Newsletter · · Ashutosh Agarwal

Per Seat Software Pricing Breaks as AI Agents Become the Buyer - Weekly SaaS / Software Podcast Recap - Week of August 30, 2026

Weekly SaaS / Software Podcast Recap for the week of August 23 to 30, 2026. Podcast synthesis on operators arguing that per-user-per-month licensing cannot survive the cost of inference, the image of an AI agent cancelling a thirty thousand dollar contract at 2 a.m., an AWS leader insisting the SaaS apocalypse is not real, and DeepSeek v4 forcing US labs to justify their premium.

Weekly SaaS / Software Podcast Recap

Week of August 30, 2026: Per Seat Software Pricing Breaks as AI Agents Become the Buyer

Window: August 23 to August 30, 2026.

Top of mind this week

The software conversation this week did not happen on the big-name investor roundtables. It happened on the operator and AI-strategy podcasts, where the people actually selling and building software are working out what AI does to their business model. And they kept circling the same, genuinely important question: the way software has been priced and sold for 20 years, a fixed fee per user per month, may simply not survive the shift to AI.

Three ideas ran through almost every episode:

  1. Pricing is breaking. The old per-seat model was built for a world where the cost of serving a customer was near zero. AI changes that, because every AI answer ("inference") costs real money in computing power. Multiple guests argued software is moving to consumption pricing (you pay for what you use) or outcome pricing (you pay when the software actually gets a result). The blunt version came from a founder who said the per-user-per-month license "doesn't really work in the age of AI at all."
  2. Agents are becoming the customer. In several discussions the person clicking buy or cancel was no longer a person at all, it was an AI agent acting for the business, and it does not care how pretty your software is. One widely-shared essay imagined an AI sales agent quietly cancelling a company's $30,000-a-year software contract at 2 a.m. because it calculated the tool was only delivering $12,000 of value.
  3. "Is SaaS dead?" got a firm answer: no, but it has to change. An AWS leader flatly said "the SaaS apocalypse is not real," and a well-known AI newsletter writer said his own team has not "ditched software-as-a-service products in favor of vibe-coded apps." (Vibe coding means building software by describing what you want to an AI instead of writing the code yourself.) The nuance: boring, reliable, deeply-embedded software wins; thin AI wrappers on top of old workflows lose.

Everything below expands on these, with the specific companies named and who said what.

1. Dominant themes

Theme 1: the death of per-seat pricing

This was the single most-discussed topic of the week.

The clearest statement came from Husein Sharaf, founder and CEO of Cloudforce (a Microsoft-backed cloud and AI services firm), on The AI Files (Aug 23). He argued the traditional software model is fundamentally mispriced for AI:

The SaaS model, the traditional software model doesn't really work in the age of AI at all. And it's because this per-user-per-month license was never designed to really consider the cost of inference. And the cost of inference is expensive.

His key insight, which is easy to miss: even though AI models keep getting cheaper per answer, total costs keep rising, because people use far more of them. "The models are becoming cheaper and more efficient, but we're using more tokens at a faster rate and we're spending more." (A token is the unit AI is billed in, roughly a chunk of a word.) His conclusion: software needs "a better way of selling and buying AI as an intelligence layer," and "most SaaS companies are not ready for that."

The same shift showed up from the AWS side on Ultimate Partner® (Aug 23), where AWS Marketplace leaders described customers moving "to a world that's going to be more consumption based or outcome based pricing." Their concrete example: Zendesk (customer-service software) pricing on resolved tickets rather than per user, so you pay when a support issue actually gets solved.

And on Revenue Builders (Aug 23), Seong Park, an SVP at Cursor (the fast-growing AI coding tool, made by Anysphere), described how consumption pricing changes even how you sell. He described customers doing "token maxing," meaning "no budget, just go," burning as much AI as they can, until "finance steps in and says, hey, what are you guys doing over there?" His point: when customers pay by usage, the vendor's job shifts from closing a deal to helping the customer actually get value, because "the real sale starts after the signature."

