The Brief Script audited 150 enterprise agent deployments. Only 22 disclosed cost or financial impact. That's 14.7%. The other 85.3% are running on faith, slides, and time-saved metrics that nobody outside the room can verify.
This isn't an AI story. It's a liquidity story.
Capital has flooded into enterprise agents because the narrative is frictionless: 62% adoption, productivity gains, autonomous workflows. But when you strip the marketing and look at the order flow โ the actual verified value โ the book is thin. Dangerously thin.
I've spent the last eighteen months doing forensic analysis on institutional crypto flows. The pattern here is identical: retail and corporate FOMO chasing a narrative that institutional capital has already priced, while the underlying liquidity quietly disappears.
Let me walk you through the evidence.
The 150-deployment audit shows 71 reported "time savings" and 50 reported "productivity gains." Those are soft metrics. Self-reported. Unaudited. The same category of data that inflated NFT wash-trading volumes in 2021 โ impressive headline numbers that mask the absence of genuine institutional interest.
Only 10 deployments reported actual revenue. Only 22 disclosed cost. If 85% of your deployments can't produce a P&L line item, you don't have a market. You have a demo.
Now here's where it gets interesting.
The Gartner estimate is brutal: of thousands of vendors claiming agentic capabilities, roughly 130 are real. That's a 2-6% true-positive rate. The rest are AI assistants and RPA tools with a new label โ concept arbitrage at industrial scale.
I've audited smart contracts for oracle latency exploits. I've watched DeFi protocols claim decentralization while running on three nodes. This is the same playbook. Centralized execution wearing a decentralized costume.
The 130 real vendors have planning, tool-use, and memory loops. The thousands of others have a wrapper and a pitch deck. And the market โ desperate for exposure โ is pricing them at parity.
Look at the autonomous tier. 57 of 150 deployments operate with bounded autonomy. Only 17 run fully autonomous end-to-end. And 118 explicitly describe human approval, exception handling, or governance review.
That last number is the tell. When 79% of your "autonomous" systems require human-in-the-loop, you're not selling autonomy. You're selling supervised workflow automation with an LLM brain.
Which is fine โ if you price it that way. But you're not. You're pricing it like a general-purpose autonomous agent, and the unit economics don't support that valuation.
Consider the Harvard working paper: AI coding tools increased code output by 30% with zero increase in problems solved. Review time jumped 49%.
Read that again. The cost didn't disappear. It migrated from generation to verification. The marginal cost of producing code fell. The marginal cost of validating it rose. If this pattern holds across domains โ and I have no reason to believe it won't โ then the entire "productivity gain" thesis is structurally flawed.
This is the verification cost trap. And it's the same mechanism that killed the 2021 NFT market. The minting was cheap. The curation was impossible. The market collapsed because nobody could distinguish signal from noise without incurring prohibitive cost.
Agent deployments are approaching that same threshold.
C.H. Robinson is the counterexample everyone cites: quote coverage from 60-65% to 100%, response time from 17-20 minutes to 32 seconds. Impressive. But note the conditions โ proprietary data, clear rules, bounded scope. This isn't evidence that autonomy works. It's evidence that bounded autonomy works when the problem is narrow enough to be governed.
The market is extrapolating from a single point.
Meanwhile, Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027. Cost escalation, unclear value, weak risk controls. That's not a correction. That's a reckoning.
The Brief Script proposes a five-question audit framework: What does the agent read? What tools can it choose? What can it write or change? What requires approval? What happens on low confidence or policy breach?
Those are good questions. They're also the same questions we should have been asking about exchange proof-of-reserves in 2022 โ before FTX made them mandatory reading.
History rhymes. This isn't 2022. But the structure of the failure โ opaque liability, unverified claims, governance theater โ is identical.
So here's the positioning. The 130 real vendors will consolidate. The thousands of wrappers will die. The governance and audit layer โ the tools that can answer those five questions at scale โ will become the actual moat. Not model capability. Model capability is commoditizing in real time.
The trade isn't "buy agents." The trade is "sell verification."
Because in a market where 85% of participants can't prove their own numbers, the only sustainable business is the one that tells you which 15% can.
Watch the financial disclosure rate. If it doesn't improve from 22/150 in the next twelve months, the cancellation cycle has already started. You just can't see it in the adoption numbers yet.
That's the thing about liquidity illusions. They look like depth until you try to exit.
The Brief Script's audit is the closest thing this sector has to an order book. Right now, the bid-ask spread is widening. The volume is thinning. And the only people still trading are the ones who haven't tried to sell.