Asia/Karachi
BlogOctober 30, 2026

Is SaaS Dying? What AI Agents Mean for Software

The evidence for and against, and what to do about it
Fawaz Israr
Is SaaS Dying? What AI Agents Mean for Software
Is SaaS dying? On the evidence available in late 2026, no, but the way SaaS companies charge for their products is under real pressure. AI agents can now do work that people used to do by clicking through software, and that weakens the per-seat subscription most vendors depend on. Vendors are answering with agents of their own and with usage or outcome pricing, and public software stocks, after a steep fall early in the year, had climbed back to a gain for 2026 by October. I work at a studio that builds this kind of software, so I have a stake in the answer. I say more about that below. First, the facts, with dates, because this topic attracts a lot of confident predictions.

What is SaaS, and why did seat-based pricing work?

SaaS means software as a service. You rent software that runs on the vendor's servers and use it in a browser or an app, while the vendor handles hosting, updates and security. It replaced boxed licences and servers kept in a back office. Most SaaS products charge per seat: a fixed monthly fee for each person who has a login. That worked well for both sides for a long time. Buyers paid little upfront and could add or remove people as they hired. Vendors got predictable recurring revenue, and their income grew whenever a customer's headcount grew. The price also tracked the value, because the more people used the tool, the more work it supported. Investors liked the model for another reason: margins. Benchmarks quoted by CloudZero put traditional SaaS gross margins at roughly 70 to 90 percent, depending on the source, since serving one more user costs the vendor very little. One assumption is buried in all of this. The customer is a human sitting at a screen. Interface design, onboarding, seat counts and growth by headcount all depend on that. The case against SaaS is a case against that assumption.

What is the case that SaaS is in trouble?

There are five arguments, and each deserves a fair hearing. The first is that agents do the work instead of people. If an AI agent reads a support ticket, checks the order system, issues a refund and writes the reply, nobody needs to open the helpdesk interface for that ticket. George Brocklehurst, a managing vice president at Gartner, said agentic systems deliver outcomes without the usual application interface, which in his words "breaks the link between user growth and revenue growth" for many vendors (as reported by CIO Dive on 6 July 2026, and in similar terms by CIO on 2 July 2026). The second follows directly: per-seat pricing weakens when fewer humans use the tool. As a made-up illustration, a support team of 50 that becomes a team of 20 plus an agent cuts the vendor's seat revenue by more than half, even if the customer is happier. The vendor pays for the improvement. The third is build versus buy. AI-assisted coding makes small custom tools cheaper to produce than they were. Retool, which sells internal-tool software, surveyed 817 builders in late 2025 and reported in February 2026 that 35 percent of teams had replaced at least one SaaS tool with a custom build, and 78 percent expected to build more in 2026. The most replaced categories were workflow automations, internal admin tools and BI tools, with CRMs and form builders at 25 percent. Treat this with care: the sample is self-selected, Retool profits from the answer, and the survey does not say whether the replacements worked out or saved money. The fourth is investor pressure. On 3 February 2026, Reuters reported that a selloff in software and data stocks deepened after Anthropic released plug-ins for its Claude Cowork agent covering legal, sales, marketing and data analysis. Thomson Reuters fell nearly 18 percent in a day, its biggest daily drop on record. A Schroders analyst said AI lets businesses do more with fewer staff, which threatens the traditional model of charging per user (Reuters report published by The Star on 4 February 2026). That reaction was about expectations, and no customer had stopped paying yet. The fifth is margins. Adding AI features adds a real cost per use, because every model call costs the vendor money. ICONIQ's July 2026 report, as summarised by CloudZero, put the average gross margin of AI products at 45 percent in 2025, with a projection of 53 percent for 2026 and 59 percent for 2027. Bessemer's figures, also quoted there, are 50 to 60 percent for AI companies against 80 to 90 percent for SaaS. A different benchmark, Benchmarkit's, found that median software gross margin was still above 80 percent, so the squeeze has not yet reached the typical company. Margin pressure is clearest at AI-native startups and at vendors who bolt on heavy AI features.

What is the case that SaaS survives?

The strongest version of this argument starts with what an agent needs in order to act. It needs somewhere to read and write: a customer record, an invoice, a ticket, an employee file. Someone has to keep that data accurate, secure, permissioned and auditable. That is what systems of record do, and agents make them more useful, since agents hit them thousands of times a day. Bain made this point in a May 2026 press release. It said SaaS companies spent two decades building moats around systems of record, and that agentic AI is unlikely to end SaaS. It expects agents to automate the human coordination work that sits between those systems. Bain estimated a US market of about $100 billion for that coordination work, of which vendors were then capturing roughly $4 billion to $6 billion (Bain, 7 May 2026). That is a consultancy's estimate of a market it hopes to advise on, so take it as a direction and not a forecast. Several other forces favour incumbents:
  • Data gravity. Years of customer history live in the existing system, and moving it is slow and risky.
  • Integrations. A mature product connects to payment, email, accounting and identity tools. A custom build has to recreate each connection.
  • Compliance and security. Audit trails, access controls and certifications take sustained work to keep up, and a homegrown tool inherits all of it.
  • Distribution and trust. A company's finance team will sign a contract with a known vendor more readily than it will rely on a tool a single employee built over a weekend.
  • Switching costs. Staff are trained, processes are built around the product, and reports depend on its fields.
Building a tool is one cost. Owning it for years, with its bugs and security patches, is a larger one. That is inference on my part and not a measured result, but it matches why the Retool numbers concentrate in small internal tools and not in core platforms. Even Gartner, the source of the scariest 2026 number, was not predicting extinction. According to CIO, Brocklehurst called the change a metamorphosis and said SaaS will persist in a different form.

