Manouk Draisma

The through-line

Across the LinkedIn posts, Manouk Draisma keeps circling back to one question: does anyone actually know what their AI is doing, and what it is costing them to do it. Early posts frame this as an observability problem, teams wanting to see "what their LLMs were actually doing in production" as "a safety net for after something went wrong" [12]. Over the following months the question sharpens twice. First it becomes a testing question: "Testing / evaluating your AI (agents) has gone from a 'we'll get to it eventually' line item to the first question on the table" [12]. Then it becomes a spend question, as enterprises she talks to can generate a bill but not an explanation: "I saw the latest AI bills, but I have no clue where all this is actually going" [1], and "a good third of it wasn't going to shipped code at all" [2]. The constant underneath all of it is that the invoice, the trace, or the backlog ticket is never the real unit of value. The real asset, in her framing, is the corrections and judgment that get poured into the system and usually leak out unnoticed [9].

On AI spend and where it actually goes

Draisma repeatedly returns to the gap between what companies pay for AI and what they can explain about that payment. She quotes a fintech CTO who is "rolling out AI faster than almost anyone" yet still says "I saw the latest AI bills, but I have no clue where all this is actually going" [1]. Digging into Claude Code bills with a customer, she found "bloated context, skills nobody used, MCPs wired up wrong, compaction quietly burning tokens by default, and a fair few people running side projects on the company account" [2], concluding that "spend on its own is a KPI that punishes the wrong thing" [2]. She frames the underlying anxiety plainly: "Developers and leaders alike are watching token spend increase month over month, without clarity on where the waste is coming from" and worry about "what would the same usage cost when" VC-subsidized token prices end [6]. Even at the level of enterprise governance, the question enterprises keep opening with has shifted from which agents are approved to "can you list every AI agent running in the company right now?" [1].

On testing and shipping with proof

The second recurring theme is that observability after the fact is no longer enough; teams want proof before something ships. "Teams don't just want to see what their AI did anymore. They want to know whether it actually works, before it ships" [12], including partners who admit they were previously "shipping on vibes, no evals" [12]. This extends to how she treats her own product's documentation: "We run simulations that check whether Claude can read our docs, instrument LangWatch, and write its own scenarios, entirely on its own. If it can't, that's not Claude's problem. That's our docs failing the test" [10]. Her conclusion from that exercise is a broader claim about the audience for documentation itself: "In the agent era, your docs aren't for humans anymore. They're an interface for AI agents. So we test them like one" [10].

On who gets to improve the agent

Draisma frames a structural mismatch: "Agents are doing legal, health, and financial work now. But the engineers building them aren't lawyers, doctors, or bankers. The domain experts are. And until now, their contributions died in a backlog" [8]. Her fix is to let non-engineers make reviewable changes directly: someone "describes the change in plain language. What lands is a pull request, with tests, best practices included. Reviewed by your engineers, or it doesn't merge" [8]. She's explicit that this doesn't remove engineers from the loop, it just changes what reaches them: "an engineer gets a reviewable change instead of a ticket. That split matters more than it sounds" [7].

On what counts as the real asset

Building on Satya Nadella's "Reverse Information Paradox," Draisma argues that the knowledge handed to a model in daily use, not the model itself, is the durable value: "You pay for intelligence twice. Once with money. Once with the knowledge you hand over to make the model any good... It leaks. Trace by trace. Correction by correction. Eval by eval" [9]. She insists this is not byproduct but the core asset: "The corrections are the asset, the value... That is not exhaust. That is the company. It is the thing a competitor could never buy" [9]. This reframes competitive advantage away from model choice: "the real question is not which model is smartest this month. Models change every few weeks" [9]. On a related but separate worry about AI and headcount, she argues engineers remain the bottleneck rather than becoming obsolete: "AI didn't reduce the need to build. It increased the scale of what everyone can build... the bottleneck just moves... And the moment a company wants to move faster than that ceiling allows, it hires more engineers. Not fewer" [11].

From the stage

In interview and podcast appearances, Draisma discusses ground the written posts don't touch: her own career path from junior account manager at an e-commerce startup through acquisition by Lightspeed, its IPO, and further acquisitions across European markets, and how integrating acquired companies with different cultures and processes contrasts with the "unified commitment" of startup leadership versus the more structured hiring needs of larger organizations [13]. She also lays out forward-looking predictions not present in the posts: she expects AI to be incorporated as a decision-making partner within organizations over the next five years, and expects smaller language models to become more available and competitive within fifteen years [14].

Takeaways

• Don't treat an AI vendor invoice as a KPI on its own; without session-level detail it "punishes the wrong thing," since a third of observed spend went nowhere near shipped code [2].
• Push past "which agents are approved" and ask "can you list every AI agent running in the company right now" to get an accurate picture [1].
• Test documentation and interfaces as if an agent, not a human, is the primary reader; that's now the real audience [10].
• Route domain-expert corrections into reviewable, tested pull requests rather than backlog tickets, so their judgment doesn't die unused [7][8].
• Treat accumulated corrections, evals, and traces as the company's defensible asset, more durable than any specific model choice [9].
• Expect AI to raise the bar for output rather than shrink engineering headcount over the next 2-3 years, since competitors scale up too [11].