Assem Chammah

Assem Chammah's core argument is that most enterprises are stuck somewhere in a predictable AI maturity cycle, and naming the stage is the first step to getting out of it. He describes it plainly: "Stage 1: DIY... Stage 2: Buy off the shelf" . The DIY stage produces mediocre outputs that lack adoption or break against edge cases, usually because someone was "tasked to 'do AI' alongside their usual workload" without the expertise or time to do it properly . Buying off the shelf feels safer and gets procurement involved, but the product is standardized while the business is not, so it "covers the common path and stops at every exception" . His point is not that either stage is wrong, only that they are waypoints, not destinations, and only 16% of Copilot pilots reach production, evidence he cites that off-the-shelf alone does not close the gap .

A second recurring theme is the distinction between what a company has written down and what it actually knows. On building an internal knowledge base or "second brain," he is blunt that the intentional build of org charts, SOPs, and documentation "is the written version of the company, which is about 20% of what the company knows, being generous," and that the share of that which is even up to date is "closer to 1%" . The other 80%, he says, "never got written down and is inside your employees' heads as they do day to day tasks," in the exceptions, gotchas, and thresholds nobody documents . This same logic shows up in how he tests file-reading architectures for agents, where he built a fictional 46-document drive, planted contradictions, and found that the best pattern uses a lead agent dispatching sub agents per folder, each starting with "an empty context window" and reporting back "what it found, and which files contradict the index" .

Chammah returns often to adoption and fear as the real bottleneck, not capability. He argues teams "aren't directly vocal" about fearing redundancy, but poor adoption is the tell . His prescription is specific: "Before the rollout starts, tell each person what their work becomes once the agent takes their current task. Name the actual job," because "'Upskilling' won't land" . He backs this with a Berkeley finding that 77% of employees using AI at a 200-person tech company said the tools increased their workload, using it as evidence that freed hours refill with the same work unless leadership decides in advance where that time goes . He extends the same instinct to tooling choice, insisting employees should be asked directly, since "whoever runs a process knows which three hours of their week are dead" .

On the technology itself, his consistent claim is that the model is a commodity and the surrounding system is what actually creates value. He points to Claude Code as proof, calling it "1.6% model and 98.4% harness," and concludes flatly that "the model is a commodity" while "the work is everything you build around it" . This shapes his advice on cost: he frames model selection as an ROI calculation, not a spec sheet comparison, telling leaders to measure what percentage of a workload a current model finishes end to end with no human intervention, do the same for a frontier model, and only upgrade when "the cost of the human hours you remove is greater than the extra token spend" . He has applied this literally, describing how he cut a client's AI bill from $100k to $60k a month by moving comparable workloads from Opus 4.8 to Sonnet without losing quality .

He is also firm that governance should never mean banning tools. "Never, ever, ever ban ChatGPT or Claude," he writes, because banning simply pushes usage onto personal accounts where the company loses all visibility, a pattern he calls Shadow AI . He cites the same phenomenon elsewhere, noting that 78% of employees already use AI their company never approved because "they found better tools than the one you rolled out," and that a shadow-AI breach costs roughly $670k more, concluding "banning it doesn't work" . His alternative is a center of excellence that owns the approved stack and governance while pushing accountability for efficiency gains down to individual VPs and team leads .

Chammah treats agents as living systems that require constant tending rather than one-time builds. Describing a sales agent that has generated over $6M in revenue across 182 versions in fifteen months, he states the "big takeaway" is that "the majority of the ROI from AI comes in the tweaking down the line," and that "you can't build an agent once and ship it, and you can't specify one up front either" . He applies a similar swarm methodology to non-technical problems like grant selection and recruiting, where a "Rubric Miner" interviews human judges to reconstruct the scoring logic that "used to live in their heads," on the premise that a written application is "maybe 20% of the picture" and parallel agents are needed to investigate the other 80% .

Finally, he is skeptical of over-relying on model vendors as implementation partners, and sees opportunity in unconventional corners of enterprise change like private equity carve-outs. On vendor risk, he notes OpenAI's enterprise API share falling from 50% to 25% in under two years while Anthropic rose from 12% to 32%, and warns that "hiring the model vendor as your implementer ties your operating model to their roadmap" . On carve-outs, his view is that the forced rebuild of IT,