You're Not Implementing a CRM. You're Paying Someone to Discover Your Company.

What to decide before you sign the next implementation, in the year AI vendors started sending engineers to live inside their customers.

By Ricardo Segura · OPEX Consultoria · November 2026

1. The second line of the proposal

The proposal is on your desk, and it has two lines. The first is the license: so many users, so much per month. The second is the implementation: process mapping, configuration, integrations, data migration, training.

Almost all the negotiation goes to the first line. The money tends to be in the second. In my experience it is larger than the license, and the customizations that come afterwards are expensive. I looked for an independent number for that ratio and didn't find one. What exists are estimates from consultancies and vendors, ranging from one to five times the annual license value depending on complexity, with a few models coming in below one. Take it as an order of magnitude.

Now look at what the second line buys. For months, an outside team interviews your people to find out how your company works. Who approves a discount. When an order becomes an exception. Why that one customer has terms no other customer has. You are paying for someone to discover what your company knows and never wrote down. At the end, part of it becomes fields, rules and workflows inside the system. The rest leaves in the consultant's head.

Which leads to the question this article answers only at the end, because the answer depends on how we get there:

When the project is over, what stays with your company?

A disclosure before going further. OPEX sells organizational memory architecture, so I have a side in this discussion. That is why I set aside a whole section, near the end, for what cuts against my position.

Until recently you could treat that question as a contract detail. In 2026 you no longer can, and the reason showed up first in the job market.

2. The engineer who moves into your company

The role is called Forward Deployed Engineer, or FDE. It is the engineer who works inside the customer, next to the people who run the operation, to configure the AI, connect it to the company's data and systems, and solve business problems. The model was popularized by Palantir and was, until recently, a niche job title.

In 2026, US job postings with that title grew more than 1,000% between January and August compared with the same period a year earlier, according to Lightcast data published by Fortune. The median advertised salary is above $188,000, against roughly $145,000 for a traditional software engineer.

What stands out is who is hiring. In April, Salesforce launched an FDE partner network with Accenture, Deloitte and more than thirty other firms, and explained why: few companies have the technical foundation to turn pilots into systems the business can actually run on. In May, Anthropic and OpenAI announced services companies to deploy AI inside customers. In October, CNBC reported a $1 billion commitment from AWS for the same purpose.

Read that as a buyer. The companies that make the best models and the largest systems reached the same conclusion: the product, on its own, doesn't make it to production. What is missing is someone who understands the customer's pain, the customer's process, the customer's exceptions. So they are sending people inside the customer to learn exactly that.

It is the second line of the proposal, now with a high salary and a billion-dollar budget.

If that knowledge is worth so much to them, why isn't your company the one cataloguing it first? The answer starts with an uncomfortable fact: the system you already have was never built to hold it.

3. What your CRM knows and what it never knew

A CRM is a system of record. It knows what stage the opportunity is in, what it is worth, who the contact is, when the last activity happened. It does that well, and the company still needs someone to do it.

What it doesn't hold was described precisely in an essay from Foundation Capital, published in December 2025, that became a reference point in this debate. The example is simple. A renewal agent proposes a 20% discount. Company policy caps renewals at 10% unless an exception is approved. The agent gathers the serious incidents the customer went through, finds an earlier approval by a vice president in a similar case, and finance signs off.

The CRM ends up with a single fact: a 20% discount. Who authorized the deviation, on what basis and on which precedent, none of that is kept.

The authors call the missing piece decision traces, and they give the set of them, linked across customer, policy and time, the name context graph: a living record that makes precedent searchable. It is what I called, in the manifesto, a decision as an immutable event with its why. Venture capital arrived at the same place by another road and put a trillion-dollar label on the opportunity. The label is their thesis, not a market measurement.

The argument that matters to a buyer is structural. Systems of record keep the current state. Data warehouses receive the information after the decision has already been made. The context is lost before it reaches either one.

That context is what the consultant digs up during implementation and what the FDE goes looking for inside your company. The big vendors have noticed the same thing, and they are moving faster than would be comfortable for my thesis.

4. The vendors are not standing still

It would be easy to say the big systems are walled gardens. In part they are. In part they no longer are.

