Executive summary
- One mechanism. Producing artifacts — specs, designs, prototypes, code, synthesis — got cheap. Everything that was valued or priced on artifacts is finding a new price.
- Product management: postings up 14%, PM headcount at tech companies still 28% below its 2022 peak, and a separate “AI product manager” title listed at roughly twice the pay. That’s a different job, not a premium.
- Design: the same curve in another vocabulary. Output is commoditized, judgment is scarce, and designers with AI skills earn a 10–15% premium — well below the 56% that AI skills fetch across jobs in general.
- Software pricing: 97% of SaaS CEOs plan to retire seat pricing while 94% say it still fits their value. The practice is hybrid pricing, up from 25% to 37% of companies in a year, not the outcome pricing everyone talks about.
- Valuations: about $285 billion came off software, data and financial stocks in a single February day on an argument, and by June most of the fall had come back — a repricing of a belief, not of a business.
Judgment is what’s left in every one of these, and you have to show it, not claim it: write down what you expect, then go back and check.
In February 2026, roughly $285 billion vanished from software, data and financial stocks in a single day.
The trigger was a product launch. The reasoning, as far as anyone could reconstruct it, was a single syllogism: if one agent does the work of five people, and your software charges per person, your revenue depends on a number that is about to go down.
That’s a strange thing for a market to notice all at once. The logic had been there for two years, and nothing about the arithmetic changed that week. What changed was that enough people applied it at once, and a price held up by an assumption fell to where the assumption no longer held it.
I’ve been thinking about that event since, because it’s the loudest case of something quieter happening everywhere else. Picture the same week from three other desks. A designer with twelve years of experience is told her portfolio doesn’t show judgment. A product manager with my background finds that his title has split into two jobs with a hundred-thousand-dollar gap between them. A CRO somewhere is explaining to a board why net revenue retention is flat despite record usage.
Four conversations. One mechanism.
It existed to reduce the risk of spending expensive engineering time on the wrong thing. A working prototype now does that faster and more accurately.
A record of finished things. It shows what you chose and hides what you rejected — and rejecting is the half that stayed expensive.
Headcount stood in for value received, reliably enough to build an industry’s financial model on. Agents do the work with nobody logged in.
A price held up by an assumption, which falls the moment enough people apply the arithmetic at the same time.
Output got cheap. Everything priced on output is repricing. And almost nothing in how we work, hire, or charge was designed for that.
This piece is the long version of that claim: the mechanism, then the four places it’s showing up, then what I think it means for anyone whose living depends on one of them. I’ll give numbers where they exist and say so where I’m speculating.
Part one: the mechanism
What “output” means here
I want to be precise, because “AI makes things cheaper” is true enough to be useless.
The thing that got cheap is not intelligence, insight, or good decisions. It’s the production of artifacts. Specifically:
Got cheap
- A written document that describes something
- A visual representation of an interface
- A functioning prototype of an idea
- A synthesis of a pile of unstructured text
- A first draft of nearly any kind of code
- A summary of what happened
Did not
- Deciding which of the six is worth producing
- Knowing what good looks like before you see it
- Rejecting the option that reads well and is wrong
- Owning the call when the evidence is incomplete
- Noticing the thing nobody in the room asked about
- Being right, and having written it down beforehand
Every one of these used to take a trained human hours to weeks. Every one now takes minutes, at a quality between “acceptable” and “better than the median professional,” depending on the artifact and the professional.
That’s the whole input to everything below.
Why cheap artifacts reprice labor
Most professional value in software was, in practice, tied to artifacts. You were valuable because of what you could produce, and your seniority roughly tracked how good your artifacts were and how little supervision they needed.
This was never the stated value. Every design leader and every product leader would have told you, correctly, that the real work was judgment. But the part of the work you could see, judge and pay for was the artifact. That’s what you put in a portfolio. That’s what you showed in an interview. That’s what a manager reviewed on a Friday.
When artifacts become cheap, that whole way of judging people stops carrying information. The portfolio still looks good; it just no longer tells anyone apart. The case study is still polished; anyone can polish one now.
Here’s the part that matters: the market doesn’t stop evaluating people. It starts evaluating something else. The something else is judgment, which was always the real product but was never measured directly, because the artifact was a good enough proxy.
