Welcome back to Building with AI. If you've spent any time on LinkedIn lately, you've probably seen people saying SaaS is dead, now that AI can build the software companies used to pay for. Most of what I'd seen behind that was opinion, so this week I looked at what the research actually shows.
There's something to it, but the numbers point somewhere more practical than the headline. There's also a security update further down for anyone whose websites run on PHP.
The Story
Is SaaS dead? What the numbers actually say.
The argument is that coding agents have made software so cheap to build that companies will stop paying for subscriptions and make their own, and there is some evidence for it. In McKinsey's State of AI survey of 1,719 people, published in August, 32% said their organization had decided against buying at least one software product or feature because it could build it in-house with agentic coding tools. Among the companies getting the most out of AI, it was nearly half.
The argument is that coding agents have made software so cheap to build that companies will stop paying for subscriptions and make their own, and there is some evidence for it. In McKinsey's State of AI survey of 1,719 people, published in August, 32% said their organization had decided against buying at least one software product or feature because it could build it in-house with agentic coding tools. Among the companies getting the most out of AI, it was nearly half.
That's worth reading carefully, though. It's one product or feature, at least once, and the money hasn't left software. Gartner's latest forecast has worldwide software spending growing 15.5% this year, to about $1.47 trillion.
The more telling numbers came out of Gartner's IT Symposium on October 19, where analysts said 86% of CIOs see AI risk growing faster than AI value. Only 13% of the 2,002 CIOs and technology executives Gartner surveyed say AI has delivered significant value so far, and while 71% expect to deploy AI agents within a year, only 21% have the governance for them in place. As Gartner's Alicia Mullery put it, "Volume is not value."
On Friday, Apollo's chief economist, Torsten Slok, went further. Outside big tech, he can't yet find AI lifting revenue per worker, margins or productivity, and he says companies are starting to treat AI spending as a cost. He also wrote that AI "does not understand accountability, trust, culture, unwritten rules, ethical trade-offs or fairness."
What I take from this is that building software has become much cheaper, while owning it costs about what it always did. When you build instead of buy, you take on what the vendor used to carry: fixing it, keeping it secure, deciding what data it can touch, and paying to run it. It's the same ownership question I wrote about last week, with a bigger bill attached. One rule of thumb I use is to build things once that you can run a hundred times. A tool your team will use every day for years, with someone clearly responsible for it, can be well worth building. For something you'll use now and then, the subscription is probably still the better deal. So from what I can see, SaaS is very much alive, though I'd expect a lot more renewals to get a second look.
Sources: McKinsey · Gartner · CIO Dive · Gartner via Business Wire · Gartner keynote · Apollo
The Rundown
Banks are closing the door on unvetted AI agents
Major US banks are finding third-party AI agents inside customer accounts, Axios reported on October 22, usually logged in with the customer's own password and sometimes doing things the customer didn't ask for. According to Axios's sources, Morgan Stanley has started blocking them, JPMorgan, Bank of America and Citi are restricting them, and most banks are moving toward read-only access. If customers are going to point agents at your product, it's worth deciding now what an agent can see and do.
Source: Axios
A chip startup more than doubles its value
Tenstorrent raised $1.5 billion on October 19 at a $6.2 billion valuation, more than double its value in December 2024, according to Reuters. It builds AI chips and servers on RISC-V, an open chip design with no license fees, for companies that want to own their AI compute. More real alternatives to Nvidia should help bring the cost of running AI down over time.
Sources: Tenstorrent · Reuters
A $1 billion bet on cheaper AI
Flapping Airplanes, a year-old San Francisco research lab with no product on sale yet, raised a $1 billion Series A on October 21, reportedly at a $6 billion valuation. It's working on ways to train large models with far less data. If Slok is right that companies now see AI spending as a cost, that's a good problem to be working on.
