Channel|Quarterly SaaS Winners and Losers Tracker 3Q26

Hi,

We’re launching a new series on how advances in AI are affecting SaaS companies and their growth prospects. For this first issue, we spoke with 31 companies, and we will share below their views on who are the winners and losers in the agentic AI era.

We’d welcome your feedback and suggestions on our coverage. If there is a SaaS/ software name that you are interested in learning more — share their names here (it takes about 30 seconds). Your input will help shape our research for the next quarter!

Btw, we’ve also prepared a downloadable slide deck if you prefer that format. Contact sales@funda.ai if you have any questions on our institutional offerings.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

Executive summary

1. Today, the systems of record are winning. Most interviewees using Salesforce or ServiceNow, and every one using Snowflake or Databricks, reported a rising bill for the 4 software platforms for three reasons: enterprises buy the vendors’ native AI agents because they sit on the data and inside the workflow; third-party agent access is billed per action; and agents query data far more often than people do, on platforms that bill by compute. Alongside them, the vendors of agent infrastructure are winning too: agents need gateways, agent identities, extra security and audit trails, a demand that did not exist two years ago. The losers are single-purpose tools, glue software, BI, knowledge and marketing tools, for one of two reasons: a coding agent rebuilds them in days, or employees no longer open them because they ask an assistant instead.

2. Enterprise software carries three layers of cost: building it, running it, and acquiring the domain knowledge behind it. Longer term, the systems of record will face sharper competition from AI-powered in-house builds and AI-native vendors, because AI has sharply cut the cost of building software and partly cut the cost of running it. The early signs are already visible: of ~30 enterprises, 15 have replaced single-purpose tools with in-house builds, most of them in recent quarters, and some have gone further to BI front ends and non-core modules such as Salesforce Marketing Cloud; AI-native challengers now exist in every category. Two incremental changes drive this. (1) AI has made software cheap to build, and Opus 5.5 now enables affordable long-running development with minimal human intervention, so a small team can build what used to need a vendor; the same agents are starting to take on running work such as refactoring and upgrades, though interviewees see less help there than on building. (2) SaaS has become much more headless and open to third-party agents, which erodes its interface moat and lock-in effect over time.

3. However, AI cannot cut the cost of acquiring domain knowledge, nor will it help replace the accountability that incumbents provide in running the software, so the incumbents still hold an advantage even as competition sharpens. No interviewee is replacing a core system yet: risk sign-off takes years, vendors carry pooled industry knowledge, regulators expect a vendor to answer for the system, and a failed migration costs more than the license it saves. For Salesforce, ServiceNow and Snowflake, the sharper long-term competition is manageable: what they sell is years of industry experience and an understanding of customers’ workflows, which neither an in-house team nor an AI-native start-up has.

Part 1 · Winners and losers in the agentic AI era

Today’s winners are the systems of record. They win because enterprises buy their native AI agents (1.1), third-party agent access is billed per action (1.2), and agents consume far more data than people do (1.3). A second group of winners is the vendors of what agents newly require: they win because agents need gateways, identities, extra security and audit trails, a demand that did not exist two years ago, and because payment and advertising platforms that put an agent behind a fee earn more per transaction (1B). Today’s losers are the single-purpose tools, glue software, BI and reporting, knowledge tools and marketing tools that agents can rebuild or bypass: they are easy to rebuild in-house, or employees no longer open them and ask an assistant instead (1C).

1A. The systems of record are net winners

Most interviewees using Salesforce or ServiceNow, and every one using Snowflake or Databricks, reported a rising bill; the few flat or falling cases are enterprises that refused an AI add-on on price or rebuilt a non-core module (2.1).