What is actually changing in 2026?

Here is a dated sequence of verified events, taken from the sources linked in this post.
  • 12 December 2025: The Register reported that Salesforce CEO Marc Benioff said seat-based pricing, with some caveats, was becoming the norm for its AI agents, after customers asked for predictable costs.
  • 3 February 2026: The selloff described above, driven partly by the Anthropic plug-ins.
  • 17 February 2026: Retool published its build-versus-buy report.
  • Late January to April 2026: The S&P 500 software and services index lost more than 26 percent from late January to its April low, according to Reuters on 6 October 2026.
  • 27 April 2026: GitHub announced that from 1 June 2026 all Copilot plans would bill by token usage through "AI Credits". GitHub said the old premium request model was no longer sustainable because Copilot now runs long, multi-step agentic sessions that need much more compute (GitHub changelog).
  • 7 May 2026: Bain's report on agentic AI as an opportunity for SaaS.
  • Late June 2026: The software recovery began, per Reuters.
  • 1 July 2026: Gartner said up to $234 billion of enterprise application software spend was at risk from agentic AI through 2030, about 20 percent of enterprise SaaS spending (as reported by CIO and CIO Dive; I could not open Gartner's own page, so I rely on their coverage).
  • 26 August 2026: Salesforce reported second-quarter results for fiscal 2027.
  • 6 October 2026: Reuters reported that the software index had reached its highest level since November 2025.
That last item is the one that surprised people. The index rose 1.3 percent that day, posted its biggest July to September quarterly jump since the second quarter of 2020, and was up about 5 percent for the year. Reuters attributed this to strong earnings from Salesforce, ServiceNow and Accenture, partnerships with AI labs, and a view among analysts that disruption fears had been overstated. A strategist from Zacks said the real test could come in the second half of 2027, when more data centre capacity might make AI coding a bigger threat to traditional software firms (Reuters via Kitco).

Are software vendors really leaving per-seat pricing?

Partly, and the picture is mixed. Some moved early. Zendesk announced in August 2024 that customers would pay only for issues its AI agents resolve on their own, with a starter amount included. GitHub moved Copilot to usage-based billing, as above. Brocklehurst told CIO that incumbents should shift from interface-based value to outcome-based value. Others went the other way. Benioff's December 2025 comments, reported by The Register, described a seat-based licence for Agentforce called the Agentic Enterprise License Agreement, which bundles credits into one package priced by number of seats. Gartner analyst Jan Cook said buyers still fear unpredictable AI bills, and warned that seat licences can hide fair-use limits and credit top-ups. A Forrester analyst warned that buyers could end up paying for users who barely use the product. Salesforce's own numbers show why this is hard to read. In its second-quarter fiscal 2027 release of 26 August 2026, the company reported revenue of $11.3 billion, up 11 percent, and Agentforce annual recurring revenue above $1.5 billion, up over 240 percent. Two caveats apply. The company widened the Agentforce ARR definition that quarter to include other AI offerings, so the growth rate is not directly comparable with earlier quarters. And revenue included $456 million from the Informatica acquisition, so organic growth was lower than 11 percent. Next to full-year revenue guidance of $46.1 billion to $46.4 billion, $1.5 billion is a small but fast-growing slice. My reading is that pricing is moving toward hybrids: a platform fee or seat fee for access, plus usage or outcome charges for what agents do. Buyers want a cap on the bill, and vendors want the revenue to follow the work. That is an inference from the examples above, and a single clean answer may not appear for years.

What could happen next? Four scenarios

None of these is a forecast. They are ways the next few years could plausibly go, with what would have to be true and what to watch.
Text
Scenario 1: Repricing, not replacement
  Most vendors add agents and move to hybrid pricing.
  Seat counts flatten, usage revenue grows, margins dip a little.
  Needs: agents that help staff but do not replace whole teams.
  Watch: net revenue retention and gross margin in vendor earnings.

Scenario 2: The agent layer wins
  Horizontal agent platforms sit on top of many apps.
  Apps become databases with APIs, and the interface matters less.
  Needs: agents that work reliably across systems, and vendors
  that allow them full API access on fair terms.
  Watch: vendor API limits, and contract clauses about agent access.

Scenario 3: Custom builds eat the long tail
  Small tools (workflows, admin panels, simple reports) get
  rebuilt in-house. Systems of record stay put.
  Needs: cheap AI-assisted builds that stay maintainable.
  Watch: churn at low-priced SaaS tools, not at the big platforms.