Theme 2: when the buyer is a machine

A striking, recurring image: the customer of the future is an AI agent, not a human, and that breaks a lot of assumptions software companies rely on.

On The AI Daily Brief (Aug 23), host Nathaniel Whittemore read from a set of essays by builders and investors. One, by investor Tina Ha ("Boring infrastructure will win"), captured it best. She describes software being cancelled overnight:

Your customer decides to stop using your customer relationship management software at 2 a.m. Why? The sales rep realized that your $30,000 annual contract for your CRM only gives you $12,000 in value... There was no meeting, no negotiation, because the sales rep was an agent. Agents are not loyal. They are rational actors... They are, in a way, ruthless. Your software can be easy to use and look good, but AI agents neither see nor care.

Her conclusion is important for anyone holding software names: as AI models all become similarly capable, "companies will compete less on having the best model, and more on the systems that connect those decisions to real-world outcomes." The winners, she argues, own the boring plumbing, meaning task routing, data access, workflow rules, and compliance, which act like "toll roads": "If you don't want to pay, you've got to build your own bridge. And that could take years and cost hundreds of millions for compliance." She specifically flags regulatory approval, banking, and compliance systems as durable, and the idea of headless software built only for machines to talk to each other, with no human users at all.

The same theme showed up practically on Ultimate Partner®: AWS said buyers now discover software through AI ("agent mode") using natural language instead of category searches, which means a software company's product listing has to be readable by an AI agent, "almost an API contract," not marketing copy. They called this designing for GEO, essentially search-engine optimization but for AI agents.

Theme 3: AI's real effect on software is to split winners from losers, not to kill software

Nobody serious said software is finished. The nuanced view was that AI widens the gap between good and bad.

On Topline (Aug 23), the hosts and guest Rick Smolen (CRO of ShipHero) debated AI in sales and landed on a barbell view: "the best will become 10x better and the average will become in relation worse." One host compared AI to debt: "It's like debt... It can make performance exceptionally good or it can bankrupt you."

The AI Daily Brief essays echoed this for software companies themselves. Sumit Singh, a former Andreessen Horowitz partner now running a firm called WorldBuild, warned: "Founders who AI-ify existing workflows will lose." His argument is that just bolting AI onto an old process is a trap he calls skeuomorphic, making the new thing look like the old thing (like early phone apps with a trash-can icon that looked like a real bin). The winners, he said, invent workflows that were never possible before, the way Uber and DoorDash did with smartphones: "What becomes possible now? What work can we invent that only AI makes possible?"

Theme 4: the economics of AI itself keep getting cheaper, and someone wants to own the meter

On Keen On America (Aug 23), tech investor Keith Teare argued the price of AI "over time will tend towards zero" for light, individual uses (a student, a school, a hospital patient) while heavy business users "running 3,000 agents" will always pay. The discussion turned to OpenRouter, a service that automatically sends each request to the cheapest model that can answer it well, squeezing token costs down. And it highlighted Stripe (the payments company) pushing to meter intelligence, becoming the billing layer that charges for AI usage the way a phone company bills for data. As Teare put it, metering "just means they need to get paid for the investment." He also gave a rare hard number on how unprofitable consumer AI still is: he pays about $100 a month for Anthropic's Claude, but reckons "if they charged what they would have to charge to be profitable on you individually, you'd probably have to pay $200 to $400 a month, maybe even more."

2. Key debates

Debate A: does AI make salespeople far more productive, or just flood the world with AI slop? The wedge is whether AI is a genuine productivity multiplier in go-to-market, or a shortcut that produces worthless work.

