At this year’s company retreat, we went paddling on the Kolpa river. The best part wasn’t the paddling - it was the hours of riverside conversation with colleagues I don’t work with every day. One question kept coming back: “Igor, what will we be doing in five years?”
It didn’t come out of nowhere. LinkedIn is full of freelancers who “vibe-coded” a CRM, a support platform or a marketing automation tool over a weekend - something thousands of engineers at HubSpot, SAP and Salesforce have been building for twenty years. If we take them at their word, much of what we call business platform implementation today belongs in a museum. My answer by the river came a little too quickly: “We’ll be context engineers.”
Then I got home and looked at the stock market. People with far more money at stake are asking the same question.
HubSpot has lost roughly half its market value this year alone. The business, meanwhile, hasn’t fallen apart: revenue grew 23% in the first quarter and 20% in the second. It now has 306,446 customers, roughly seven thousand more than three months earlier. And when it reported better-than-expected results on 5 August, the stock lost almost a fifth of its value the next day.
The sell-off even has a name: “SaaSpocalypse” - the fear that AI will devalue subscription software because companies will build more and more of it themselves. Sound familiar? Those are my LinkedIn freelancers, dressed as equity analysts.
The market isn’t panicking without reason. Growth is cooling, new customer numbers are below plan, and AI is shrinking seat counts and compressing user interfaces. HubSpot has already started moving part of its AI offering away from classic seat-based pricing toward credits and outcome-based pricing.
If HubSpot mostly sells seats and interfaces, that’s a problem.
If it sells something else, the story gets a lot more interesting.
In June I wrote that building your own CRM is a company’s most expensive side project. Today I’d go a step further: the most valuable part of a CRM isn’t the application. It’s organizational memory.
You can build an interface surprisingly fast these days. Forms, pipelines, reports and workflows - AI will produce them ever faster and cheaper. What it can’t produce is ten years of customer history. It doesn’t know why John cancelled his contract three years ago. It doesn’t know that companies in a certain industry only buy after the third meeting. It doesn’t know which proposals died on price and which ones won on service. And it doesn’t know what “an active customer” even means at your company.
At FrodX, every third project is a platform migration, so I know first-hand what actually moves. We transfer the records in a few weeks. Meaning takes the longest to build.
You can buy the model. You can’t buy the context.
This year, HubSpot CEO Yamini Rangan told investors something more important than most of the company’s AI demos: the more agents run on HubSpot, the more valuable its context becomes. And the more valuable the context, the stronger the platform.
That is a very different defence of the CRM from “more features”.
HubSpot calls it “Growth Context”: the context that lets an agent understand not just the data, but what the data means.
Four acquisitions over the past few years show this is more than a marketing story. Clearbit added data about companies and contacts: who they are. Cacheflow added transactions, payments and renewals: what they buy, and when. Frame AI added conversations and other unstructured signals about the relationship: what they say. Dashworks added documents, email, meetings and data from other apps: everything else the company knows about them.
None of these acquisitions adds another big user interface. Every one of them extends the memory.
HubSpot started as an application people log into to get work done. It may be turning into a platform fewer people log into - and more agents use.
If so, seat count is no longer the most interesting metric.
We see this every week in our work at FrodX. HubSpot’s Prospecting Agent can write a surprisingly good sales email. It can also write very nicely formatted spam.
Give it just a name, a job title and an email address, and it improvises. Give it website visits, downloaded content, past conversations, proposals and meetings, and it changes. It starts to behave like a salesperson who prepared properly before the call.
I’ve written before about how this same agent found me a customer. Not because FrodX has smarter AI than anyone else. It had years of our context.
HubSpot sees the same pattern with larger customers. Jotform, a form-building platform with more than 35 million users, taught the Prospecting Agent its positioning and handed it the entire prospecting motion: 625,000 credits a month. In a direct test, the agent qualified leads on par with human reps.
Customer Agent, meanwhile, according to HubSpot, now resolves an average of 70% of conversations on its own; the best teams reach 90%.
The models got better, yes. But the same models are available to your competitor - and to a freelancer with a free weekend.
The difference is no longer just the intelligence of the model. The difference is what the agent knows.
Traditional SaaS mostly produces transactions. You open a ticket. You move a deal. You send a campaign. You log a call.
A good agentic system should leave the company a little smarter after every completed task. A support agent resolves an issue and links the symptom to the cause. A sales agent learns which argument moved the buyer.
The next agent can already use that.
With every use, the system accumulates organizational capital - and that is the part you can’t copy over a weekend.
This changes the economics of business software. Until now, we measured its value by the number of features, users and processes. If I’m right, the market may be putting too much weight on how many people use the application - and too little on how much useful context the system accumulates and how well it can apply it to the next decision.
Which is why, with AI, I care less and less about how much work it automates, and more and more about a single question: after the work is done, does the system know more than before?
If not, you have automation.
If yes, you’re building an asset.
There is one more problem: context has a shelf life.
Change a product, a process or a definition, and part of the old context becomes wrong. Add a new system, and the agent suddenly sees only half the story. Someone changes what a field means, and the documentation stays the same. A data point without a definition means two different things to two departments. And the agent doesn’t know any of this unless someone tells it.
Anthropic measured how fast this happens on its own analytics agent. Without active maintenance of the documentation that explains the data model, its offline accuracy slid from roughly 95 to 65 percent in a single month.
So context is not a data migration you do once.
It’s infrastructure.
It needs to be connected, cleaned, explained, constrained and repaired - and above all, it has to keep up with the business. More data does not mean more context.
Context only exists once data has meaning.
Today I would answer my colleagues a little differently.
I don’t know whether we’ll implement HubSpot and SAP Engagement Cloud in five years the way we do today. Almost certainly not. AI will automate the configuration: connecting systems, mapping data schemas, writing integrations, documenting changes. Interfaces will shrink, some applications will disappear, fewer people will log into a CRM every day. Much of what the market sells as implementation today will become faster, cheaper and less of a differentiator.
And that’s fine.
It would worry me if our value were in clicking through an interface.
It isn’t.
At FrodX, we don’t start with the fields you need in your CRM. We start with where the company is leaking growth: customer acquisition, conversion, retention, growing existing customers, or the cost of serving them. Technology comes after. AI doesn’t change that order. It changes where the value is created.
The harder problem sits before the implementation and after it. Someone has to decide what the company is optimizing for. Then, what the system needs to remember about customers to get there. What is a good customer? Which signal means buying intent? When may a service agent approve compensation on its own, and when must it hand the decision to a human? And then someone has to check whether any of it actually moves the numbers.
That’s why, to me, context engineering is not a new name for CRM implementation. It’s connecting a business goal, processes, data, rules and technology into a system that helps people and agents make better decisions. Implementation is part of it. The result is the measure.
So the answer from the Kolpa still feels right - only the reason has changed. Not because AI won’t be able to build context. It will. Our job will be to make sure it builds the right context for the right business problem.
Because context without a business goal is just a very well-organized pile of data.
P.S. FrodX also makes its living implementing HubSpot and SAP Engagement Cloud, so read my thesis knowing where I stand. The market is currently betting that AI reduces the value of SaaS platforms. I’m betting it mostly reduces the value of their interface. If I’m right, the most valuable part of the platform will be the part the user never sees.