On May 15 at 10:13 PM, Aaron Levie posted on X that he is fully forward deployed engineering pilled. It is Friday evening when he writes it; by Saturday morning the post has 83 thousand views and a serious conversation underneath it. What Levie is saying in public is what CTOs of large companies have been saying in private for the last three quarters.

What matters more is what Tanmai Gopal, co-founder of Hasura, posted in reply hours later. He named the part of the FDE model that does cost money every day, and that your next AI contract will quietly bury in the fine print.

Levie’s argument, in its buyer-side version

Levie is CEO of Box. He does not sell consulting. He sells enterprise SaaS. His tweet is not pitching his own service; it is describing what he observes from the other side of the contract. That gives it weight.

His argument, plain: “In software, you deliver a stable piece of technology to a customer and they adopt it and that’s that… In AI, you’re delivering something that is constantly evolving both due to the nature of the new capabilities and best practices that emerge, but also because the underlying models change so much that they can meaningfully change the workflow as a result of their upgrades.”

From there he jumps to the economic conclusion that is not obvious: “it’s far more logical that one vendor can share best practices across thousands of companies more efficiently than every single company can learn and manage these best practices themselves.”

That is the logic of distributed moat. A vendor seeing a hundred implementations per quarter has a learning floor no individual customer can buy with their own internal hours.

For a mid-sized B2B company in Central America the translation is direct. If you are still thinking “we will hire an internal AI engineer and figure it out in-house,” the opportunity cost is high. Not because of your team’s technical ability. Because of the speed at which what needs to be known keeps changing underneath.

The crack Tanmai Gopal opened

Hours later, Tanmai Gopal answered. Worth reading his credential first: Tanmai is co-founder of Hasura, a GraphQL infrastructure company with real enterprise customers and significant operational experience with technical deployments. He is not a Twitter commentator. When he says something does not work, it is because he tried it and it broke.

His reply, word for word: “I was very FDE pilled last year but I ran into problems that were hard to solve. The problem was capturing and maintaining ‘shared context’. It’s spread across multiple peoples, changes every day and hence is prohibitively expensive to capture.”

The objection is surgical. Tanmai is not arguing whether FDE creates value. He is naming the budget line that breaks the model the moment you try to industrialize it: customer context is not an artifact you capture once in week eight of onboarding and file away. It is alive. It changes with every new decision, every meeting where someone says “we are not going down that road anymore,” every customer whose intake quietly modifies the playbook. Keeping that alive costs money every day.

And if you do not keep it alive, the model collapses on its own. When the FDE rotates off the account — because big companies do rotate them; that is how the model is wired — the context walks out with the badge.

The part that genuinely compresses

Here you have to separate two things public discourse keeps mixing together.

The FDE function itself — a human embedded in the customer, translating between technology and business — has been done by consulting firms since 1990 under three different labels. I wrote yesterday about exactly that: the craft changes its business card every fifteen years but the actual work is the same.

What changes now is not the function. It is the technical possibility of capturing the context that function produces, and keeping it alive in a form that does not die when the person rotates. That part is genuinely new.

In 1995 the Andersen consultant filled physical folders with notes that stayed in his laptop. In 2015 the Accenture strategist set up a corporate Confluence nobody ever opened again. In 2026, for the first time, we can run a persistent layer underneath the embedded human that ingests transcripts automatically, extracts open items with citations to the source, flags contradictions when they appear, and translates decisions into living artifacts. We have been running this in production for over a month for a real services company: senses, memory, nerves, immune system.

That layer is what breaks Tanmai’s objection. It does not eliminate the cost. It changes the nature of the cost. It stops being a human cost that repeats every time the team rotates and becomes a cost of keeping the infrastructure alive. And infrastructure compounds. Every new engagement starts further along the curve because the brain already captured precedent, patterns, and resolved contradictions.

The four questions a CFO actually should ask

The typical CFO at a Central American company of 80 to 800 people does not need a framework to evaluate AI. They need four concrete questions to ask any vendor walking in with an AI transformation deck.

What is the monthly cost of keeping shared context alive, separate from implementation hours? If the answer is “we don’t charge for that,” the vendor’s model depends on the customer absorbing the cost on their own, without seeing it. It arrives in the month-18 invoice dressed up as expansion.

What happens to the captured context when the FDE rotates off my account? If the answer is “the next one picks it up” without a concrete system that shows exactly how, what is being sold is a transfer miracle that does not exist.

What is the renewal logic if the underlying model changes next quarter? If Claude 5 or Gemini 3 forces a rebuild of half the deployment, who pays for that? The answer matters most when the answer is “we absorb it,” because that is only sustainable if the vendor has an abstraction layer that does not depend on a specific model.

If I cancel at month six, which artifacts remain my property and which leave with you? This question cleanly separates the vendor selling sovereignty from the vendor selling dependency. Both are legitimate offerings, but knowing which one you are buying is the only thing a CFO is obligated to understand before signing.

These are not adversarial audit questions. They are the questions an honest vendor welcomes, because they make clear what is being sold and why it is priced the way it is priced.

How we wire this at IQ Source

I will be specific about how we answer each of the four questions, because the point of this post is not to preach in the abstract; it is to show the wiring.

AI Maestro is two months of discovery. What we deliver is not a report; it is a Process Reality Map — a map of how the company’s processes actually work, not how the manual claims they do — plus an AI Opportunity Score per process and an honest Go/No-Go gate at the end. Sometimes the answer is do not deploy agents; that is part of the value. The Process Reality Map is shared context captured as an asset, not as a deck that dies in a Drive folder.

Technology Partner is the build phase: integrations into a legacy HubSpot or SAP, evaluations, guardrails against hallucinations, change management on the human side. During this phase the brain keeps absorbing everything that happens: every decision, every scope change, every meeting transcript. The context stays alive without anyone having to remember to capture it.

If our human counterpart at IQ Source rotates in month ten, the next person starts by reading the brain, not by running re-discovery interviews. That is what structurally lowers the cost Tanmai identified as prohibitive. We do not eliminate it. We move it to a budget line a CFO can see, predict, and approve.

What changed in 36 years

I have been in computing since 1990, when I started at fifteen on a Commodore 64. Five cycles of enterprise knowledge management: filing cabinets, relational databases, corporate intranets, Notion-style wikis, now AI brains. The first four times the underlying problem was the same: when the human who knew left, the knowledge left with them. The tool never closed the gap because the tool was not a nerve. It was a shelf.

What changes this time is not that the human stops rotating. What changes is that the system underneath the human can capture, keep alive, and deliver the context without being asked. That is the part specifically possible in 2026 and not before. Levie is right that FDE becomes a core competency. Tanmai is right that shared context is the part that breaks the cheap version. You don’t have to choose between them. You need the Levie model running on infrastructure that already solves Tanmai’s objection.

The question of the next contract

When you sign your next contract with an AI vendor, read it line by line until you find where the word “context” sits. If it is not there, ask where it is. If they tell you “we cover that in implementation hours,” mark the clause in yellow. If they tell you “we have a concrete system that keeps context alive and we will show it to you in demo,” ask for the demo before you sign.

That demo is the cheapest test you will run this quarter. And it will cleanly separate the AI vendors building the distributed moat Levie is describing from the vendors selling hours with a new business card.

Audit my context capture with AI Maestro