  • Side A (huge productivity gains): One Topline host argued AI is a real multiplier and that some AI-native companies already have reps doing "net two, three, four times what quotas used to be." The bull case: reps spend only "20 to 30 percent of their time doing revenue-generating activities," and AI hands the rest of that time back.
  • Side B (it's mostly slop, and it can't do the human part): Rick Smolen (ShipHero) pushed back hard. AI note-takers and AI-written emails, he said, cannot capture "the tone... the nuance... that little look that one person gave another." He warned reps who outsource their thinking to AI will see "their performance go backwards." Host Sam Jacobs described getting an AI-written cold email so generic he replied asking, literally, "are you a real person?" The counter-counter: enterprise selling is physically capped, "there's only so many trips you can take in a month," so AI cannot 5x a rep who has to get on planes. Tellingly, Smolen said ShipHero has "not raised quotas on anybody on the team."

Debate B: is the per-seat software model dead, or just evolving? The wedge is whether AI breaks the economics of subscription software, or just adds a new pricing option.

  • Side A (the model is broken): Cloudforce's Husein Sharaf, for whom per-user-per-month "doesn't really work in the age of AI at all" because it ignores the cost of inference.
  • Side B (SaaS is fine, it's just adding usage-based options): An AWS Marketplace leader on Ultimate Partner® said plainly, "the SaaS apocalypse is not real," framing consumption and outcome pricing as an addition to how software is sold, not a replacement. On The AI Daily Brief, Every CEO Dan Shipper said his own company "hasn't ditched software-as-a-service products in favor of vibe-coded apps," and argued AI actually increases demand for expert human work: "The more we automate, the more expert human work there is to do."

Debate C: should a US hospital or university ever run a cheap Chinese AI model? The wedge is whether frontier AI is now good enough and cheap enough from China that paying US labs' premium prices is hard to justify.

  • The setup (Sharaf, The AI Files): In one week, OpenAI launched GPT-5.5 (strong, but "double as expensive as GPT-5.4"), while China's DeepSeek launched v4, and "not just that China has caught up, it is that they now probably have better frontier intelligence and too cheap for US labs to defend."
  • Sharaf's answer: He would not recommend regulated customers use DeepSeek's public service, over data-security concerns, but because it is open-weight (you can download and run it privately), a few of his customers run it in a locked-down environment, mostly for research. His real point is economic: DeepSeek "forces the Western counterparts here in the US to really justify their premium." So the debate is not switch to China, it is that US labs now have to prove they are worth the price.

Debate D: which software companies win the AI era, the flashy model builders or the boring infrastructure? The wedge is whether value accrues to whoever has the best AI, or to whoever owns the unglamorous connective plumbing.

  • Side A (boring wins): Tina Ha's "boring infrastructure will win," where as models converge the edge shifts to workflow orchestration, compliance, and regulated systems (banking, wire transfers), which behave like toll roads.
  • Side B (vision and craft win): Another essayist, Noah Breyer, argued "software companies will outperform software factories," that building great software is less like stamping out identical car parts and more like Andy Warhol's studio, where everything serves one creative vision. The hardest problem "is still creating a vision and keeping an entire team... building toward it." The synthesis both sides shared: raw AI capability is becoming a commodity, and the durable value is in the human and institutional systems built around it.