They opened the door. Salesforce and HubSpot already offer MCP servers, the open protocol I covered in The Executive Arsenal, and HubSpot has connectors for Claude and ChatGPT that read and write CRM records. In practice, whichever AI assistant your company chooses can already query and update the system. The argument that these systems don't talk to anything has aged.

They are building the memory. At Dreamforce 2026, according to analyst coverage, Salesforce presented a context and memory layer shared across agents, with general availability indicated for December. In May I wrote that there was no off-the-shelf product for the Second Brain. I need to update that sentence: the container is becoming a product.

And they are making money. In the quarter that ended in July 2026, Salesforce reported revenue of $11.3 billion, up 11%, and Agentforce annual recurring revenue above $1.5 billion, up more than 240%. Whoever announces the death of the CRM hasn't read the earnings release.

But notice one detail about the open door. In May 2025, Salesforce changed the terms of use of the Slack API. Third-party applications can still read messages in real time, and they lost the right to store and index them over the long term. Querying stayed allowed. Taking the data with you became forbidden.

Put the pieces together. The vendor offers the system, the memory layer and the engineer who comes to live with you. It is a complete offer, and for many companies it will be tempting. But it changes the nature of the decision: you stop choosing a piece of software and start choosing whom you will depend on.

5. Who ends up depending on whom

The hasty reading of all this is that the AI solves the problem and the engineer just installs it. The facts point the other way. The model is becoming a commodity: any assistant already connects to the big systems, and swapping one for another costs less and less. Knowledge of the process still belongs to each company.

AI executes well whatever can be codified and has volume. It doesn't know which problem is worth solving, or why the company decides the way it does. Someone has to know. I see three answers to who that someone will be, each with a different cost.

The vendor's engineer. It is rented knowledge. They learn the process, build, and leave. What they learned goes with them or stays recorded in the format of whoever sent them. It solves the lack of engineering, which is real, and creates a new dependency.

Your own trained people. Better, because it stays in-house. But it can be the fifth pain from the manifesto wearing a different badge: the leader who becomes a cognitive bottleneck. If what the trained person knows stays only in their head, the company has swapped the irreplaceable senior manager for the irreplaceable AI specialist.

Written memory. It is the only one of the three that reduces the other two. With processes, exceptions and reasons on record, the outside engineer becomes a temporary translator, and the trained person becomes the memory curator I described in the previous article and stops being a single point of failure.

In that architecture, AI is infrastructure, people own the process, and the FDE is scaffolding: it goes up to build and comes down when the structure stands. The vendors themselves admit where this is supposed to end. AWS says its engineers work with the customer's team so that the customer can maintain and expand the systems on its own. Anthropic includes in its training program engineers who are already employed by customers and partners.

On paper, it holds together. The board will ask a different question: does it pay?

6. Where the money showed up

It does, for a few. And the profile of those few keeps repeating.

In McKinsey's 2026 global survey, 80% of the people who use AI at work say their own productivity improved. Only 37% attribute any effect on EBIT to AI. And the high-performing group, which attributes at least 5% of EBIT to it, stayed flat at 6%. The distance between "my productivity improved" and "the company's results changed" is the picture of thirty uncoordinated ChatGPTs.

What do the 6% do differently? Nearly three quarters of them redesigned their workflows. Among the other companies, about one quarter did. The Wharton–GBK study from October 2026 adds a finding any board can use: those who measure AI by growth report strong returns almost three times as often as those who measure efficiency alone, 35% against 13%.

Nobody redesigns a process that hasn't been described.

The 2025 MIT NANDA study, known for the headline that 95% of generative AI pilots fail, explains the other side. More than half of the budgets go to sales and marketing tools, while the biggest return showed up in back-office automation. And the tools stall in corporate use because, according to the report, most generative AI systems don't hold on to feedback, don't adjust to context and don't get better with use. That is the description of a memory problem.

The most solid work I found shows what happens when that problem is solved. Brynjolfsson, Li and Raymond, in a study released in 2023 and published in 2025 in The Quarterly Journal of Economics, followed about 5,000 customer support agents who started using an AI assistant. In the 2023 version's numbers, issues resolved per hour rose 14% on average and 34% among novices and the less skilled. Among the most experienced, the effect was minimal. The authors' explanation, presented as suggestive evidence: the model spreads the best practices of the most capable agents and shortens the experience curve for the new ones.