So a generation of professionals who rightly optimized for the old proxy are finding that it no longer tracks the thing it stood for. That’s not a skills gap. It’s a measurement problem, and it feels like betrayal.
Why cheap artifacts reprice software
The same mechanism, one level up.
Software was priced on seats because headcount was a reliable proxy for value received. More people using the tool meant more work done with it, which meant more value, which justified more money. The proxy was so reliable that an entire industry’s financial model — net revenue retention, land-and-expand, all of it — assumed that customers grow their headcount.
Agents break the proxy in the most direct way: work gets done without a human logged in. Billing per human operator punishes the vendor for improving the product and the buyer for adopting it.
Read that twice. Seat pricing didn’t become slightly less accurate. It became anti-correlated with value delivery in exactly the products where the value is growing fastest.
So labor and software are the same story. A proxy that was reliable for twenty years stopped being reliable in about two. Everything priced on it is now finding a new price, and the process is messy, fast, and mostly unadmitted.
Part two: product management
The numbers
Start with the confusing part, because the confusion is the finding.
Product management job postings were up 14% year over year as of May 2026, and early in the year open PM roles hit their highest level in over three years. Meanwhile, PM headcount at US tech companies sits about 28% below its 2022 peak, as of June.
+14%
product management postings, year over year
May 2026
3+ yrs
open PM roles at their highest level in over three years
May 2026
−28%
PM headcount at US tech companies, against its 2022 peak
vs 2022 peak
Sources: Lenny’s Newsletter; Mind the Product.
The numbers
| Measure | Value |
|---|---|
| PM postings, year over year (May 2026) | +14% |
| Open PM roles | Highest in over three years |
| PM headcount at US tech companies vs 2022 peak | −28% |
Both are true. The job market crashed in 2022–23. That was mostly a post-COVID bubble correction rather than automation, though the timing let AI take the blame. It has been recovering into a smaller market ever since.
But volume isn’t where the story is. Composition is.
The Atlassian number
The cleanest data point I’ve found: Atlassian cut roughly 1,600 jobs in March, about a tenth of its staff, more than 900 of them in R&D, and said the savings would pay for investment in AI and enterprise sales.
Headcount down. The skills the money buys, reset. If you were one of the 1,600, you lived through a layoff; if you read the press release, you’d conclude the company was investing. Both accurate.
This is the shape of the whole transition, and it explains why the public conversation is so incoherent. People inside the shift experience it as loss. People reading aggregate numbers see growth. Nobody is lying.
The pay split
The number I’d put in front of anyone who doesn’t believe the job has split in two: Glassdoor lists “AI product manager” as a title of its own. The figures that circulated in late 2024 put it at roughly $245,000 in total pay against roughly $124,000 for product managers generally. The two numbers aren’t measured the same way and are nearly two years old, so read the gap as rough.
2.0×
the AI-focused figure over the traditional one; a skill premium is 15–50%
Ratio
Source: Pawel Huryn, from Glassdoor data, October 2024.
The numbers
| Role | Typical pay, US, late 2024 |
|---|---|
| AI-focused PM | ~$245,000 |
| Traditional PM | ~$124,000 |
Even discounted heavily, a gap that size is more than a skill premium, and it sits on a separate title. Premiums are 15%, 30%, maybe 50%. This is a different job, not a better-paid version of the old one.
And the “AI” in that title means something more specific than it sounds. It doesn’t mean you work at an AI company. It means you can do the things that don’t have artifacts: define quality for a system that gives different answers each time, own the bet rather than the roadmap, and show the thing rather than describe it.
What actually got removed
Three supports under the traditional PM role, all removed within about two years:
Specs were the only way to show an idea
The PRD existed to reduce the risk of spending expensive engineering time on the wrong thing. An idea now becomes a clickable, working thing in an afternoon.
What went with it: the document as the unit of persuasion.
Synthesis took a week
Now an hour, which means research can happen inside a decision rather than arriving after it.
What went with it: the research readout as an event.
Experiments were too expensive to run
So teams argued instead, and much of what looked like product process was argument, because the meeting was cheaper than the test.
What went with it: the meeting as the cheap option. It is now the expensive one.
