Have you noticed how quickly a market story can harden into gospel? One quarter everyone treats software platforms like cash machines. The next quarter the same names get treated like leftover tasks waiting for an agent to wipe them out. I keep coming back to a simpler question, the one too many portfolios skipped: if intelligence gets cheap, what still stays scarce?
Why Cheap Intelligence Moves Value Toward Data
For about a year, software names absorbed a brutal repricing. Call it fear, call it fashion, call it a tidy slogan. The working assumption was blunt. If an agent can draft the email, update the ticket, or nudge the deal, why pay for the old stack? That framing treats a company as a pile of chores. It is neat. It is also incomplete.
The better test is uglier and more useful. Does the firm own something an agent cannot work without? Not a pretty interface. Not a feature checklist. Something stubborn. Trusted records. History. Permissions. The messy trail of what a customer actually did last Tuesday.
That is basic economics dressed in new clothes. When an input floods the market and the price of that input falls, value slides toward the scarce complement. Intelligence is the input getting cheaper. Trusted proprietary data is the complement that does not magically appear because a model got a new benchmark score.
Lower cost of intelligence increases the value of incumbent data.
I have found that investors love the first half of that sentence and underweight the second. Models leapfrog each other every few months. They look more capable. They also look more interchangeable. Open alternatives keep arriving. Well-funded labs keep shipping. Improvement cycles may even speed up if self-improvement loops get real traction. Fine. Then the moat is not “we built a slightly smarter brain this quarter.”
The Wrong Question Behind The Software Selloff
Markets love a simple villain. Over the years the list has included weather, trade pacts, even cultural noise that had almost nothing to do with cash flow. The latest version asked whether an agent can replace a product. Seats were supposed to melt. Attrition was supposed to spike. Pricing power was supposed to vanish.
That story vaporized a staggering amount of software market value. Roughly two trillion dollars, if you add up the damage across the group. Some of that was earned. Plenty of firms really are thin wrappers around generic workflow. Others sit on data gravity that agents still have to orbit.
Perhaps the most interesting aspect is how quickly “can AI do the task?” crowded out “where does the work get stored, audited, and reused?” An agent closing a sale still needs a place to research the account, log the call, park the contract, and tailor the terms. Without a trusted system of record, you get confident nonsense at scale. Junk in, junk out. Old line. Still true.
What The Latest Salesforce Print Actually Showed
One large customer-data platform just delivered a quarter that made the replacement narrative look sloppy. Revenue grew about 11 percent to $11.35 billion. Bookings ran hotter. Current remaining performance obligation rose 14 percent. Non-GAAP operating margin landed at 34.1 percent. Adjusted earnings came in near $5.90 a share against a consensus that sat closer to $3.27. Management lifted full-year guidance toward as much as $46.4 billion.
Those are not “one lucky beat” numbers if the surrounding mix holds. Combined AI and data annual recurring revenue reached $3.9 billion and more than tripled in a year. Agent product ARR jumped from about $100 million to more than $1.5 billion in roughly 18 months, up around 240 percent year over year. That is not a side experiment anymore.
Look at the volume underneath the dollars. Data infrastructure ingested about 104 trillion customer records in the quarter, up 355 percent year over year. Agents delivered 3.2 billion units of agentic work, nearly double the prior quarter. Work creates more records. Records make the next agent more useful. That loop is the part markets priced as if it did not exist.
| Signal | Recent Print | Why It Matters |
| Revenue | $11.35B, +11% | Demand did not roll over |
| AI + data ARR | $3.9B, more than 3x in a year | New stack is becoming core |
| Agent ARR | $1.5B+ after 18 months | Adoption is not theoretical |
| Records ingested | 104 trillion, +355% YoY | Data gravity is accelerating |
| Agentic work units | 3.2 billion, nearly 2x QoQ | Usage compounds the moat |
| Operating margin | 34.1% non-GAAP | Pricing panic looks overdone |
Skeptics said seats would shrink. Several core clouds still grew seats year over year. Skeptics said customers would walk. Attrition sat near the low end of the company’s history. Bookings more than doubled quarter over quarter in the commentary that followed the print, and contract length improved across segments. Agents, in other words, were using more of the platform, not less.
The replacement story sounded tidy. The operating numbers did not cooperate.
Agents Still Need A Place To Land
Here is the part that feels obvious once you sit with operators instead of model demos. An agent is not a company. It is a worker with no filing cabinet unless you give it one. It can draft. It can recommend. It can even execute inside a workflow. Then the output has to live somewhere that legal, finance, sales, and service all trust.
That is why the “data business” framing is not marketing fluff. If you own the system where customer context accumulates, you are not merely selling seats. You are selling the substrate. Models can swap. The substrate is sticky because ripping it out means losing history, permissions, integrations, and audit trails.
- Research still needs a reliable account record, not a hallucinated summary.
- Interactions still need a log that other teams can reuse next week.
- Contracts still need storage that finance and legal will accept.