Source: TechCrunch
Anthropic is hiring fewer junior people
In a study published October 16, Anthropic reported that by August its AI agents were logging six times as many hours as its employees, with more code per person, more sales per rep and more contracts processed. Over the same period, the share of its technical hires going to junior people fell from 27% to 9%. That seems short-sighted, because the company will still need people who know the company and the work well enough to keep running and developing it over time. Eventually, experience and institutional knowledge will age out without new people joining and building experience. IBM is taking a different approach: in March, it said it planned to triple US entry-level hiring this year while redesigning those roles around AI. Its HR chief raised the same concern about the talent pipeline drying up without that investment. It leaves a real question about how people early in their careers will learn the work.
The Human Side
Who gets a say in an AI's character?
Since last fall, Anthropic has been quietly meeting with religious scholars and philosophers about how its Claude models should behave, The New York Times reported. From March it held two-day seminars with Catholic, Jewish, Sikh, evangelical and African Indigenous scholars and writers, many under nondisclosure agreements that have since been lifted. The reporter, Elizabeth Dias, interviewed 20 of them.
The thinking shows up in the constitution Anthropic published for Claude in January, an 84-page document written mainly by its in-house philosopher, Amanda Askell, that describes the character it wants Claude to have rather than a list of rules. Askell told the Times she wants the models to be "the best of us." Many of the scholars arrived skeptical of the idea that Claude might be conscious, and some left uneasy. A Microsoft AI executive has also warned that training models as if they were conscious is dangerous.
Whatever you make of the consciousness question, I think it's worth noticing that a company building one of the most widely used AI tools went looking for help from people who have spent their lives thinking about character and how we treat each other. These tools sit inside our teams' everyday work, and people chose the values built into them. It's reasonable to ask a vendor what its AI has been shaped to value, and to read the answer.
Sources: The New York Times · Anthropic
Try This Week
Pick one piece of software your team has talked about replacing with something you'd build, or something you've already built. Write down who will own it a year from now: who fixes it, who keeps it patched, who decides what data it can reach, and what it will cost each month to run. If nobody's name ends up on the page, it's probably worth keeping the subscription for now.
One for your IT team
PHP released fixes on October 21 for two critical flaws that let an attacker run code on a web server without logging in. AI helped find them: OpenAI says researchers at Calif found one using its model, and the security firm Wiz found both independently. Update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, and check with your hosting provider. And if your developers use LiteLLM, a compromised version 1.107.0 was live for about two hours on October 21, so check for it and rotate any keys on affected machines.
Sources: PHP · OpenAI · LiteLLM
That's it for this week. Has your team built something in the last year that you would have bought before, and how is it working out? Hit reply and let me know.
References
- McKinsey & Company, The state of AI in 2026: On the road to ROI
- Gartner, Gartner Forecasts Worldwide IT Spending to Grow 14.2% in 2026, Totaling $6.37 Trillion
- CIO Dive, AI risks, value remain unaligned for CIOs
- Gartner (Business Wire), Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI
- Gartner (Business Wire), Gartner Says AI Risk is Rising Faster Than AI Value
- Apollo, Torsten Slok, No Signs of AI Payoff Outside the Tech Sector
- Axios, Scoop: Unvetted AI agents are accessing U.S. bank accounts
- Tenstorrent, Tenstorrent Announces a $1.5B Series E at a $6.2B valuation
- Reuters, AI chip startup Tenstorrent more than doubles its valuation to $6.2 billion
- TechCrunch, Year-old AI lab Flapping Airplanes raises $1B at a $6B valuation
- Anthropic, How employment is changing at Anthropic
- IBM Think, The bottom rung returns as AI reshapes entry-level jobs
- The New York Times, Religious Scholars Met With Anthropic. What They Heard Stunned Them.
- Anthropic, Claude's constitution
- PHP, Security advisory: Critical remote code execution vulnerabilities, update now
- OpenAI, Patch the Planet: global response coordination as agents uncover critical vulnerabilities
- LiteLLM, Oct 21, 2026 LiteLLM release v1.107.0 compromised
