The Death and Rebirth of the CRM

Pick the one piece of software every company pays for and no one opens. It's the CRM.

Sales avoids it. Managers nag about it in every one-on-one. And the one thing it exists to hold, what's actually happening with customers, mostly isn't in it.

So this is the dissection. What killed the CRM, what's keeping it on a ventilator, and what grows back.

Start with the body on the table. A CRM is a place where a person types a summary of work they already did, so other people can read a rough version of it later.

That's the entire function. Someone has a call, then logs a shadow of the call. Someone sends a proposal, then drags a stage to stand in for the proposal. The record is a tracing of work that happened somewhere the software couldn't watch.

That tracing is what I mean by the distortion layer. Software whose only job is to let you read your own work back, late, partial, and one step removed from the thing itself.

And the tracing was never accurate. Every sales manager has run some version of this exchange.

Run that across a whole team, every week, and the picture is clear. The CRM doesn't hold what happened. It holds a slow, edited, occasionally invented version of what happened, shaped by what someone remembered to type and whether an integration fired.

For most of three decades, companies paid people to keep this tracing by hand. Not because anyone wanted it, but because machines couldn't read the real work. Typing the shadow was the only way to get the work into a system at all.

That constraint just vanished. An AI can read the calls, the inbox, the calendar and the thread directly. The instant it can, the reason to type the shadow is gone, and so is the reason to pay for the software that collected it.

Salesforce read the situation early.

This spring it shipped a headless product. Open APIs, agent access, and an explicit bet that the value was always the data underneath, not the screen on top. Most of those interfaces had been there for years, so the announcement was more positioning than engineering. The positioning is the interesting part. Pull the screen off a CRM and what's left is a database. So what was the screen ever for?

Seema Amble of a16z wrote the cleanest answer in a piece called "Is Software Losing Its Head?". Her argument: agents dissolve muscle memory as a moat, but the operational logic underneath survives, because an agent needs the rules, the approvals and the edge cases spelled out before it can safely act. AI cheaply rebuilds the easy 80% of a system of record. The hard 20% is the part worth defending.

She's right about the 20%. I'll come back to it, because it doesn't last the way she thinks.

But look at how she describes the old moat. The Salesforce screen, she says, got thousands of reps to "enter data they otherwise wouldn't have." She means that as the strength of the product. Read it twice and it's the confession.

Reps fought the data entry because the data already existed. It sat in the sent folder and the call log and the calendar. The screen's whole purpose was to make a human copy out, by hand, what the machine couldn't read yet.

That last word is the one that matters. Yet. The machine reads it now.

So the 80/20 line isn't quite the right cut. The 80% isn't cheap to rebuild. It's pointless to rebuild, because it was only ever a copy of work that an agent can now read at the source. And the 20% was never the record. It's a different thing wearing the record's clothes.

The left column is the distortion layer. Read the underlying activity and it reconstructs itself, for about the price of the tokens. The right column never lived in the activity. It's the set of rules that decide whether an action is allowed, and an agent needs them spelled out more than a human ever did. A tired rep at least hesitates before approving a bad deal. An agent signs it at 2am unless a rule stops it.

Here is where the current fix makes things worse instead of better.

The plan at most companies is to wire the old CRM to the agents through MCP and call the job done. Watch what that actually builds.

An agent reads the inbox and the calendar and works out the state of a deal. It writes that state into the CRM, translated into stages and fields. Hours later a second agent opens the CRM and reads those fields back, to answer a question like "what's slipping this quarter?"

Two conversions, each one a step further from the truth, and both of them unnecessary. The second agent could have read the same inbox the first one read.

The agent already had the raw activity. We taught it to rewrite that activity into a 2005 data model so a second agent could decode it back out. Calling that integration is generous. It's a relay for an AI talking to itself.

MCP gets sold as the bridge that keeps the old software alive. It's actually the proof the software was always a middle step. If an agent can write the record and an agent can read the record, the record is a holding pen between two machines that were already in conversation.

That pen made sense while the reader was a slow human who needed the work flattened into rows. Take the human out of the loop and it's a ritual the data performs on the way back to where it started.

So why hasn't the CRM died already? One reason, and it has nothing to do with how good the software is.

In much of the world, an employer can't simply point an AI at its staff's email and chat. The honest version of the agent pitch is "let the model read everyone's inbox and work out what's going on." In Europe, that sentence runs straight into the law.

Under the GDPR, an employee's messages are personal data, and feeding them to an AI is a fresh purpose that needs its own legal basis. Monitoring staff systematically on that scale also triggers a formal data protection impact assessment before anything is switched on. None of it is a toggle.

Germany goes further than anyone. Section 87(1) No. 6 of the Works Constitution Act gives the works council a veto over any technical system capable of monitoring employee behaviour or performance. The Federal Labour Court reads "capable" loosely. A tool doesn't have to be built for surveillance, it only has to be able to do it, which is why even Microsoft 365 and Teams fall inside the rule. An AI reading the company inbox is the clearest example there is. Where a works council exists and hasn't agreed, the rollout simply doesn't happen. By one industry count, around 60% of AI projects in German firms stall on exactly this.

You can see how narrow the escape hatch is. A Hamburg labour court let staff use ChatGPT on personal accounts in 2024, but only because the employer couldn't see the usage. Give the company an AI that can watch what employees do, and the veto snaps back into force.

France walls it off from the other direction. The 2001 Nikon ruling held that an employer can't read messages an employee marks personal, even on a work computer, even where personal use was banned outright. Secrecy of correspondence isn't only an employment matter there. It's criminal, under Article 226-15 of the penal code, carrying up to a year in prison and a 45,000 euro fine. An AI told to read everything sweeps up precisely the correspondence the law puts off limits.

