Table stakes, not moats: what a right to play costs in 2026

Position, signed by Adam Witanowski, 25 August 2026. This is the lab's view on the table-stakes field and, by extension, on what Quantium's AI practice is for. It draws on two validation runs, on the standing-answer log — which is the most honest record of what the firm is actually asked — and on the win/loss patterns Engel has collected since March.

What is demonstrated. The firm is asked the same fifty questions. Which model for this. What does it cost. Retrieval or fine-tune. Is that vendor claim real. The answer rate on those fifty is the single metric that correlates with whether a sector team wins the next engagement (c-table-stakes-1). Every one of those answers is available to any competent client team with a quarter to spend; what the client is buying is not the answer but the fact that we already have it, dated, with the evidence attached. Accenture, Deloitte and QuantumBlack sell the same thing, and the client can tell the difference by asking a second question. On the capabilities themselves — gateway, evals, memory, auth — a client's own engineering team can build any of them in a quarter, and the good ones have (c-table-stakes-2).

What is hype. The proprietary platform. Every major consultancy has one, and in fourteen vendor and competitor assessments this year we found no case where the platform was the reason a client chose the firm. Also hype: 'our agents' as a differentiator, when the agent is a fixed workflow with a model at one step (sa-vendor-claim-agents). And hype in our own direction: the belief that a research lab is a moat. It is not. It is what makes the table stakes current, which is a different and more defensible claim (c-table-stakes-3).

What would have to be true for the position to be wrong. If the frontier labs' own consulting arms — two have been announced — delivered engagements at scale, the value of being right about model choice would move upstream to the vendor. If the questions stopped repeating — if the fifty became five hundred — the standing-answer model would break and a different mechanism would be needed. If clients stopped being able to tell a dated, evidenced answer from a confident one, the moat would be marketing after all. We see no evidence of the third, some of the first, and none of the second (c-table-stakes-4).

What we would do. Fund the answer rate as the practice's primary metric alongside lead time, and publish it. Stop investing in platform assets that a client team can replicate in a quarter; invest in the evidence that makes our choice of those assets right. Sell the position, not the platform: what we believe, why, and what we got wrong, in that order. Make the graveyard public inside the firm as the proof that the answers are earned. Re-run this position at the quarterly election; it is the one that decides what the lab is measured on.

Strong because the evidence is the firm's own question log and win/loss record, and because the position has already survived one red team pass arguing that platform assets do differentiate at the RFP stage. They do — at the RFP stage. The engagement is won or lost on the second question.