Remove those three and the coordination half of product management — the translating, the specifying, the relaying, the waiting on analysts — doesn’t get automated so much as it stops being required. That’s a harsher claim, and I think it’s the accurate one.
What’s left is the decision half, which got harder: more options to choose between, products whose outputs vary every time, and new kinds of risk nobody owned in 2019.
Part three: design
The same curve, different vocabulary
Here’s what convinced me this is one story rather than several.
Companies are moving from large teams of generalist designers to smaller, more specialized teams with stronger strategic and technical skills. Designers with AI skills earn a premium of 10–15%, well below the 56% premium PwC finds for AI skills across all jobs: the same direction, at a slower pace. Roles requiring judgment are growing; roles requiring only craft are not.
Sources: UX Design Institute (2026); PwC 2025 Global AI Jobs Barometer, close to a billion job ads. The designer figure is drawn at the middle of its range.
The numbers
| Group | Premium for AI skills |
|---|---|
| Designers (UXDI) | 10–15% |
| All jobs requiring AI skills (PwC) | 56% |
Now put “product manager” in place of “designer” in those three sentences. They stay true. The numbers differ; the shape is identical.
The artifact problem, stated more honestly than product people state it
Design writing has been clearer about this than product writing, possibly because the artifact was more obviously the job.
The formulation I keep returning to: AI didn’t replace designers, it commoditized output. Wireframes, layouts, UI variations, and visual exploration are no longer scarce. Judgment is. If your value was tied primarily to producing artifacts, the market has already moved on.
And the matching hiring shift, the part that should worry anyone with a beautiful portfolio: hiring teams are no longer asking whether a designer is talented. They’re asking whether the designer can be trusted when things are unclear.
That’s a different question, and no portfolio answers it, because a portfolio is a record of finished things. It shows what you chose; it hides what you rejected and why. When choosing was the expensive part, the portfolio was a fine proxy. Now that generating options is free, the portfolio shows the cheap half of the work.
The convergence, and what it actually is
Designers are told to become “full-context”: to think end to end, to carry weight across the system rather than within a function. Product people are told to become full-stack.
These are the same instruction, issued to two disciplines, in two vocabularies.
But I want to be careful about what’s converging, because “PM and design are merging” is the lazy version. The disciplines aren’t merging. Both roles had a large execution part and a small judgment part, and the execution part is being removed from both. What’s left in each is mostly judgment, and judgment about a product looks fairly similar whether you came to it through design or through product.
The roles aren’t becoming the same. They’re becoming more similar in composition because the same part was subtracted from each.
The part design got right first
One thing the design world has put better than the product world: taste is built from lived experience — watching real users struggle through onboarding, shipping a pricing page that flops and learning exactly why.
This is the right account of where judgment comes from, and it has an uncomfortable implication neither discipline has fully absorbed: if judgment comes from verified outcomes, then years of experience without verified outcomes don’t produce judgment.
A designer who shipped forty screens and never watched anyone use them has forty screens of experience and very little taste. A PM who ran thirty launches and never checked a success metric has thirty launches of experience and almost no product sense, because the feedback loop that builds it was never closed.
That’s most of us, partly. It was certainly partly me.
Part four: the business model
This part gets written about in different publications, by different people, for different readers, and it’s the same phenomenon.
The contradiction
My favorite number in this piece. A Cruxy survey of 300 SaaS CEOs in April 2026 found that 97% plan to retire seat-based pricing within two years. 94% said seat-based pricing currently aligns with their product’s value.
Source: Cruxy survey of 300 SaaS CEOs, April 2026.
The numbers
| CEOs who… | Share |
|---|---|
| Plan to retire seat-based pricing within two years | 97% |
| Say seat-based pricing aligns with their value today | 94% |
Sit with that. Nearly every one of them believes their current pricing is correct, and nearly every one of them intends to abandon it.
That’s not confusion. It’s a proxy failing in real time. The seat count still tracks value today, for most of them, which is why 94% is honest. And every one of them can see the point on the graph where it stops tracking, which is why 97% is also honest.
The whole industry is standing on a floor it can see cracks in.
What’s actually replacing it
Here the story and the practice diverge, which is worth knowing if you’re going to write or argue about this.
The narrative is outcome-based pricing. Charge for the result — a resolved ticket, a completed workflow. Intercom’s per-resolution model is the standard example.
