The approval queue is the new delivery bottleneck
AI adoption is stalling on governance and ownership while model speed keeps improving.
GEICO's approval gates and Tenex's 5% figure for agents in production tell the same story: process decides delivery.
Approval gates decide. Approval gates decide whether agent projects land.
GEICO's $1,000 agent day produced a vetted migration plan only because humans approved every gate, and Tenex sees just 5% of enterprises running a non-engineering agent in production.
The reports are self-reported and early, yet they point the same way.
Budget for review capacity before you scale autonomy.
Measure the job. Benchmark averages hide the waits users actually feel.
MiniMax's own table shows sign score dropping 6.6 points and p95 time to first token jumping from 1.5 to 11.3 seconds when concurrency doubles, so real users wait eight times longer.
It's a vendor pitch, but the discipline applies: measure end-to-end job cost and keep a rollback plan before swapping.
Own the patching. The expensive part of security is owning the fix.
Carrefour's fridges hit 50°C and caught fire after vendors shipped patches nobody applied.
The IBM panel reports 44% of attacks now exploiting agent identities, so treat agents as insider risks and put kill switches ahead of more detection tools.
My take
I'd rather spend budget on review queues and ownership charts than on the next model.
The day's strongest evidence keeps landing on the same wall: agents fail where process, permissions, and accountability are fuzzy.
I'm not convinced custom harnesses or output pay fix that; they just move the friction.
The durable moves are making legacy code queryable, standardizing how a team works, and counting approval time as real capacity.
That's where the durable win is.
The videos behind this brief
We watched these so you don't have to.
AI war rooms can crack legacy code, GEICO's $1,000 day showsAI Engineer
Paying engineers for output exposes AI's real bottleneck: processLatent Space
The next AI bottleneck is governance, not individual developer speedGOTO Conferences
Why a faster model can make your users wait 8x longerAI Engineer
AI finds the bugs; the patching gap still decides the damage.IBM TechnologyVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.