- Customization still needs past terms, not a generic template.
In my experience, people overrate first-contact magic and underrate second-contact memory. The second contact is where money sits. Renewal. Expansion. Escalation. Handoff. Agents amplify that loop only if the memory layer is clean.
When Model Labs Buy The So-Called Obsolete Stack
If frontier labs could run the enterprise without the incumbent layer, they would. Instead, nine of the top ten AI companies in the company’s telling now run on its CRM and collaboration suite. Combined spend from that cohort jumped 435 percent year over year. That is an awkward data point for the disintermediation camp.
Partnerships followed the same logic. The headline collaboration paired a leading model shop with the CRM incumbent and framed the models as dependent on customer systems rather than replacements for them. You can argue about branding. Harder to argue with the direction of travel: the people building the brains are paying for the filing cabinets.
That is the tell. When the supposed winners of “pure intelligence” write bigger checks to the system-of-record vendor, the value map is shifting in public. Not overnight. Not for every ticker. But the direction is clearer than the slogans from last winter.
A Flywheel Markets Still Price Too Cheaply
Start with agents doing work. Work creates events. Events become records. Records make the next agent cheaper to aim and safer to trust. Safer agents get used more. Usage deepens the data store. The data store raises switching costs. It is not poetry. It is inventory.
Agent flywheel, stripped down: Agents produce work Work produces trusted records Records improve the next agents Better agents pull more spend into the same system
Once that loop is running, raw model quality still matters. It just matters less as a standalone equity story. You can swap the engine. You cannot casually swap the mileage log, the service history, and the signed paperwork.
Management is backing that view with capital, not just slides. A $25 billion buyback, described as the largest in company history, is a loud way to say the balance sheet and the thesis are aligned. Buybacks are not proof. They are a signal of confidence when the market has spent a year treating the category as roadkill.
Not Every Software Name Gets This Gift
This is the part where a cheerleading piece would stop. I will not. Plenty of software firms offer functionality and little else. If the product is a thin task layer with weak data rights, weak workflow lock-in, and no unique context, agents can route around it. Some already do.
Balance sheets matter too. A company with sturdy free cash flow and light leverage can fund the messy middle of an AI rebuild. A company servicing a heavy debt load may spend the next two years feeding lenders instead of feeding product. That split will get wider, not narrower.
- Ask whether the firm owns proprietary data agents actually need.
- Ask whether customers trust that data enough to let agents write back into it.
- Ask whether cash flow can fund integration work without emergency dilution.
- Ask whether switching costs come from history, not from a pretty screen.
- Ask whether usage creates more data, or just more chat transcripts nobody stores well.
If the answers come back soft, the SaaS scare was not a hallucination. It was a filter. The market applied it with a sledgehammer. Filters still have a use.
Commoditized Models Change Who Captures The Margin
Watch the model race for a minute and you can see why “the lab wins everything” feels dated. Leadership rotates. Benchmarks get stale. Prices get competed down. Capable open weights from well-resourced overseas labs add another pressure valve. Recursive improvement, if it arrives in a meaningful way, would compress the uniqueness of any single model even faster.
None of that makes models worthless. It makes them look more like electricity and less like a luxury brand. Electricity is essential. Utilities still have to fight over who owns the meters, the wires, and the billing relationship. In enterprise AI, the meter is usage inside a trusted system. The wire is integration. The bill is recurring revenue tied to data plus agents, not to a one-off demo.
I’ve sat through enough vendor pitches to know the difference between a prototype that wows a board and a workflow that survives a quarter-end close. The second one needs permissions, lineage, and a place where a human can still say no. That unglamorous layer is where incumbents with real customer graphs can reassert pricing power.
Pricing Power Was Supposed To Die. It Did Not.
The bear case needed margin compression. Instead, the print showed expanded profitability alongside faster bookings. That combination is hard to dismiss as accounting theater when remaining performance obligation is also up and guidance moves higher.
Could one quarter fade? Of course. Enterprise cycles still exist. Budgets still freeze. Implementation still slips. I would be foolish to pretend otherwise. Even so, the mix of AI ARR, record ingestion, and agent work units is a different shape than a classic license bounce.
Net new annual contract value growth was described as the strongest in four years. That line matters because it pushes back on the idea that customers are only defending old spend. New spend showed up. It showed up in the layers tied to agents and data, which is exactly where the thesis said it should show up.
How To Read The Broader Software Complex From Here
Generalize carefully. Do not turn one winner into a blanket bid for every subscription ticker. The market already tried a blanket theory in the other direction and look how tidy that was until it was not.
Sort the universe into two rough buckets. First, firms whose product is mostly execution logic. Second, firms whose product is a living archive of customer reality. The first group has to prove that agents make their interface more necessary, not less. The second group has a shot at charging rent on the archive while models get cheaper around it.
- Customer graphs with years of interaction history.
- Workflow systems that already sit in the path of revenue or service.
- Permission models that enterprises will not rebuild for fun.