The practice is hybrid. A predictable base plus variable consumption on top. It’s the most common primary structure among 230+ software companies surveyed, at 37%, up from 25% a year earlier. And credits — which sit between seat access and pure outcome pricing — were the defining innovation of 2025, with adoption growing 126% year on year.
Outcome pricing is where the enthusiasm runs ahead of practice. In the same survey, a quarter of investors favored it, while the companies themselves moved to hybrids and credits.
- Hybrid (base plus consumption) as the primary pricing structure
29%
already use AI credits; another 33% plan to within a year
Credits
+126%
growth in credit-based pricing, year on year
Credits
26%
of investors favor outcome-based pricing
The narrative
Sources: Kyle Poyar, 2026 State of B2B SaaS and AI Monetization (230+ companies); PricingSaaS 500 Index.
The numbers
| Measure | Value |
|---|---|
| Hybrid as primary structure, a year earlier | 25% |
| Hybrid as primary structure, now | 37% |
| Using AI credits | 29% (+33% planning) |
| Credit-based pricing growth | +126% |
| Investors favoring outcome-based pricing | 26% |
So everyone is talking about outcomes, and almost everyone is shipping hybrids. The honest read is that outcome pricing is harder than the discourse admits. You have to define a “result” precisely enough to bill for it, and have metering the customer trusts. Those are product problems, not pricing problems, and most teams have nobody who owns them.
The trajectory
Gartner’s projection: at least 40% of enterprise SaaS spend shifts to usage, agent, or outcome-based models by 2030. The shift has started: the share of companies pricing mainly by the seat fell from 21% to 15% in a single year. IDC forecasts 70% of software vendors moving away from pure per-seat models by 2028.
Supply is ahead of demand: 83% of AI-native SaaS companies already offer usage-based pricing. The lag is on the buyer side. If you’re buying software now, that’s a negotiating window, and it closes as the models standardize.
- A year earlier
- 2025
survey, one year apart21%15%
Sources: Maxio; IDC; Gartner (via Deloitte TMT Predictions 2026); Kyle Poyar, 2025 State of B2B Monetization. The 2028 and 2030 figures are forecasts; the seat-pricing change is a survey result.
The numbers
| Measure | Value |
|---|---|
| AI-native SaaS offering usage pricing (Maxio) | 83% |
| Vendors off pure per-seat by 2028 (IDC, forecast) | 70% |
| Enterprise SaaS spend on usage/agent/outcome by 2030 (Gartner, forecast) | ≥40% |
| Companies pricing mainly by the seat (Poyar) | 21% → 15% |
The buyer got sharper at the worst possible moment
Meanwhile, across the table, buyers audit utilization and refuse to renew shelf-ware.
And vendors have been raising prices 20–37% at renewal by bundling AI features into existing products or migrating customers to more expensive AI-inclusive SKUs — a practice the people paying it now call “the AI tax.”
Index: last year’s price = 100. The gold band is a range. Source: Tropic.
The numbers
| Price, indexed | |
|---|---|
| Last renewal | 100 |
| This renewal, AI bundled | 120–137 |
So vendors don’t know what to charge for, buyers are newly rigorous about what they’re getting, and a price increase lands on top of both. That’s unstable, and it’s where much of 2026’s commercial friction comes from.
Part five: the valuations
The $285 billion
Back to where we started. The February selloff was the market applying the labor argument to the revenue model in one move: if an agent replaces five humans, charging per head makes no economic sense.
The line is the iShares software ETF (IGV), indexed to its November peak = 100, at the three points reported: the peak, the April 10 low and early June; straight lines join them. Sources: Bloomberg, via Yahoo Finance (Feb 2026); Bespoke Investment Group (June 2026).
The numbers
| Measure | Value |
|---|---|
| Lost on Feb 3 | ~$285 billion, software, data and financial stocks |
| Software stocks, Nov 2025 peak to Apr 10 low | −36.6% |
| Rally from the April low, by early June | ~+46% |
| Drawn as an index (Nov 2025 peak = 100) | 100 at the peak; 63 at the April 10 low; 93 by early June (63 × 1.46) |
I don’t think the market was wrong about the mechanism. I think it was wrong about the timing, and what came next says so. The selling ran on for weeks: software stocks fell about 37% from their November peak to an April low, then rallied about 46% by early June. That was a repricing of a belief, not of a business. Revenue didn’t move on that schedule. What moved, twice, was the confidence interval around future revenue.
