Practical AI · Episode 54 · Deep dive

Chess,
or scrambling?

The biggest names in business software all moved this month. The interesting question is whether they know what they are doing.

What happened

Four moves in six weeks

The evidence

Eleven moves in seven months

Feb 3
A lab ships Claude Cowork. Within 48 hours, a reported 285 billion dollars comes off software stocks.
April
HubSpot rewrites its own pricing. Free trials and outcome based pricing, knowing it will slow sales down.
May
HubSpot opens everything. Any agent can drive the CRM. Full API parity.
Jun 16
Revenue Hub. Quotes, contracts, billing and payments onto the records agents read.
Jul 21
Block releases Buzz. Humans and agents in one open source workspace.
Jul 23
Agent Hub. Halligan: marketing app 2006, suite 2012, CRM platform 2018, agent platform 2026.
Jul 29
The Claude connector gets much bigger. Leads, campaigns, emails, quotes, revenue.
Aug 5
Guidance lowered, and Google's former cloud apps president joins the board. Same day.
Aug 25
Claude's memory gets much better. Chat and Cowork share one memory. It now builds while you talk instead of after.
Aug 26
Claudeforce. Salesforce inside Claude. Claude inside Salesforce and Slack. Announced with earnings.
Aug 27
YouSpot. One dollar a month. 154 slots gone the first day.

Nothing on this list existed before February. That is the fact both readings have to explain.

Reading one

It looks like chess

Reading two

Or it looks like
staying relevant

None of that is failure. It is what smart people do when the threat is real and nobody has solved it yet.

Bet one

Salesforce sells permission

The move

Stop fighting for the screen. Get inside every screen instead.

What they are really selling

Governance. Their customers have compliance teams, approval chains and auditors. An agent cannot touch that data without something vouching for it. Salesforce is selling itself as the thing that vouches.

The risk

It is still a pilot. Nobody outside it has run a quarter on it.

And the bigger one

If the model companies build their own governance, Salesforce is selling something its partner can make itself.

Bet two

HubSpot sells access

The move

Marry nobody. Let every agent drive the CRM equally.

What they are really selling

Access, not permission. Their customers mostly have no admin, no IT and no compliance team. Governance is not their problem. Getting something working this afternoon is.

And a second bet, at one dollar

A CRM for the one person company, built by the co-founder inside an internal group. It goes below HubSpot's own floor, to a customer HubSpot has never sold to.

His own boundary

"It's not for GTM teams. I recommend HubSpot." A company funding the thing that could undercut it, and telling you when not to buy it.

In their own words

Three sentences,
two companies, one idea

"Here, the UI is the AI."

Marc Benioff, Salesforce. August 26.

"We are big believers in the idea of headless. Not big believers in the notion of humanless."

Dharmesh Shah, HubSpot. Q1 earnings call.

"Under the hood, it's actually implemented as an AI harness with CRM tools inside."

Dharmesh Shah, launching YouSpot. August 27.

That last one is the whole story in one sentence. Not a database with AI added. An assistant, with the database as a component inside it.

The part nobody mentions

It is also a fight
over the meter

Software has been priced per person for twenty years. Agents do not need seats.

"Prospecting agent, one dollar per qualified lead. Customer support agent, fifty cents per resolved conversation. Both agents work well and now have aligned incentives."

Brian Halligan, HubSpot co-founder. May 2026.

Nobody is buying a license for a person in that sentence.

The market · one

How far below
their own record

Salesforce HubSpot
0% -80% Salesforce -31% HubSpot -71%

Salesforce peaked near 363 dollars in December 2024. HubSpot peaked at 881 in February 2025.

The market · two

Then Wednesday happened

Salesforce HubSpot
0% +25% Salesforce +22.6% HubSpot +7.9%

Salesforce's second best day in its history. HubSpot rose almost eight percent on no news of its own, which tells you the whole category was being repriced.

The counterweight

Over ten years, HubSpot
is the better investment

+16% a year
HubSpot over ten years. And it is down 71 percent from its high.
+12% a year
Salesforce over the same ten years.
23 vs 92
Price to earnings. HubSpot is still four times more expensive after falling 45 percent.

This is a repricing of the future, not a failing company. Almost nobody makes that distinction.

Two things nobody reported

Read the filing,
not the press release

HubSpot's slowdown was a decision

"The first was deliberate. We knew this would create some near term headwinds, but we believe it's the right long term trade off."

Yamini Rangan, chief executive, August 5

It ran everywhere as a miss. They chose slower growth now to move customers onto AI pricing sooner.

Salesforce owns a piece of Anthropic

Not just a partner. An investor since 2023, in a stake valued around five billion dollars in June.

And in the same quarter

The legal filing records 2.7 billion dollars of unrealized gains on that stake. It is not in the press release. All disclosed, nothing improper, and unmentioned in every story about the announcement.

Meanwhile, in the real world

Almost nobody fills in
the CRM they pay for

76%
of companies say less than half their CRM data is accurate and complete.
65%
of sales reps spend five or more hours a week typing into a CRM by hand.
3%
have automated that data entry.
6%
of everything added to my own CRM this year was typed in by a human. The rest arrived on its own.

That gap is the prize. Every move on these slides is aimed at it. And the last number cost nothing, took no migration, and already works.

Two AIs, same week

I asked another AI
the same question

Where we agreed

  • Both bets rest on one assumption: that an AI remembering you is not the same thing as a company's shared record of what is true.
  • Holding the record is the weakest of the three moats. It is the most copyable thing either company owns.
  • The reacting reading has real evidence. Neither of us could rule it out.

Where we split, and I lost

  • I said the strongest moat was permission: being the thing that decides what an agent may do, and keeps the receipt.
  • It said distribution wins, because permission only pays off if agents actually start acting. Today they mostly draft and wait for a human.
  • That is the better answer. A moat that depends on a future that has not arrived is not a moat yet.

It also stated four things as confirmed that turned out to be secondhand quotes from articles, not company filings. Two AIs reading the same week produced sharper thinking and the same sourcing problem a human has.

So which is it

Both readings fit
the same facts

If it is chess

  • They become invisible plumbing that every agent has to route through, and they get more valuable, not less.
  • The 2026 crash was the market misreading a transition.

If it is scrambling

  • Agents get good enough to keep their own memory and the record becomes a commodity anyone can hold.
  • The crash was the market being early, not wrong.

Wednesday was the market choosing chess. It has not chosen wrong. And the only part of this whole story with proof behind it is the boring one: connect the record to the assistant, and stop typing.