- Data products that can feed agents without leaking the crown jewels.
Those traits do not guarantee a rerating. They do give you a map when the next scare headline lands. And there will be another scare headline. There always is.
What Operators Quietly Optimize For
Talk to people who run revenue teams and you hear a less cinematic wish list. They want fewer tools that pretend to be a source of truth. They want one place where the account is not a rumor. They want agents that can act without creating a second, shadow CRM in a chat window.
That preference is why “agents using more of the platform than ever” is not just a victory lap. It is a distribution detail. If the agent lives inside the system of record, the vendor captures usage. If the agent lives outside and only peeks in, the vendor becomes a passive database with a shrinking user interface. The quarter argued for the first outcome, at least for now.
Collaboration seats growing alongside service and sales seats is another quiet tell. Agents do not only hit the CRM object model. They hit the places where humans still argue about what the CRM means. Threads, tickets, deal rooms. The messy social layer around the record. Ignore that layer and you will misread stickiness.
A Practical Checklist Before You Rewrite A Software Thesis
If you manage money, or if you just want a cleaner way to argue at the dinner table, try this sequence. It is not a model. It is a filter I actually use when the narrative gets loud.
- Is the core asset a task or a memory bank?
- Do agents write back into that memory bank, or only read from it?
- Is unit growth showing up in data volume, not only in demo counts?
- Are the model vendors customers, partners, or both?
- Can the firm fund integration without starving the dividend of optionality, meaning buybacks, R&D, or both?
Answer those without the slogans and a lot of tickers look less doomed. A smaller set looks more interesting than the index did during the panic. That is usually how these rotations start. Not with a manifesto. With a print that refuses to fit the joke.
Risks That Can Still Break The Story
Let me be plain. Execution risk is real. Agents can create support load. Data platforms can get messy faster than governance teams can clean them. Customers can stall after a pilot because legal wants a longer look at write permissions. None of that is exotic.
Competitive risk is real too. Other platforms will try to become the landing zone. Some will win specific workflows. A model lab could push harder into system-of-record territory. Switching is painful, not impossible, especially for mid-market accounts that never fully implemented the last stack.
Macro risk does not vanish because one software name printed well. If budgets freeze, AI line items slip with everything else. The difference is that a company with a deepening data flywheel can wait. A company selling only novelty cannot.
Abundance in models does not erase scarcity in trusted context. It advertises it.
Why The Category Panic Felt So Convincing
It felt convincing because demos are vivid and databases are boring. A generated email looks like the future. A permissions matrix looks like homework. Markets pay for theater until the quarter arrives and the theater has no remaining performance obligation attached to it.
It also felt convincing because some products deserved the hit. There are tools that automated a narrow job and called it a platform. Agents can swallow those jobs. Good. Creative destruction is not a villain. Misapplied destruction is. Painting every subscription business with the same brush was the mistake.
I keep a private rule for moments like this. If the popular question can be answered with a product screenshot, it is probably the wrong question. Screenshots do not capture switching costs. Earnings, bookings, record volume, and who is paying whom do.
What “Structural Winner” Should Mean In Practice
Structural does not mean the stock only goes up. It means the company’s scarce asset becomes more valuable as the abundant input gets cheaper. That is a slow claim. It shows up in mix shift, in multi-year contracts, in customers who add agents without subtracting seats, in model shops that become bigger accounts instead of replacements.
On that definition, a firm sitting on one of the largest enterprise customer-data repositories starts in a better seat than a firm selling a clever form builder. The first firm can meter intelligence against context. The second firm watches intelligence walk off with the form.
Will every quarter look like this one? Unlikely. Will the argument keep coming back whenever model prices fall again? Yes. That is the point of a complement. It keeps getting more important each time the input gets easier to buy.
A Cleaner Way To Talk About AI Winners
Stop asking who has the smartest model this month. Ask who owns the scarce input the smart model still has to beg for. In consumer toys, that might be distribution or unique content. In enterprise selling and service, it is customer truth that has been cleaned, permissioned, and reused for years.
That framing also keeps you honest about hype inside the winner set. Data quality is uneven. Some “records” are junk with a timestamp. Volume without trust is just a bigger haystack. The companies that win will be the ones that make write-backs safe enough for a chief revenue officer to live with.
So yes, I am more constructive on software names that pass the complement test than I was during the slogan phase. Not because models failed. Because they succeeded enough to become ordinary. Ordinary inputs do not collect the rent. Complements do.
The Question Worth Keeping On Your Desk
If raw intelligence keeps getting cheaper, who gets paid to make it usable on Tuesday morning inside a real company with real customers and real auditors? That is the job description hiding under the earnings print. It is not glamorous. It is operational. It is also where a lot of the next decade’s software margin will sit if the flywheel keeps spinning.
Keep the checklist. Ignore the catchy collapse narrative when the bookings refuse to collapse. And when the next model drops and the timeline fills with eulogies for software, ask the only question that aged well this time: what does the agent still cannot operate without?