That’s what happens when a proxy loses credibility. The underlying thing may be fine. The measure everyone used to value it is no longer trusted, and until a new measure takes hold, prices swing on argument rather than evidence.
The 95%
Now the number that complicates every optimistic reading of all of the above.
A 2025 MIT study found that 95% of enterprise AI pilots showed no measurable impact on the P&L, a contested figure drawn from 52 interviews and a survey of 153 leaders. Meanwhile, 64% of product teams have already integrated AI into their products.
Product teams that have shipped AI into the product
64%
Nearly everyone is doing the work.
Enterprise AI pilots with no measurable return
95%
Nearly none of it is demonstrably working — and both markets have already repriced as though it is.
Put those together and you get the state of the industry in 2026: nearly everyone is doing AI product work, nearly none of it demonstrably works, and the labor and capital markets have both repriced as though it does.
This is the most interesting fact in this essay and the least written about, because it doesn’t flatter anyone. The optimists have to explain the 95%. The skeptics have to explain why the repricing is happening anyway.
My reading: the repricing is right about direction and early about magnitude. The capability is real; the organizational ability to turn it into outcomes isn’t there yet. And the gap between the 5% and the 95% is largely a product problem — what is this for, what does good output mean, how would we know if it worked — rather than a model problem.
Which is, conveniently or inconveniently, exactly the set of questions the job I’ve spent fifteen years doing is supposed to answer.
Part six: what it means
Six conclusions, in descending order of confidence.
1. The proxy is gone in every domain, and nothing has replaced it yet
Portfolios, case studies, seat counts, and revenue multiples were all proxies that worked. All four are degraded, and no replacement has settled in any of them. That’s an unusually uncomfortable moment, and it’s why everything feels contested: we’re arguing because we don’t have an agreed measure.
2. Judgment is the residue, and it has to be demonstrated rather than claimed
Every domain arrives at the same word. None has solved how to show it. A finished artifact doesn’t. What does: the bet you made and what you predicted, the options you rejected and why, the thing that went wrong and how fast you found out.
That’s an argument for a different kind of professional record than any of us have been keeping.
3. Verified outcomes are the scarce credential
If judgment comes from closed feedback loops, then the scarce thing isn’t experience — it’s checked experience. Predictions you made, wrote down, and scored.
Almost nobody has this, in any of these disciplines, because the tools and the culture never required it. So it’s cheap to acquire right now and expensive to fake. If I were advising anyone in product or design, at any level, this would be the whole advice: start writing down what you expect to happen, and go back and check.
4. The middle is being hollowed out, in all four domains
Junior PMs face the sharpest competition. Companies are folding generalist designers into smaller specialized teams. Mid-market SaaS is caught between usage pricing it can’t meter and seat pricing that’s dying.
The pattern is consistent: the top of each market is fine, the bottom is being automated, and the middle — the competent, artifact-producing professional or the competent, undifferentiated product — is where the pressure lands.
5. Whoever solves “what does good mean” wins the next five years
Outcome pricing requires defining a billable result. AI features require defining acceptable output. Hiring requires defining demonstrable judgment. Enterprise pilots fail because nobody defined success before starting.
Four problems, one shared blocker: the absence of a rigorous, agreed definition of good. That’s not a technical capability. It’s a product skill, and it’s been the least glamorous part of the job for twenty years.
6. The correction will overshoot in both directions
The market repriced software in a day on an argument, then spent the spring taking much of it back. The 95% suggests the argument was early. That partial reversal is what usually happens when a real structural change arrives inside a hype cycle. What I’d expect next is the slower, more permanent version of the same shift.
For anyone making career decisions off this: don’t bet on either the panic or the dismissal. Bet on the mechanism, which is durable, and be patient about the timing, which isn’t.
The short version
Producing artifacts got cheap. Everything valued or priced on artifacts is finding a new level.
For product people: postings up, PM headcount down 28%, generalist roles converted into specialist ones, a new title paying roughly double.
For designers: the same curve in different vocabulary — output commoditized, judgment scarce, portfolios showing the cheap half of the work.
