Receipts, But No Ruler

Four sectors just published before-and-after agent outcomes — and the common pattern is time collapse, not a headcount cliff

Three weeks ago I told you the human-workload receipt I wanted had not surfaced. This week I need to update that on air, because the evidence got a lot better and a lot more specific.

BNY Mellon's Eliza platform is now reported at more than 20,000 AI agents across 125-plus use cases. In the cited case studies, legal contract review dropped from four hours to one — a 75% reduction — and financial planning work fell by 60%, with agents cross-referencing databases, validating regulations, and communicating through Microsoft Teams.

JPMorgan's agentic blueprint reports 30% to 40% reductions in routine operational workflow time and $1.5 billion to $2 billion in annual cost savings. A Dutch financial institution's AI-driven KYC onboarding reportedly cut onboarding time by 90% while reducing compliance staff workload by 30%.

And the healthcare receipt is the one I cannot stop staring at. A Veterans Health Administration site using Nova Insights agents dropped day-of-surgery cancellations from 17% to under 2%, 30-day readmissions from 11% to near zero, and survey administration from 172 hours to 2 per cycle. That is not an incremental gain. That is a category change in what a small team can do.

Notice what these receipts do not show. They do not show a public headcount cliff. They show workflow compression, staff redeployment, avoided administrative drag, and structural advantage for organizations that moved early. That is the strongest evidence week I have had yet for the extension thesis.

The compliance sting in the tail is healthcare. Imprivata's new survey says 83% of healthcare organizations are running AI in multiple departments, and 72% have tools or agents deployed without formal IT approval. So yes, the receipts are here. But some of the receipts the industry still lacks are probably already running inside organizations that have not formally admitted they are running them.

Your agent's persistent memory is effectively authorizing actions — and the major IAM vendors still have not named that surface

Last week I told you the permission memo is inside the agent now. This week the paper behind that line got harder to ignore, not easier.

The arXiv paper on endogenous authorization laundering formally shows that persistent memory functions as an effective authorization policy. Under incremental memory updates, memory-writing models created false authority for up to 50.2% of unauthorized requests — 51.0% in finance specifically. Once that false authority existed in memory, executor agents acted on it in 98.6% of trials.

What matters here is the source of failure. No attacker is required. The paper calls it endogenous because the problem emerges from the agent's own summarization and consolidation steps. Your system can launder authority from the inside.

Now compare that to the market posture. Okta, SailPoint, and CyberArk have all done credible work treating agents as privileged identities. But in the material cited here, none has published architecture that explicitly names agent memory as an identity-bound authorization surface in response to this finding. Cohesity's September launch is about recovering memory and configuration state after failure — important, but resilience is not governance.

The closest bridge is Okta's integration with MemU's Agentic Memory Framework, and even there memory is framed as behavioral context, not as the place effective authorization can be fabricated.

So the business instruction is simple even if the engineering is not: if your agent has persistent memory, treat that memory like an access token. Scope it. Audit it. Make it revocable. Nobody has the finished product for this yet. That gap is the opportunity and the risk at the same time.

Eighty-eight percent of AI pilots fail, and there is still no benchmark for multi-vendor agent assemblies

Two Sundays ago I said the thesis flinched because the failure data was too loud to ignore. I am not walking that back. I am updating it with a second uncomfortable fact: our measurement layer is still behind the deployment layer.

IDC research reported by CIO found only 4 of 33 AI proof-of-concepts reached production — an 88% failure rate. Forrester's September line is even cleaner: companies are chasing, few are catching.

The architectural read is not random. The H1 2026 retrospective of 100 documented agentic deployments says the orchestrator-plus-specialists pattern is becoming dominant in high-complexity, high-blast-radius workflows, while single-agent systems still dominate by volume for simpler tasks. The market is sorting by consequence, not ideology.

But here is the seam problem I raised in the special edition and it is still unowned: there is no benchmark for heterogeneous multi-vendor agent pipelines. MCP-Atlas evaluates real MCP servers, but within single-provider contexts. Agent Spec checks cross-framework portability while holding the model constant. OrchestraBench studies orchestration failure modes in homogeneous setups. ZipBench Zoo compresses benchmark coverage, but it is compressing the same single-vendor evaluations.

As of September 14 — the seam I told you to watch — no lab, standards body, or open-source coalition has shipped a benchmark for systems where agents from different vendors and frameworks interoperate under shared orchestration. And that matters because the real enterprise architecture is heading toward exactly that mess.

So yes, the failure rate is real. And yes, we are still partially blind about why when the pipeline crosses vendor boundaries. That does not excuse the failures. It explains why teams keep learning them expensively.

The token cost collapse is real, and model routers are turning agent ROI from a pilot debate into a production math problem

Back on September 13 I told you the constraint calendar moved again. It moved again.

OpenAI cut the price of GPT-5.6 Luna by 80% on July 30, taking per-token fees down to roughly two hundredths of a cent per thousand tokens in the framing cited this week. Meanwhile NVIDIA-backed inference microservices reported sub-50-millisecond P95 latency for routed model calls in early August, and Replit said its router architecture cut average task latency by 30% while reducing cost per task by 25% versus static deployments.

Put that against the older economics. Multi-step agent workflows could easily burn millions of tokens and run ten to twenty dollars per task at prior rates. If you had a ten-thousand-task monthly workflow, the business case six months ago looked very different from the one in front of you now.

The counterweight is not that cost stopped mattering. It is that token consumption is exploding. Fortune warned agentic workloads could drive a 24-fold rise in token consumption by 2030, and HBR's cost-shock warning is the governance version of the same story: subsidies are giving way to explicit usage billing.

That means every ROI sheet you built with static model assignments is aging badly. The move now is not just use a cheaper model. It is route each step to the cheapest model that can do that step well enough. If Replit's numbers hold up more broadly, routing is not a nice optimization. It is table-stakes production infrastructure.

Nobody has won the agentic platform war, and Forrester's honest answer is that most enterprises are building agentlakes instead

In February I predicted the development camps would become visible and debated by mid-2026. That prediction is now confirmed. What I did not fully anticipate was the resolution: not a winner, but a tangle.

Forrester's Q3 Wave on AI Platforms explicitly told readers to recalibrate. Its landscape work on agentic development platforms says the market is expanding well beyond code generation, with orchestration and autonomy as the real differentiators. And its 2026 predictions gave the most honest enterprise word I have heard in months: agentlake.

That is not a compliment. It is an admission. Enterprises are stitching together wrapper agents from Microsoft, DIY agents from internal engineering teams, and vertical SaaS agents from domain vendors, then calling the resulting patchwork an architecture.

Mike Gualtieri's framing is the useful binary: runtime architecture versus agent sprawl. That is the decision line that matters. Not whether you are ideologically pro-wrapper or pro-DIY, but whether you have a governed runtime or a pile of autonomous initiatives making each other somebody else's problem.

The wrapper camp — Microsoft's ecosystem in particular — has the clearest public production receipts right now through Copilot Studio TEI work. That does not mean the war is over. It means the wrapper path currently asks the fewest organizational questions up front, which is why large companies keep taking it.

So Prediction One gets the stamp this week. The camps are visible. The uncomfortable answer is hybrid. And hybrid without governance is just a prettier word for sprawl.