The system works. The people do not.
That was my first note, taken at 2 a.m. while I was parsing a vendor deck that claimed 'full autonomy' across a Fortune 500 procurement stack. I had asked for the tool-call manifest. I got a marketing PDF and a demo video. According to an audit of the brief script, only 130 out of thousands of vendors claiming agentic capabilities can survive a basic architectural review. The rest are RPA scripts wearing a trench coat.
That's a 2 to 6 percent ratio of truth to theater. In any other market, that would be a scandal. Here, it's a funding round.
Context: The Surge That Cannot Be Audited
Enterprise adoption of AI agents is reportedly at 62 percent. That number gets repeated in boardrooms, on earnings calls, and in pitch decks. What the number actually measures is anyone's guess. A production deployment? A pilot? A Slack bot that queries a database? The definition is elastic enough to include all three.
I've spent the last four months stress-testing the deployment architecture behind these numbers. The pattern is consistent: massive self-reported adoption, near-zero verifiable financial disclosure. Of 150 audited production deployments, only 22 disclosed any cost or financial impact. Only 10 reported revenue attributable to the agent. The rest report 'time saved' and 'productivity gains' โ soft metrics that are structurally immune to falsification.
Time saved is a claim. Revenue is a fact. The market is currently funding claims.
Core: The Architecture of Bounded Autonomy
The Brief Script audit segmented deployments by autonomy level. The distribution is instructive:
- Fully autonomous end-to-end: 17 of 150.
- Bounded autonomy: 57 of 150.
- Human-in-the-loop, explicitly described: 118 of 150.
My background is in quantitative due diligence, not product marketing. So when I read 'agentic AI,' I don't see reasoning loops. I see an orchestration layer โ LLM inference plus tool calls plus guardrails plus exception routing. This is integration engineering, not architectural breakthrough. The production-grade agent of 2026 is a workflow engine with a probabilistic brain and a lawyer on retainer.
What the Deployments Actually Do
The C.H. Robinson case is the most cited success. Quote coverage went from 60โ65% to 100%. Response time dropped from 17โ20 minutes to roughly 32 seconds. That is genuinely impressive. It is also deeply bounded: proprietary pricing data, rule-governed quote generation, a narrow domain with clear success criteria. The agent reads structured data, calls a pricing function, and writes a quote. This is not general intelligence. This is a very fast clerk that never sleeps.
The success stories share a shape: highly structured inputs, proprietary data, verifiable output. The moment you strip away any of those three, the productivity claim collapses into a demo.
The Verification Tax
Harvard's working paper on coding agents should be printed and taped to every VC's monitor. Code volume increased 30 percent. Problems solved: zero growth. Review time increased 49 percent. Read that sequence again.
The agent generated more code. Humans had to inspect all of it. Net new problem-solving capacity: flat. What actually happened is that the cost of production shifted from generation to verification. In accounting terms, you didn't reduce the cost function; you moved the expense line and added a new one. The engineers didn't disappear. They became auditors of machine output, and the audit queue grew faster than the generation queue.
I have seen this pattern before. In 2020, I simulated Uniswap v2 liquidity pool dynamics and found that the constant product formula created asymmetric risk during volatility events. The math was elegant. The retail LP outcomes were not. The same elegance trap applies here. The demo is coherent. The unit economics bankrupt.
The Orchestration Moat That Isn't There Yet
Suppliers jockey for position on model quality. That fight is already commoditized. The real engineering problem is in the governance layer: who defines the boundary, who routes the exception, who owns the audit log. The audit's five-question framework โ what does the agent read, what tools can it call, what can it write, what requires approval, what happens on low confidence โ is the closest thing to an industry standard I've seen. It is also not a product. It is a checklist. That gap between checklist and product is where the next twelve months of venture capital will be spent, and mostly wasted.
Contrarian: What the Bulls Are Right About
The bulls will tell you the 40 percent cancellation forecast is noise, that survivors capture the entire market, and that early adoption always looks messy. They are partly correct.
The 'bounded autonomy' majority โ 57 of 150 โ is not a failure of ambition. It's the correct engineering posture. If you are running agents against production systems with write access, you constrain the blast radius. That is not timidity. That is discipline. The 17 fully autonomous deployments operating without mature financial measurement are the reckless ones, not the pioneers. The bulls are right that deployment beats theory. They are wrong to conflate deployment with proof.
What the bulls consistently miss is that the verification tax is structural, not transitional. The 49 percent review-time penalty isn't a bug that better models fix. It's the cost of operating probabilistic systems in deterministic workflows. You can reduce it with better tooling, but you cannot eliminate it without eliminating the agent. The code compiles. The reality bankrupts.
Takeaway: The Accountability Ledger
I don't trust the audit; I trust the exploit. My 2017 experience with an integer overflow in a vesting contract taught me that whitepapers are hypotheses, not contracts. The same is true of agent adoption metrics. A 62 percent adoption rate with a 14.7 percent financial disclosure rate is not a market. It's a confession.
The next eighteen months will produce a measurable event: either financial disclosure rates climb, or project cancellations validate the 40 percent forecast. The market will tell us which, and the transaction โ unlike the mistake โ will be permanent. Until then, ask every agent vendor the same question: show me the P&L line where your system appears. If they can't point to it, they're selling you a spreadsheet in a costume.