against that happening fast.
The squeeze
The SaaS apocalypse is a middle squeeze.
The very top of the market survives because the platform companies have real structural advantages underneath them. Microsoft has an infrastructure moat: Azure, Windows, the productivity rails every enterprise already runs on. Google has a data moat: search, ads, the knowledge graph, the largest aggregated view of human intent on the planet. Amazon has a distribution moat: AWS underneath half the internet, and a retail and logistics network that physically moves goods across most of the developed world. Those advantages compound in the AI era because AI needs compute, AI needs data, and AI needs distribution. The Mag 7 winning the AI era is the least surprising prediction in the room. The one place they’re exposed is the data center build-out, where the costs are astronomical.
The more interesting question is who else survives. Incumbency without a structural moat is a different story. Healthcare holds up in places where the moat is real: EPIC and CERNER have spent decades building integrations, compliance posture, and the brutal switching costs of clinical systems. A horizontal SaaS vendor with a customer list and no comparable moat is buying time, not winning. The customer list goes stale, the AI features get matched, agents take over people’s work and a competitor with a cleaner approach takes the install base.
The long tail of vertical SaaS also survives. Small, focused, AI-native vendors in transactional categories or vertical niches that the public corpus doesn’t cover. There’s nothing for a new entrant to replicate because the moat is domain knowledge some of which doesn’t exist online, yet.
The middle dies. Companies between $10 million and $1 billion in ARR, sitting in horizontal categories with documented workflows (corpus), accessible data, and no transactional teeth, are in serious trouble. Many of them are owned by private equity and are structurally unwilling (or unable) to sacrifice the cash-flow business to pivot to AI-native architecture with agents. They’re too small to have the distribution advantages of the giants. They’re too big to rebuild themselves casually. They’re too undifferentiated to defend. The next twenty-four months will sort out which of them find a real pivot and which become collateral damage.
If a company’s product is reporting, that company is finished. If it’s horizontal workflow automation, the same applies. If it’s a system of record in a category the AI can rebuild from public corpus, it has a window: long enough to matter, short enough to require a real plan.
Where Pipe17 sits
Pipe17 lives at the boundary between the digital world and the physical world. Order operations is the meeting point, and it fails the AI corpus test in all the right ways.
Logistics is poorly documented in public. The workflows aren’t in books. The APIs don’t behave as documented. The tribal knowledge sits inside operators, 3PLs, merchants, and the veterans who’ve been running fulfillment since the early 2000s. An AI trying to learn order operations from the open web learns nothing useful.
Order operations is also transactional. Orders are money in motion. Inventory is money in motion. Fulfillment is money in motion. A hallucinated shipping address becomes a chargeback. A double shipment burns through inventory and margin. The cost of being wrong is denominated in dollars per error, and customers want vendors with contracts and accountability on the other side.
The value in order operations isn’t on the happy path. The happy path is easy to draw on a whiteboard: an order comes in, inventory is available, the warehouse ships it, the customer gets a tracking number. The complexity in there is a killer,, e.g.. a fulfillment location rejects the request, a carrier fails, a marketplace sends incomplete data. The ERP and the commerce platform having inconsistent data that needs reconciliation, a warehouse accepting an order that it later cannot ship, a customer changes the order after it has already started moving, a SKU mapping is wrong since an API changed, a bundle partially allocates, a 3PL behaves differently than its documentation says it will, etc. . Commerce operations is messy, physical, fragmented, and operationally unforgiving.
The useful frame: some SaaS isn’t really SaaS at all. It’s a network wearing a SaaS interface. Twilio wasn’t valuable because of the dashboard. It was valuable because it made the telecom network programmable. Stripe wasn’t valuable because of the merchant UI. It was valuable because it made payments programmable. The same pattern applies in commerce operations. The world doesn’t need another workflow screen. It needs a programmable, reliable execution layer between the places orders originate and the places orders get fulfilled.
In an AI era that becomes more valuable, not less. Agents generate demand. Agents make recommendations. Agents take instructions. At some point something has to execute, and that something has to know which system is authoritative, which warehouse can ship, which inventory is real, and what to do when the answer isn’t clean. Pipe17 has to live in the execution layer, not in the interface.
The call
Everyone will use AI to build. That part is settled.
The losers are predictable. Productivity apps with documented workflows and accessible data. Categories where the TAM is mostly labor and the AI can take the labor too. Middle-tier horizontal SaaS that can’t outship the incumbents and can’t outmaneuver the startups. Static reporting in any form.
The winners are platform vendors with real structural moats (Microsoft on infrastructure, Google on data, Amazon on distribution). Systems of record the agents themselves depend on, for a window measured in years. Vertical specialists in domains the AI can’t bootstrap. Transactional systems where accountability is part of the product. Networks wearing SaaS interfaces.
Order management is money in motion. AI doesn’t eat that category. It joins it.