What we heard

“We reduced our ServiceNow seats by about 16%, but we’re spending 6-7% more on the AI Control Tower and 13% more on AI agents within ServiceNow.” — CIO, US diagnostics company

“Usage comes down before contract spend comes down.” — Enterprise analytics director, US healthcare group

The one losing the job is the human, not the software.” — Technology executive, US regional bank

1.1 Reason one: enterprises buy the vendors’ native AI agents, because they sit on the data and inside the workflow

Enterprises are buying the systems of record’s own agents and AI tools (Agentforce, Now Assist, AI Control Tower) rather than building equivalents, because the vendor’s agent sits next to the data, runs inside the existing workflow and permissions, needs no development or deployment, and produces a result the business can measure. These add-ons are priced per action or per token, so the bill rises with use.

What we heard:

“Your AI agent has to be close to where the data is. That’s where you get more value.” — Chief Data Officer, Australian healthcare distributor

“The agent takes all the data and summarizes it: you’re talking to this customer today, here are the key points to bring up. And it takes notes.” — CIO, US diagnostics company

“Spending on ServiceNow is going up because of Now Assist. We have been incrementally buying more of their services.” — Digital lead, global pharma company

“We just didn’t hire the two more people we would have needed to handle the volume of calls.” — Chief Data Officer, Australian healthcare distributor

1.2 Reason two: third-party agent access is billed per action, so work that bypasses the interface still earns

Enterprises want one AI layer that works across systems and lets them choose the model, so many build or buy agents outside the suite and connect them through MCP and APIs. Vendors like Salesforce, ServiceNow and Workday have accepted this and charge for it per action. Agent traffic that never opens the interface still generates revenue for them.

Our illustration of what this is worth today: at Salesforce’s published rate of roughly $0.10 per agent action, an agent handling 2K tickets a month with about five actions each would bill about $1K a month. The revenue from third-party agent calls is real but still small; what matters is that it grows with every agent added and does not depend on seats.

What we heard

“If I own the harness layer and the context, switching the LLM, or even the backend system, is relatively easy.” — Head of Data and AI, Singapore telco

“ServiceNow is a SaaS vendor; they’re not going to give you the information for free. The pricing model is something we still need to study.” — Head of Data and AI, Singapore telco

“Everything is becoming headless.” — Digital lead, global pharma company

1.3 Reason three: agents query data far more often than people do, and data platforms bill by compute

Data platforms are the cleanest beneficiaries. Three sources drive the growth: agents querying through MCP, data engineers using Cortex to ship more with the same team, and unstructured data such as call and chat logs being loaded for agents to use. None of it depends on headcount, and no interviewee has cut usage or spend.

What we heard

“Any time we build an agent, it gets a Snowflake MCP to go and query the data.” — Enterprise analytics director, US healthcare group

“An agent can actually execute dozens of queries in a single workflow.” — Head of Product, Data and AI, US e-commerce retailer

“The more machine-learning and AI use cases we build, the faster our Databricks bill goes up.” — CTO, European telco

1B. The second group of winners: vendors selling what agents newly require

Agents create four kinds of demand that enterprises did not have two years ago, met partly by new tools and partly by new modules from existing vendors: a gateway to route and meter model calls, with orchestration, observability and evaluation tooling; an identity and least-privilege access for each agent; more security capacity, because attackers use AI too; and audit trails and data governance, because agents touch more systems than people did. Each is small, but it is bought as shared platform cost and grows with every new use case. A fifth group is not software at all: PayPal and Pinterest put an AI agent behind a fee-based service and earn more per transaction when the agent performs better.

What we heard

“We’re building this AI gateway. Those are the tools we can leverage for the second, third and fifth use case. That’s why about 60% of the budget goes to the common platform.” — Consultant, US pet insurer

“Who’s going to monitor the behavior of all the agents? I don’t know yet.” — Chief Digital and Technology Officer, UK local authority

Up close: six winners

What we heard

“We can manage what a model is allowed, and what kind of workload matches which model. Per model, per kind of task, a quota.” — Consultant, US pet insurer, on TrueFoundry

“For human users you typically have a hundred entitlements. For agentic, you want to be very specific: for this table, which column you can modify, and for how long, and after that the permission is revoked. That’s why we’re adopting Okta’s agent identity.” — IT consultant, US fintech bank

