Back to blog
· 10 min read

Re: Founding

HubSpot calls its reset a refounding. What Intercom, Salesforce and a new CRM called Dreamteam show about starting a software company again.

“If we were starting HubSpot today, knowing what we know, what would we do?”

Dharmesh Shah calls this the core question in his first note on the HubSpot refounding. He co-founded HubSpot in 2006, and on 12 November his co-founder Brian Halligan came back as CEO with the words “This is a refounding, and I can’t wait to start!” Shah has been posting notes about it since. On the third day he wrote that the emails between him, Halligan and Duncan Lennox, HubSpot’s president of product, “have been flying”. A company was being founded again, over email.

When I first read his question, my reaction was doubt. “Refounding” is a lovely word, and I have seen a lot of AI announcements this past year that were really layoffs with a new name. So I wanted to know whether refounding is a real need or a new word for cutting staff, and I went and checked what the companies doing it have actually changed.

Intercom rebuilt everything

Intercom is the case I learned the most from, because it started early and finished. Eoghan McCabe came back as CEO in October 2022 after five quarters of slowing growth. ChatGPT arrived a few weeks later.

Fin’s own account of the decision, written by Anne Marie Kingsland, says: “AI might actually be both the death of our company and also the best chance we have of survival.” Over the next three years, a later essay says, “we rebuilt almost everything we’d spent the previous eleven years building, including the name.”

When the rename came in May 2026, McCabe wrote that the company “is now finally, fully reborn and its name is Fin”, and admitted he should have “pulled the trigger on this change much sooner”. A month later Salesforce agreed to buy it for about $3.6 billion.

The thing I noticed is how much of the company changed together: the product, the price, who uses it (an agent answering customers instead of support staff), the organisation and the name. That is a long way from a memo.

Memos that change one thing

Most AI announcements of the last two years changed far less. Tobi Lütke’s Shopify memo in April 2025 made “reflexive AI usage” a “baseline expectation” for staff. Duolingo said it would stop using contractors for work AI can handle, then walked it back after the backlash. Fiverr cut about 30% of its staff to become “AI-first”. Salesforce cut its support team from about 9,000 to about 5,000, with Marc Benioff saying “I need less heads” (TechRadar).

These change how a company works, or how many people do the work. The product, the price and the customer stay where they were. So I put the cases against five questions, and counted how many answers each company has changed:

What we sellHow we chargeWho or what uses itHow we are organisedName and values
Shopify memoYes
Duolingo memoYes
FiverrYes
Intercom to FinYesYesYesYesYes
HubSpot (announced)YesYesYesYes

HubSpot’s answer on name and values is a small one so far. On the third day of the refounding, Shah wrote that the company keeps its mission, shortened to “Help millions grow better”, but one word in it now means something else: “In v1, better meant ‘with inbound marketing’. Now, it means ‘with AI’.” When I first read it, the change looked tiny. Then I saw it was the whole refounding in small: keep the name, and change what it means.

The memos change one answer. The refoundings change most of them at once, and the biggest change is in “what we sell”. For a software company right now, that question has a particular shape.

Records, agents and the middle

There is a line I keep hearing in software these days: in future, software will be only two things, systems of record and agents. The records hold the truth, and the agents do the work.

The best-known version is Satya Nadella’s, on the BG2 podcast in December 2024. Business applications, he said, “are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents.” I read the transcript of the full answer, and he goes one step further: once the logic lives with the agents, “people will start replacing the backends.”

I don’t agree with the “only two things” part, though. Even a system of record needs something on top that an agent can call: an API, an MCP server, and the rules about what an agent is allowed to change. What gets thinner is the part built for people to click through, the screens and step-by-step workflows that most of us have spent our careers building. The middle stays, but more and more of it is written for agents.