For software: 97% of CEOs intend to abandon a pricing model 94% of them think currently works, and the practical answer so far is hybrid rather than the outcome pricing everyone talks about.
For the market: $285 billion repriced in a day on an argument and largely restored by June, against a contested study finding 95% of enterprise AI pilots with no measurable return.
One mechanism, four vocabularies. The people who read it as four separate stories will keep being surprised by each one.
I’m writing from inside this, not above it. I’m a product person with fifteen years of the old proxy and a working theory about the new one, testing it by rebuilding my own job in public. If the theory’s wrong somewhere, I’d rather hear it.
— Chris
Chris Direduryan is a product leader with 15+ years across advertising, fintech, SaaS, and early-stage startups. He writes Bets and Ships. Figures current as of September 2026; sources linked throughout.
Sources
Where the numbers came from. Each row opens the article or report itself; where the original is not public, the row says who reported it.
- State of the product job marketLenny’s NewsletterProduct managementNewsletter2026PM openings and the AI pay gap
- Product management trendsUserpilotProduct managementArticle2026PM postings up 14% a year
- What the data tells us about tech and PM lay-offs in 2026Mind the ProductProduct managementArticle2026PM headcount below its 2022 peak
- UX designer salaries in the US: updated for 2026UX Design InstituteDesignArticle2026What designers with AI skills earn
- Why the UX job market is changing in 2026 (and most designers aren’t ready)UX PlanetDesignArticleJan 2026“AI didn’t replace designers. It commoditized output.”
- What can’t AI do in design in 2026The Fountain InstituteDesignArticle2026Taste built from lived experience
- What actually gets senior product designers hired in 2026The Fountain InstituteDesignArticle2026What hiring teams now ask
- TMT Predictions 2026: the AI gap narrows but persistsDeloitte, with Gartner’s forecastPricingReport2026At least 40% of SaaS spend off seats by 2030; seat share 21% → 15%
- The 2026 state of B2B SaaS and AI monetizationKyle Poyar · Growth UnhingedPricingReportMay 2026230+ companiesHybrid as the primary structure, 37%
- How SaaS pricing changed in 2025PricingSaaS 500 IndexPricingData2025500 companiesCredit models 35 → 79, +126%
- What’s the endgame for SaaS pricing models after the AI panic?Cruxy’s survey, as reported by UserpilotPricingArticleApr 2026300 CEOs97% plan to retire seats; 94% say seats fit today
- The death of per-seat pricingIDC’s forecast, as reported by The SaaS CFOPricingArticle202670% of vendors off pure per-seat by 2028
- The AI tax: how AI is driving software price increasesTropicPricingReport2026Renewals up 20–37% with AI bundled in
- MIT report finds most AI business investments failMIT NANDA’s “GenAI Divide”, as reported by Virtualization ReviewAI adoptionNewsAug 202595% of pilots with no measurable return
- Anthropic AI tool sparks US$285 billion rout in software, financial services, asset management sectorsBloomberg, via Yahoo FinanceValuationsNewsFeb 2026~$285B lost in one day, Feb 3 2026The February selloff
- Software strikes backBespoke Investment GroupValuationsNewsletterJun 2026IGV −36.6% peak to Apr 10 low; ~+46% sinceThe fall and the rebound
- Atlassian slashes 10% of workforce to ‘self-fund’ investments in AI and enterprise salesCNBCProduct managementNewsMar 2026~1,600 jobs; 900+ in R&DAtlassian’s cuts
- The Fearless Future: 2025 Global AI Jobs BarometerPwCDesignReportJun 202556% wage premium for AI skills, all occupationsThe all-jobs AI premium
- The state of B2B monetization in 2025Kyle Poyar · Growth UnhingedSoftware pricingNewsletter2025Seat-based as primary model: 21% → 15% of companiesSeat pricing’s share falling
- 2025 SaaS pricing trends reportMaxioSoftware pricingReport202583% of AI-native SaaS offer usage pricingUsage pricing on the supply side
The tool
Scorecard
Conclusion three says the scarce credential is checked experience: predictions you wrote down before the outcome and scored afterward. This is that, built — and the uncomfortable finding is in the case study: of eleven recorded predictions, one settled cleanly.
Read the case study →