It is the knowledge that lived in the heads of the best people, made available to everyone and measured in a real operation.

In Brazil, the obstacle comes earlier. In a September 2026 survey by Zappts, 55.1% of companies have no structured mapping of the return on AI, and 19.8% measure it clearly and on a recurring basis.

Those are other people's numbers. I have a case of my own, a small one, that I can tell from the inside.

7. The math I did in my own company

In April 2026 I retired the CRM at OPEX. The reasons are on record in the decision written that day. The subscription didn't pay for itself in a one-operator business. The system's model assumed one seller and one buyer, and every deal of mine has four parties: buyer, intermediary, manufacturer and rep. The available AI connector had limitations that broke the routine. Today the source of truth is a collection of text files, with dated decisions, written policies and a daily ritual running on top of them.

The operation is small, it has one decision-maker, and I write code. The decision itself records the condition under which I would revisit it: the day there is a sales team and the operation stops being single-user. On that day, permissions, audit trails, reporting and scale weigh in again, and a CRM is back on the table.

The difference is how I would arrive at that conversation. With the process described, the exceptions catalogued and the pains ranked by return. The vendor wouldn't need to discover how the company works, only translate what is already written.

Before I say what I would do in your place, I owe you the list of what may be wrong in all of this.

8. Where this thesis may be wrong

Those who build alone tend not to get as far. In the MIT study, solutions bought from specialized vendors or built in partnership reached deployment about twice as often as those built internally. Fortune cites about 67% for the first group. Understanding your own pain is different from knowing how to build the solution.

And yet the best build more. In the McKinsey survey, nearly half of the high-performing companies build their solutions in-house, against 31% of the rest. My reading is that both things hold: the minority that has engineering and a described process builds and wins, and the majority that doesn't is better off buying.

The most cited case doesn't say what people say it does. Klarna is remembered as the company that swapped its CRM for AI. The CEO told a different story: they did not replace software with a language model. They built an internal stack, with a graph database, to bring into one place the knowledge scattered across their systems. About the rest of the market, he was direct: "Will all companies do what Klarna does? I doubt it." And he added that he didn't see it as the end of Salesforce.

Storing is the easy part. Forrester analyst Charles Betz points out that pieces of this decision layer have existed for decades and that the real difficulty is the fidelity of what gets stored. Without governance, the record becomes a diary everyone writes in and nobody reads. And many of the most valuable decisions happen in meetings and on phone calls, leaving no digital trace.

The data is fragile. Almost everything I cited about returns is executives answering questionnaires, and McKinsey and GBK sell consulting in this area. CNBC notes that the gains attributed to FDE projects have not yet been independently verified. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. I found no study comparing companies that started with memory against companies that started with the system. My own case is a single one. And the three dependencies in section 5 are my reading.

None of this overturns the order I am about to propose. All of it limits how much I can promise with it.

9. What to sign, and in what order

Back to the opening question: when the project is over, what stays with your company?

What stays is what was already written, in the company's name, before the project began. The rest stays with whoever discovered it.

That is why the order matters more than the choice of vendor.

First, memory. One area, one decision-maker, ninety days. Record decisions with their why, write down the unwritten rules, note every exception on the day it happens. Here AI does the manual labor that used to make this unworkable: it transcribes the meeting, reads the email, proposes the record, chases what was left open. The person curates.

Second, diagnosis. With three months of trail, the pains show up without interviews. Where the process breaks. Which exception keeps repeating. Which approval stalls the sale. Each one gets an estimated return, and the list is ranked by it. This is where you find out how much of the problem was a badly designed process, and a badly designed process inside an expensive system is still badly designed.

Third, system. Buy, renew, switch or build. With the specification written by the company itself, the second line of the proposal shrinks, the comparison between vendors becomes objective, and the memory is born in a format that survives replacing any of them.

There are cases where the order is reversed. A large sales team, a regulated industry, audit and permission requirements from day one, or no in-house engineering capacity at all. In those cases the CRM comes first, and the memory layer is built alongside it, from the start, in the company's name.

Buying is not the problem. Brazil buys: in the TIC Empresas 2025 survey, CRM use rose from 25% to 31% of companies, and AI use from 13% to 17%. The problem is buying without knowing what your own company already knows.