HubSpot’s own quarterly report, filed on 4 November, eight days before the refounding, shows this happening to one company. Under risk factors it says customers “increasingly authorize third-party AI platforms and agents” to work inside HubSpot through its “Model Context Protocol server”, which “may reduce direct engagement with our platform.” It also warns that its new usage credits “may substitute for, rather than add to, existing seat and hub spend.”

The customer numbers show the pressure, though the filings do not say what caused it. HubSpot ended September with 311,704 customers, but the number it added each quarter had shrunk: 10,752 in the first quarter of 2026, 6,988 in the second and 5,258 in the third, going by its quarterly reports. I added them up against the year-on-year change in the 10-Q, and they match exactly. The third-quarter count is also the first to include subscribers to YouSpot, a new personal CRM, so the real slowdown may be a little steeper. Jason Lemkin of SaaStr wrote: “Customers were mostly staying (101% NRR). But new ones fell by half in just 2 quarters.”

That is the need, in the company’s own words and numbers. The people clicking through the product are being replaced by agents, and the price per seat starts to go with them.

Refounding is choosing a layer

Read that way, Shah’s question gets narrower for a software company. If software is becoming records, agents and the layer the agents call in between, which part do we want to be?

The companies that have gone furthest have each picked an answer.

Intercom picked the agent, and renamed the whole company after it. Fin charges $0.99 for each conversation it resolves.

HubSpot chose the records and the layer agents call on top of them, although it builds agents of its own too. Shah wrote that HubSpot is “leaning strongly into a headless version”, one that AI agents in Claude and ChatGPT use through MCP. Halligan’s first post called it “the Customer Harness Company”. The records and the rules stay at HubSpot, and the agents can come from anywhere.

Salesforce did both, in order. It rewrote the layer on top of its records for outside agents with Headless 360 in April, and then it bought Fin, closing the deal in September. It kept its records and bought someone else’s agent.

Two warnings for whoever holds the records

HubSpot’s choice is the one I am least sure about, because owning the records was not enough last time.

In travel, Amadeus and Sabre held the airline and hotel inventory, the system of record for the whole industry. Then online travel agencies built the screens people actually booked on. Ashu Garg told this story on Foundation Capital’s podcast, and the share prices still tell it: in late 2026 Amadeus was worth roughly $25 to 30 billion and Sabre under $1 billion, while Booking Holdings alone was worth about $120 billion. Market value is a rough measure, and travel had other forces at work, but the gap is hard to miss.

Bars of market value in late 2026. Booking, which built the screens: about $120 billion. Amadeus, which held the records: about $27 billion. Sabre, which held the records: under $1 billion. Caption: market value in late 2026, by who held the records and who built the screens.

The records stayed with the old companies, and the biggest prize went to whoever owned the place people came to. HubSpot is betting that its records stay valuable even when the agents, and the place people talk to them, belong to Anthropic and OpenAI. I don’t know yet whether that bet is right.

The second warning came on 16 November. Girish Mathrubootham, who founded Freshworks, answered HubSpot directly: “A new harness on an old horse doesn’t make it a racehorse.” His point is that a new layer for agents does not help much when the record underneath is the problem. CRMs were built for people typing into forms and fields, so “a 30-minute customer conversation gets reduced to a handful of fields and notes,” and no agent can recover what was never captured. “You cannot build an AI-native future on a human-native foundation,” he wrote. He makes a second point too: “The CRM doesn’t understand the business it serves.” It stores contacts, accounts and opportunities, but not “your products, your people, your processes”.

On the same day Dreamteam, a CRM started in 2025 by three former Freshworks leaders, with its engineers in Chennai, announced $53 million from Together Fund, Accel and Premji Invest. It pulls context from emails, meetings and conversations into what it calls a Company Brain, and runs its agents on that. The same TechCrunch article gives the opposite view too: Andreessen Horowitz argued earlier this year that existing CRMs will probably stay where customer records are kept.

