In January, I lost a deal. The client decided to build their GTM CRM themselves, with an outside consultant. His offer: higher upfront cost, lower running costs. And above all, full automation. They came back to us recently. In between, they lost half a year.
When she returned, the director was unusually generous with me. She sent me material on how to win the next pitch against vendors who build “from scratch”. For the first time in a long while, someone who buys from us handed me something this useful in writing. Her conclusion, after half a year with a “vibe coding” wizard who never delivered: before signing, he promises a miracle. After signing, he turns the project into a “transformation”. You don’t get a solution. You get ideas on how to get to the miracle yourself.
When you buy a CRM platform, you’re not buying features. Any consultant can click those together today with a few prompts and a little “vibe coding”. You’re buying something far less photogenic - the data model, the definitions, the integrations, the AI layer, all still correct six months later.
Build your own CRM and you don’t get a CRM - you end up running a software company on the side.
Anthropic, a company with some of the best AI engineers in the world, recently published how it built self-service business analytics internally. The result reads like a fairy tale. Claude now answers 95% of their business analytics questions, at roughly 95% accuracy. A director asks in plain language and gets the answer, no analyst needed.
Then come two numbers that bring the fairy tale back down to earth. Without “skills” - the written instructions that route the agent to the right tables and definitions, like what an “active customer” even means - accuracy didn’t exceed 21% on their tests. And when they left those instructions untouched for a month, accuracy fell from about 95% to 65%. The data model changes every day. The description the AI reads goes stale within weeks.
Anthropic’s fix wasn’t a smarter model. It was engineering discipline. Every change to the data model has to update the AI documentation in the same step. Today that holds for about 90% of their changes. For all of it, they needed a dedicated data and engineering team. Their own repository, CI pipelines, evaluations, and constant maintenance.
Databox CEO Pete Caputa said publicly that it would have been cheaper for Anthropic to buy his company than what building it themselves cost them. Marketing provocation? Partly. But the point holds: building isn’t the expensive part. Keeping the system accurate is.
Analytics isn’t CRM. But the problem is the same. In both cases the AI has to understand what “customer”, “lead”, or “revenue” means in your data - and hold on to that understanding as the data shifts. Every agent sitting on top of your own data fails at the same point. Anthropic was simply the first to measure it and publish.
Anthropic can afford that team. The typical B2B buyer we work with at FrodX is a company with 30 to 300 employees. No AI engineers, no data team, and most often no GTM strategist either. These are exactly the companies solo consultants now talk into building their own CRM. The argument: lower running costs.
Here’s where it gets tricky. The running cost of an in-house system isn’t server hosting. The running cost is an engineer. Someone has to fix the integration when the ERP changes its API. Someone has to update the lead definition when you change your sales process. Someone has to make sure that in six months the AI agent isn’t answering based on a world that no longer exists. The system Anthropic built with its own tools lost 30 percentage points of accuracy in a single month. Without maintenance. It will be no different for you - there just won’t be anyone to notice.
And now ask yourself one more thing. The CRM is your core GTM system. Are you prepared to depend on a quarter of one outside developer for it? When you build on HubSpot, an entire network of partners stands behind you. Someone has to implement it at nearly 300,000 companies. Replacing FrodX isn’t hard. Replacing your solo wizard is your next half year.
Building software has never been cheaper. Maintaining it has never been more expensive.
At FrodX, a typical HubSpot implementation for this segment costs €20,000 to €30,000. Infrastructure: around €800 per user per year. For a ten-person GTM team, that’s 10 × 800 = €8,000 a year. Since April, you pay for HubSpot’s AI agents by outcome: the Customer Agent at $0.50 per resolved conversation, the Prospecting Agent at $1 per lead recommended for outreach. If the agent doesn’t do the work, there’s no bill.
Now compare that with the “cheaper in the long run” side. A quarter of one developer already exceeds the entire annual infrastructure of a ten-person team. Conservatively, €12,000 to €15,000 a year. And that’s just maintaining the code. The semantic layer, the evaluations, the security, and everything Anthropic described above come on top of that. (What it costs when the consultant you still know in January is gone by July, I’d rather not get into.)
Dharmesh Shah, HubSpot’s co-founder and CTO, recently described their twenty-year constraint. They build for companies you can’t send an army of engineers to, and for buyers who won’t pay a hundred thousand dollars a year. With nearly 300,000 customers, there’s no other way. So everything Anthropic built by hand - the semantic layer, the data model, the agents, the evaluations - has to be built into the product and maintained centrally, for everyone at once.
That’s what you actually buy with a licence. Anthropic’s team - except you split the cost with nearly 300,000 other companies.
The director from the start left me one more piece of advice in that material, and I now use it in every comparison like this. When someone promises you that “everything will be automated”, ask two things. First: which business number will move, and when will we measure it? Second: who maintains the system in six months, and what does that cost?
AI does sales administration superbly today. Prospect research, drafting outreach, follow-ups, CRM entries. What it doesn’t do is build trust, navigate internal politics, or close hard deals. A vendor who doesn’t know that difference, or won’t say it out loud, isn’t selling a solution. They’re selling bait.
The client from the start lost half a year. A system built in-house with the best tools in the world lost 30 percentage points of accuracy in one month without maintenance. Maintenance waits for no one. The only question is whether you pay the platform for it, or ignore it until it catches you.
Which brings us back to where we started: build your own CRM and you don’t get a CRM - you end up running a software company on the side.
igor.pauletic@frodx.com
P.S. If you have your own development team, AI engineers, and a GTM strategist, building your own is a perfectly legitimate path - Anthropic is proof that it works. Just be aware that you’ve become a software company. And act like one: with a repository, evaluations, and a maintenance budget. Not with one consultant and a few prompts.