3. Specific names: who was discussed, and the angle

Public companies

  • Microsoft (MSFT). Mixed to constructive. On The AI Files, Sharaf noted Microsoft's exclusive grip on OpenAI has ended (OpenAI models are coming to AWS too). He does not see this as Azure losing its edge, arguing Microsoft's strength is giving developers every tool and model choice: its "ethos in this moment is to embrace all the models, bring everything into Azure... and let people choose." He also noted Microsoft-partner ecosystems benefit as Anthropic's Claude becomes available on Azure.
  • Amazon / AWS (AMZN). Constructive as the AI distribution layer. On the same day, Amazon announced OpenAI models coming to AWS in limited preview. On Ultimate Partner®, AWS leaders positioned the Marketplace as the emerging default way enterprises buy software, with AI-driven (agentic) discovery giving smaller vendors "an equal opportunity" to be found. Bull angle: AWS is wiring itself into how software gets discovered, bought, and billed.
  • Salesforce (CRM). Cited as a cautionary tale on messaging. Sharaf recalled Salesforce a year ago talking up cutting support headcount and quotas because of AI agents, "selling to the boardrooms rather than the people who actually will have to adopt it," then reversing to announce it is "creating new jobs, a thousand new internships." The lesson he drew: how you talk about AI to actual users makes or breaks adoption. Salesforce's core product is also the archetype in Tina Ha's agent-cancels-the-contract-at-2-a.m. warning.
  • Chegg (CHGG). Bearish, already broken. Held up as the clearest victim of generative AI so far: the homework-help company is, in Sharaf's words, "practically dead," with the stock going from "$200 per stock and now it's 99 cents" as students use ChatGPT directly. (Those figures are the speaker's characterization, not verified market data.)
  • Zscaler (ZS), and by name AppDynamics and Wiz. Referenced as elite enterprise sales organizations. On Topline, the pedigree of OpenAI's new revenue chief ("he was AppDynamics, Zscaler and Wiz") was cited to argue today's best enterprise sellers are better than any past generation. No investment stance was expressed on the stocks themselves; they were name-checked as talent benchmarks.
  • Google / Alphabet (GOOGL). Light mention. Referenced as a hyperscaler whose Gemini models can be pulled into other platforms, part of the "more competition than people expected" landscape, framed as good for buyers.
  • Oracle (ORCL). Passing mention as one of the big cloud and data-center operators.
  • MongoDB (MDB). Passing mention only (a Cursor sales leader's former employer). No investment view.

Private companies and AI labs

  • OpenAI. Still frontier, but no longer alone and getting pricier. Launched GPT-5.5 (strong at coding and agents, but roughly double the price of the prior version). Lost its Microsoft exclusivity, with models now heading to AWS. Hired a heavyweight revenue chief. Broadly framed as still leading but facing real competition on both quality and price.
  • Anthropic. The surprise winner narrative. Sharaf said few expected Anthropic "could not only catch up to OpenAI, but actually surpass OpenAI in a lot of ways," and its Claude models are seeing heavy enterprise adoption, now available on Azure. Counterpoint from the operators: Claude's writing still produces generic output, and one Topline host said AI-drafted follow-ups "all comes out the same" and that he never clicks send. And economically, consumer Claude is still sold below cost (see the $100 versus $200 to $400 point above).
  • DeepSeek (China). The price disruptor. v4 launched the same week as GPT-5.5, described as possibly better frontier intelligence at a price US labs struggle to match. Controversial for regulated buyers on data security, but its real impact is forcing US labs to justify premium pricing.
  • Stripe. Positioning to own AI billing. Highlighted for its push to meter intelligence, becoming the payment and metering layer for AI usage, sitting "in the middle of the traffic stream for tokens."
  • Cursor (Anysphere). Case study in the new selling model. Its consumption pricing drives token-maxing behavior and a go-to-market approach where value delivery after signing matters more than the initial sale.
  • Blitzy. AI coding tool worth knowing. Pitched in an ad read, so treat as vendor claims, as different from other AI coders: it spends days first mapping your entire existing codebase into a knowledge graph before writing anything, and claims to deliver "over 80% of entire software epics autonomously" for Fortune 500 engineering teams.
  • Cloudforce. The source company. A Microsoft solutions partner backed by Microsoft's venture fund (M12); recently took outside investment and built out a sales and marketing engine. Builds AI for regulated industries (government, healthcare, education) on Azure; its platform (Nebula One) lets non-technical staff build their own agents.
  • Every. A "SaaS isn't dead" data point. The roughly 30-person media and software company (CEO Dan Shipper) says it still runs on SaaS products and still hires human writers, editors and engineers rather than replacing them with agents.