So there is one more answer to Shah’s question: rebuild the record itself. HubSpot is betting its old records are worth opening up to agents. Dreamteam is betting they have to be replaced. Of everything I read for this post, this is the disagreement I found most useful, because I can see both sides of it: the record HubSpot holds is huge, and I also know how much of a customer conversation never makes it into a CRM field.

The third layer: a place to check

Then I read the founders again, more slowly, and noticed that the screen had not gone anywhere. In a second post Shah wrote, “Sometimes the interface that we so fastidiously worked on for years just needs to fade into the background.” His word is “fade”, and in the same post he calls a “hybrid conversational/chat interface” the “better gateway”. One reply said: “chat can do the digging. i still want a link to the record.”

Nadella says it too, a few minutes before his CRUD remark: agents need permissions so that, “when it comes to actions, it is verifiable.” Andrej Karpathy put the split of work plainly in his talk at Y Combinator in June 2025: “usually they are doing the generation and we as humans are doing the verification.” His tool for faster checking is the screen. “GUIs are very useful for auditing systems,” he said.

Foundation Capital’s Ashu Garg and Jaya Gupta give an example of what goes missing without one. A renewal agent offers a customer 20% although policy caps renewals at 10%, gathers the evidence for an exception and gets Finance to approve it. The CRM records “20% discount”. The reasoning and the approval are kept nowhere unless something else keeps them.

That is the loss Girish described above, seen from the other end. In both, the record keeps the result and loses the why. With people doing the work, the why lived in their heads and their emails. With agents doing it, it has to live somewhere.

Girish’s other point is a second gap, before the agent acts at all: it has to understand the business. The patch and the rebuild are both promising this. HubSpot’s own 10-Q calls its Smart CRM “the foundational context layer that combines customer data with AI”, and Dreamteam’s Company Brain is its answer to the same problem. Context also means knowing what the company’s words mean. Jamin Ball of Altimeter points out that sales, finance and legal can each have their own ARR number, so an agent has to be told which one counts.

At least one reader is asking for this out loud. Under Shah’s third-day note, they asked whether agents will be “first-class users of the CRM, with permissions and audit trails in the open, or will that stay buried in settings?” And Dreamteam, for all its agents, leaves “sales teams retaining control over what gets sent to customers”, in TechCrunch’s words. I read that as a sign the checking seat is real, even if nobody has built a company around it yet.

So the middle I said would stay has two sides. One faces agents: the API, the MCP server, the rules. The other faces people, and its job changes from typing data in to checking what the agents did. That is the third layer I would add to the two: systems of record, agents, and a place where a person checks what the agents did. In the cases I looked at, every refounding bet on the records, the agents, or both. None of them has refounded around the checking place yet, and I think that seat is still open.

Put together, I can picture one way an agent’s work could go round. It is one possibility, and the companies above are each building different parts of it.

What it costs, and the question for the rest of us

None of this is cheap. When Brian Chesky reset Airbnb, “almost the entire exec team left. They left or they didn’t want to work for me,” he told Fortune. Chegg cut 22% and then about 45% of its staff and has still shrunk. The most complete refounding here ended with the company being sold, and whether that counts as success depends on who you ask.

So my doubt at the start came out half right. A lot of what gets called AI transformation is a memo or a layoff. But where agents change the product, the price and the user all at once, which is what I found at HubSpot, Intercom and Salesforce, refounding looks like a real need. The companies going furthest are choosing a layer, and they are doing it on a clock. Dreamteam’s co-founder Senthil Kanthaswamy gives new companies “two to five years” before the established CRMs finish adapting. If he is right, that is roughly how long an older company has to refound.

The people inside are the ones who live through it, rebuilding eleven years of work in three, as Fin’s staff did. Shah’s question works for them too, and for anyone who builds software. We can ask it of the product we work on: if we were building it today, knowing what we know, which layer would it be: the records, the agents, or the place where a person checks their work?

Co-written with AI. Credit the prose, blame the opinions.