“Hackers are using a lot of AI. The number of alerts increased dramatically, so we are also using AI to monitor the alerts.” — Chief Data Officer, Australian healthcare distributor

“Since Mythos came up, the time to patch is being shortened. AI systems can exploit vulnerabilities sooner, and smaller vulnerabilities get connected into a big one.” — Chief Data Officer, Australian healthcare distributor

“We save a lot more details now than humans would in this auditing software. It’s a dramatic usage increase in their API consumption.” — Consultant, US pet insurer, on AuditBoard

“If agents are cars, Collibra is the highway. More cars are getting on the highway.” — Digital lead, global pharma company

“Now they put an AI agent behind it, so it means more business for us. A little less than 3% before; with the new AI agent it’s 3.7%.” — Sales and marketing AI lead, consumer-goods company (Canada unit), on PayPal

“Pinterest converts much better than any other social audience. In the pilot we are at 4%. That’s a good sales opportunity for us.” — Sales and marketing AI lead, consumer-goods company (Canada unit)

1C. Today’s losers: software that agents can rebuild or bypass

The losers share one of two traits. Either they are easy to rebuild, because they do one narrow thing on data the enterprise already holds, so a coding agent reproduces them in hours to weeks; or they are easy to bypass, because their value was a screen that people opened, and people now ask an assistant instead. Most of the lost spend goes to in-house builds, some to general AI assistants, and a little to AI-native point tools. Large suites lose seats and non-core modules in the same way, but net gain today (1A).

What we heard

“Those smaller ones we’ve already replaced with AI, and we’ve given notice on those licenses.” — Technology executive, African investment bank

“I was able to create an app within probably two hours, or one hour even. I don’t need to buy any other software.” — Chief Data Officer, Australian healthcare distributor

“Some new automation that would have become another bot three years ago now goes straight into the agent workflow itself.” — Enterprise analytics director, US healthcare group

“There is no real meaning for you to get licensed for Claude as well as pay a license fee for Tableau, Power BI and ThoughtSpot.” — Enterprise analytics director, US healthcare group

“You can ask a question and get the report without using those tools, even through Snowflake. The BI tools will eventually be killed.” — Technology executive, US regional bank

“With AI coding, doing front end is a lot easier right now. We do not even need a layer of system to manage the content.” — Head of Data and AI, Singapore telco

A framework for Parts 2 and 3: three layers of cost in enterprise software

Enterprise software carries three layers of cost: building it, running it, and the domain knowledge behind it. Part 2 shows that AI has cut the first layer and part of the second, which is why competition is rising. Part 3 shows that AI has not yet affected the third, which is why the incumbents retain an advantage, though it is narrowing.

Part 2 · Longer term: sharper competition from in-house builds and AI-native vendors

In the longer term: the systems of record will face sharper competition from in-house builds and AI-native vendors, driven by two incremental changes: AI has made software cheap to build, and software has gone headless, which erodes the interface moat over time.

2.1 Early signs: replacement has already begun

Early signs: parts of the systems of record are already being replaced by AI-powered in-house builds and AI-native vendors.

In-house builds. Replacement follows a clear order, from the software that is easiest to rebuild and cheapest to operate toward the core: single-purpose tools first, then the interface layer (dashboards, reports and front ends people used to log into), then non-core modules of the big suites, and it stops there. Across the ~30 enterprises, 15 have an AI-powered in-house replacement done or in production, five more have one planned, but none has touched the core modules yet. Interviewees expect core modules to be replaced only gradually, over a three-to-five-year horizon.

AI-native vendors. AI-native vendors have appeared in every enterprise software category. None offers the full suite of a system of record, but each competes directly with one of its modules on price, flexibility and customization.