Put those together and the CRM's real job comes into focus. It's a laundering step. The work gets copied into a system the company is allowed to read, because the place the work actually happened is fenced off by law. The distortion layer survives because the source is forbidden, not because the copy is any good.

The American case is the giveaway. Under the federal wiretap statute, the ECPA, employers can read communications on their own systems through the business-use and consent exceptions, and staff have no reasonable expectation of privacy on company kit. A handful of states, New York since May 2022, plus Connecticut and Delaware, ask only for notice, not permission. By 2024 some 78% of US employers already monitored digital activity. The thing Europe is still wrestling with, America mostly settled by handing employees a policy to sign.

That policy is the future of the entire arrangement. Not a CRM. A checkbox, where the employee agrees that their employer's AI may read their work communication, and the agent then reads the source with nothing in between.

The barrier propping up the CRM is real. It's also temporary. A law that forces work through a copy can be answered with a single consent, and once it is, the copy has no reason to exist.

This isn't a bet on repeal. Works councils will keep negotiating these terms, the EU will keep adding process, and the consent will arrive wrapped in agreements and assessments. But the direction runs one way. Every signature that lets an agent read the source pulls another brick out from under the CRM.

Strip the CRM down to what survives a world where the agent reads the source, and the leftovers are small.

A number that has to commit somewhere exact: the booked deal value finance reports on. The part that actually sends the messages, because sending is an action with consequences and a summary isn't. And the rules from a moment ago, the limits and approvals an agent checks before it acts.

Now the 20% Amble files as the moat. I said it doesn't last the way she thinks. Here is why.

The 20% is undocumented context. The exception nobody wrote down, the approval that lives in one person's head, the compliance rule passed on by word of mouth. On day one it's genuinely hard to copy, and that's real protection.

But it depreciates, for three reasons that all point the same way.

The reconstruction tools keep improving. Computer-using agents already pry context out of systems that were never built to hand it over, and each model is better at it than the last. What needs a human this year gets read by a machine the next.

The context writes itself down as it gets used. Every time an agent hits an exception and a person resolves it, that resolution is logged. Run that for a year and the undocumented 20% slowly hardens into a written rule sitting on the rails, portable like everything else.

And the people who hold the context are the ones agents replace. The day the veteran who knew the edge cases leaves, those edge cases have to be written down to survive at all. Once they're written down, they stop being a moat and start being a file.

So the 20% isn't a moat. It's a runway. It buys an incumbent time measured in years, not quarters, and the only thing worth doing with a runway is taking off before it ends.

That's the part most incumbents will get wrong. They'll guard the leftover 20% as if it were the prize, while it erodes underneath them. The ones that make it will spend the runway buying something that doesn't depreciate: the action the agent takes, and the customer the agent runs through. The record was never going to be the moat. It was the distraction.

Step back from the table. The work is the source. The agent reads it. A short list of rules says yes or no. An action commits, and a number lands where finance needs it. Nothing in that loop is a screen anyone logs into, and nothing in it is a copy anyone types.

The CRM as a place people visit is finished. What comes back is smaller, quieter, and barely looks like software: a few rules, one engine that acts, one number that counts. The reporting, the data entry, the pipeline nobody trusts, all of it goes.

Three decades of paying a tax to read our own work through a worse copy of it. The copy is what dies. The work was always the thing.

None of this was ever really about the CRM. The CRM is just the first organ to fail.

Step back and look at what an agent reading all your work actually hands a company. Total visibility. Every call, every commit, every message, seen and understood. There's a word for visibility that complete, and it isn't a soft one.

Pointed at a person, that's 1984. The telescreen that never blinks, the watcher in every room. It's what free societies are built to keep out, and it should stay out. Surveillance that total over a human life is a nightmare, full stop.

Pointed at the work, the same eye is common sense. The machine that would be a horror in your living room is a relief on your pipeline. Nobody wants Big Brother in the bedroom. Everybody wants Big Brother to write the status report.

1984 for a citizen is a horror. 1984 for a workflow is just efficiency. The whole question is whether the eye is on your life or on your work.

Reporting only ever existed to feed a blind machine. The status slide, the weekly update email, the ticket dragged across a board, the deal stage you remembered to change. Each one is a person narrating their own work to a system that couldn't watch it happen. Give the system eyes and the narration is dead weight.

So the watching falls on the work, and the time that used to go into describing it comes back. The written reporting doesn't get streamlined. It disappears. What shows up isn't Big Brother in the room. It's a Friday with nothing left to file.

One ritual outlives all of it, the one that only looks like reporting. The standup, the Monday sync, the meeting where a team faces itself in a room. That was never for the machine. It's theatre, and people need the theatre. An agent can tell you what shipped overnight. It can't make a group of humans feel like a team. So the meeting stays. The paperwork it used to generate does not.

Which is why the CRM is only first in line. Project management is next. It goes the same way.

A ticket in Linear, or a card in Monday, is a person retyping work that already exists in full. It's in the branch, the pull request, the review comments, the message where someone actually decided. The board traces over all of it, kept current by people who would rather build the thing than narrate it.

An agent reads the source. It knows what shipped, what's stuck, and what moved, with nobody touching a card. The status the board existed to display becomes a question you ask, answered straight from the work.

The purest reporting tool of all sits one layer over: the dashboard. A dashboard owns nothing. Metabase, Looker and Tableau draw a picture of numbers that live in a warehouse, refreshed whenever the pipeline last ran. Ask the agent for last week's figure and you get it from the warehouse directly. The picture on the dashboard becomes a window nobody walks up to. What survives is the agreed definition of the metric, because the agent needs it to avoid inventing the number. The charts that displayed it are done.

Same diagnosis, same cause of death. A tool whose only job was to hold a copy of work the machine can finally read for itself.