Scenario 4: Fast disruption
  Agents become reliable enough that seat counts fall quickly
  across support, sales and back office work.
  Needs: a big jump in agent reliability and enterprise trust.
  Watch: seat reductions at renewal. Some analysts point to the
  second half of 2027 as the test.
The first scenario is the most consistent with what has been verified so far. The fourth cannot be ruled out, and I would not bet my plans on it either way.
A laptop, a white coffee mug, a handwritten notepad with a pen and a smartphone on a wooden desk

What does this mean for businesses that buy software?

If you run a business or a team that pays for software, the practical points are few. Ask whether an agent could do through an API everything a person does through the screen. Brocklehurst advised CIOs to check this, and to read contracts as carefully as the technology, because licensing terms can block an AI plan and existing contracts stay in force. Also ask who owns what agents learn from corrections and exceptions in your data. Watch usage-based bills. They can rise without anyone deciding to spend more, and GitHub's own reason for changing its pricing shows how quickly agentic use can raise costs. Ask for spending caps and alerts before you sign. Pilot custom builds on small, low-risk tools first, such as an internal admin screen or a simple report. Keep systems of record, payments and anything with legal exposure on products that have a support team and a security record. If you do decide to build, the question of who builds it matters as much as what; I compare the options in freelance developer vs agency vs in-house.

What does it mean for SaaS founders and developers?

For founders, the uncomfortable point is that charging per seat for a product whose job is to save people time contains a contradiction. Test your pricing against a world where your best customer needs fewer seats next year because your product works. Options include pricing by the outcome you deliver, by usage with a cap, or by a platform fee with a seat component. Bain advises making products and data "agent ready", which in plain terms means a complete, documented API, clear permissions and an audit trail. If an agent cannot do what your interface does, a competitor's product that allows it will be easier to automate. And the data and context you hold are the hardest part to copy. If you are choosing a stack for a first product, the MVP tech stack post covers the basics, and none of it has changed because of agents. For developers, some of the work moves to new places. Someone has to build the integrations, the agent workflows, the permission models and the custom systems for companies that decide to build. I looked at the wider question of AI and programming jobs in will AI replace programmers, and the broader benefits and risks of AI in a separate post. Here is my stake, stated plainly. I am a final-year Computer Science student, and I work in web development at Hadi Dev Lab, a studio that builds agentic AI systems, SaaS platforms, custom CRMs and mobile apps. If SaaS shrinks, custom software work may grow, and if agents spread, agentic systems work grows. Take my reading of the evidence with that in mind. The AI features of my own final year project, GoCart, are still in development.

What should a founder decide to build now?

Use this as a set of questions and answer each one in writing before you commit months of work.
Text
Founder checklist: what to build now

1. Who is the user in two years: a person, an agent, or both?
2. Can an agent do everything through your API that a person
   can do through your screens?
3. What do you own that an agent or a custom build cannot copy
   cheaply: data, integrations, compliance, a trusted brand?
4. Does your price still work if the customer needs half the
   seats next year because your product works?
5. What does one agent action cost you in model fees, and is
   your gross margin still healthy after it?
6. Can you cap a customer's bill so usage pricing does not
   frighten buyers?
7. Which small tools could a customer rebuild in a weekend,
   and what do you offer beyond them?
8. What do you want to be the system of record for?
9. What would make you change course, and which number
   will tell you?

What do I make of all this? My take

I am a student developer. I am not an analyst or an investor, and the first thing the evidence has taught me is how quickly confident takes go out of date. In February, the story was a collapse. By October, the main index was up for the year. Neither the panic nor the relief tells you much about what customers will do in 2028. What the evidence supports, in my view, is a pricing and margin problem. I do not see an extinction event in it. The link between seats and revenue is weakening, and vendors who sell a screen and nothing else are most exposed. Vendors who hold the data, run the integrations and carry the compliance burden look much safer, and Bain, Gartner and the Reuters reporting all point that way. Salesforce choosing seat-based licensing for agents suggests buyers still want predictable bills, which tempers the idea that seats are finished. What remains uncertain is agent reliability and speed. If agents get good enough that a company can cut support or back-office seats by half without trouble, scenario 4 arrives faster than most plans assume. I would change my view if seat reductions started showing up in vendors' renewal numbers, or if enterprise buyers began routinely writing agent access into contracts. I would lean further towards the survival case if the cheap custom builds in Retool's survey turned out to be hard to maintain after a year. This is not financial advice.

Want to talk about this?

I am a full stack developer in Rahim Yar Khan, Pakistan, a final-year Computer Science student, and I work in web development at Hadi Dev Lab. If you are weighing a SaaS product, a custom build or an AI feature and want a second opinion, you can message me on WhatsApp or email. You can see my work and read more about me first. Cover photo from Unsplash.
Share this post:

Recent posts

Thumbnail of Developer Skills to Learn in the AI Era: A Student Guide
Fawaz Israr
November 2, 2026
Developer Skills to Learn in the AI Era: A Student GuideCareer
Thumbnail of Will AI Replace Programmers? What Coding Tools Change
Fawaz Israr
November 2, 2026
Will AI Replace Programmers? What Coding Tools ChangeAI