If the proposal is already on the table and there are no ninety days, take five questions to the meeting:

1. What will the implementation discover about our company, and where is that written down when the project ends?
2. Can I export, beyond the records, the history of decisions and exceptions, in a format another tool can read?
3. Does the contract allow the AI I choose to read and keep this data, or only the vendor's AI?
4. How does the price behave when agents do the work of users: per seat, per consumption or per outcome?
5. If the vendor's engineer comes to live here, who on my team learns from them, by what date, and what gets documented in our name?

And CRMs, do they survive? They do. The revenue numbers don't show an industry dying, and not even the authors of the context graph thesis are betting on that. What changes is where the value sits. Recording the state becomes infrastructure, necessary and increasingly similar from one vendor to the next. The value moves to whoever holds the context of the decisions.

The vendors will build the second brain into their systems, and that will tend to shorten implementation. If your company's memory is born in there, the project gets faster and the exit gets more expensive. If it is born in a format that belongs to the company, readable by any model and any system, the CRM becomes a replaceable part of an architecture that is yours.

You're not implementing a CRM. You're paying someone to discover your company. Discover it first, in writing and in your own name. Then sign.

Memory before intelligence. And, for the same reason, memory before the system.

If you have one of these proposals on your desk, let's talk before you sign. No agenda, no deck, no proposal. Just a conversation between operators.

Ricardo Segura is the founder of OPEX Consultoria, which provides organizational memory architecture services and therefore has a commercial interest in the thesis of this article.

References

1. CRM implementation cost estimates (sources with a commercial interest): hyphadev.io and codleo.com
2. Fortune (Preston Fore), on the growth in forward deployed engineer postings, Lightcast data, Sept. 3, 2026 — fortune.com
3. Salesforce, FDE Partner Network announcement, Apr. 15, 2026 — salesforce.com
4. TechCrunch (Russell Brandom), "Anthropic and OpenAI are both launching joint ventures for enterprise AI services", May 4, 2026 — techcrunch.com
5. CNBC (Barbara Booth), "AWS, Anthropic race to deploy army of specialized engineers", Oct. 15, 2026 — cnbc.com
6. Foundation Capital (Ashu Garg), "AI's trillion-dollar opportunity: Context graphs", Dec. 26, 2025 — ashugarg.substack.com
7. Salesforce Developers, "Salesforce Hosted MCP Servers Are in Beta Today", Oct. 2025 — developer.salesforce.com ; and Salesforce, "Introducing the Data 360 MCP Server", May 26, 2026 — salesforce.com
8. HubSpot, AI tools page (remote MCP server and connectors) — hubspot.com
9. Info-Tech Research Group (Igor Ikonnikov), technology note on Dreamforce 2026, Sept. 25, 2026 — infotech.com
10. Salesforce, second quarter fiscal 2027 results, Aug. 26, 2026 — s205.q4cdn.com
11. CX Today, on the change to the Slack API terms — cxtoday.com
12. McKinsey (QuantumBlack), "The State of AI: Global Survey 2026", Aug. 2026 — mckinsey.com ; figures read in TechTimes coverage, Aug. 26, 2026 — techtimes.com
13. Wharton Human-AI Research and GBK Collective, 2026 AI adoption study, Oct. 13, 2026 — gbkcollective.com
14. Fortune, on the MIT NANDA report "The GenAI Divide: State of AI in Business 2025", Aug. 18, 2025 — fortune.com ; Virtualization Review, with excerpts from the report, Aug. 19, 2025 — virtualizationreview.com ; and DX, "The AI Divide" — getdx.com
15. Brynjolfsson, Li and Raymond, "Generative AI at Work", NBER Working Paper 31161 (2023); The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889–942 (2025) — nber.org
16. TI Inside (Matheus Bracco), on the Zappts study "Panorama da IA no Brasil 2026", Sept. 8, 2026 — tiinside.com.br
17. TechCrunch (Julie Bort), "Klarna CEO doubts that other companies will replace Salesforce with AI", Mar. 4, 2025 — techcrunch.com
18. Forrester (Charles Betz), "Context Graphs Are A Convergence, Not An Invention", Apr. 10, 2026 — forrester.com
19. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", June 25, 2025 — gartner.com
20. CGI.br / Cetic.br, TIC Empresas 2025 press release, June 15, 2026 — cgi.br