The B200 and Monzo make the AI bottleneck clear: evidence, platforms, output.
Governance, platform maturity, and cost per output now decide AI rollout speed.
Banks ship in weeks when evidence pipelines exist, Monzo's agent depends on its platform, and GPU output swings 50% by operator.
Proof is the bottleneck. Banks ship in weeks when compliance gets verifiable evidence: exploit tests prove CVEs dead, DR row-count checks prove restores work.
The IBM talk warns a 97% confidence score without an audit trail is confident garbage, and it's one practitioner's reasoning, so weight it accordingly.
Build the proof pipeline before arguing about the model.
Platform is the agent. Monzo's Slack agent authors 10% of PRs, but the talk's own point is that existing alerting, staging, and sandboxing did the heavy lifting.
It works only through an MCP gateway with approved tools; a sponsored harness talk makes the same bet, versioned rules, trace-based fixes, and human approval beat prompt tweaking.
Copy the governance model before 1,800 tasks a day.
Cost per usable output. The same B200 can deliver 50% more usable output depending on who operates it, and speculative decoding can roughly triple throughput on hardware you already own.
Both push cost per acceptable output on your real workload, though the 50% figure comes from a calculator, so verify it yourself.
Measure that before buying more GPUs.
My take
The model is the cheap part of the rollout now.
The expensive work is evidence, platform maturity, and cost per acceptable output, the unglamorous stuff you already own.
I've led engineering orgs long enough to know the review cycle becomes the project, so budget for it explicitly.
But don't turn that into a build-your-own-platform mandate.
Monzo's agent only worked because its platform was already there; the rest of us should buy a thin gateway, cap scope, and measure before scaling.
Keep the model replaceable and the proof trail fixed.
The videos behind this brief
We watched these so you don't have to.
Compliance velocity in banks comes from showing work, not better AI.AI Engineer
Confidence scores are not proof: decision AI needs an audit trailIBM Technology
Monzo's Slack Agent Writes 10% of PRs. The Platform Did That.AI Engineer
Stop tweaking prompts: engineer the harness around your production agentsCole Medin
Stop Buying GPU Hours: The Operator Decides Your Real AI CostAI Engineer
Speculative decoding is the cheapest speed-up you haven't measuredHugging FaceVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.