What we heard

“A lot of companies claim to be AI-native, but what they present to us is a UI made with a web coding tool. There’s no product behind it. I can build that too.” — Head of Data and AI, Singapore telco

“A lot of companies say they are AI, but basically it’s old software.” — CTO, European telco

“Every SaaS category now has an AI-native challenger.” — Consultant, US pet insurer

“In my industry the default is buy, not build.” — Head of Data and AI, Singapore telco

2.2 Incremental change #1: AI has made software cheap to build and run

AI has made software cheap to build and run, and Opus 5.5 now enables AFFORDABLE long-running software development with minimal human intervention.

AI has cut the cost of building software step by step, and each new model or tool release moved the line (table below). The marginal change from Opus 5.5 is that development can now run for hours with minimal human intervention, test itself and stay affordable, so a small team can sustain a project rather than only start one. Two consequences follow. (1) The cost of building software has fallen sharply; capability now exceeds what an in-house rebuild of a CRM or ITSM requires. (2) The cost of running software, from maintenance and debugging to upgrades and support, has fallen to a degree: agents are starting to take on refactoring, upgrades and support, but interviewees see AI’s help on running costs as far less pronounced than on building.

What we heard on building

“The need to write a very good prompt has reduced. I personally think the inflection point was Opus 4.5.” — Head of Data and AI, Singapore telco

“It’s the first time in my career that my team comes and tells me a project finished earlier than we thought.” — CTO, European telco

“Something I couldn’t develop before that I can develop now? I don’t think so. But I can do it a lot easier now.” — Head of Data and AI, Singapore telco

“Tech is no longer the bottleneck.” — Technology executive, African investment bank

On whether Opus 5.5 could build a CRM: “I think it’s possible.” — Head of Data and AI, Singapore telco

“What Atlassian does is not that complicated. We don’t need frontier models to build that stuff.” — Technology executive, African investment bank

“If models stopped evolving for two years, it would still work.” — CTO, European telco

What we heard on running: AI is starting to help, but running stays hard

“Before, refactoring the code was one and a half years. Now a big chunk of it is done by Claude Code; instead of my people writing it, they just validate.” — CTO, European telco

“We’re using agentic AI to analyze all the custom code, produce the use cases, the test scenarios and the success criteria, as the handover to a third party running the upgrade as a service for us.” — Technology executive, African investment bank

“If AI builds it, most of the developers won’t know the code. Who’s going to support it?” — Chief Data Officer, Australian healthcare distributor

“The maintenance of the software is more than they thought: when Java needs to be upgraded, when the database needs to be upgraded, security risks. We built the software with AI, so we don’t know every nut and bolt of it.” — Technology executive, African investment bank

“Because you can do it does not mean you should. Before you realize it, you’re maintaining a piece of software and you have to justify all the engineering spending.” — Consultant, US pet insurer

2.3 Incremental change #2: SaaS has gone headless

SaaS has become increasingly headless and open to third-party agents recently; this shift will erode SaaS’s interface moat and lock-in effect over time.

Vendors opened the door themselves, because being bypassed is worse than being called. In the short run every call is paid for (1.2). In the long run the interface stops being the reason customers stay: once agents do the work, the system behind them can be replaced module by module without end users noticing.

What we heard

“They are afraid that someone else will build a CRM very easily, so they allow AI agents to operate, even if they lose some money.” — Chief Data Officer, Australian healthcare distributor

“Salesforce’s original business model was built on a great interface. Once the interface becomes conversational, the real thing is your data.” — Chief Data Officer, Australian healthcare distributor

“SaaS is not going to be as appealing as before just because it has a UI. Every company can build a UI cheaply.” — Consultant, US pet insurer

“Why do we pay Salesforce $800 a year per user? For what? A database?” — CIO, US diagnostics company

“You can de-aggregate all the services you pay Salesforce for.” — CTO, European telco

“Salesforce chooses the underlying model. I sign a two-year deal at a high rate, while in a year there are open models that cost a tenth of that. Salesforce would make a lot of margin.” — CTO, European telco

“Their growth will not be as high as before. More like a lower-growth, price-sensitive infrastructure provider.” — Consultant, US pet insurer