The Disclosure Deficit: Agentic AI, the Oracle Problem, and the Only Number That Matters
Hook
Three vendors. Forty-eight hours. One number nobody wants to say out loud.
In the span of a single news cycle, Kyndryl announced agentic AI inside Toyota Finance's IT operations. DXC published a set of efficiency claims β 500 use cases, a 95% reduction in mean time to repair, diagnostic accuracy above 92%. HCLTech committed 25,000 engineers to training on Claude. Read the press releases and you conclude that enterprise IT has crossed a threshold. Read the footnotes and you conclude something else entirely. An audit cited in the same coverage β 150 deployments across 119 organizations β found that only 22 disclosed any financial impact. That is 14.7%. Eighty-five percent of enterprise AI deployments, by that accounting, decline to publish the value they claim to create.
I do not trust the silence, I audit the code. And when an industry that sells measurement refuses to be measured, the absence is the finding. This is not a story about whether agentic AI works. It is a story about who is allowed to decide that it works, and on what evidence. For anyone who has spent a decade arguing that verification is the product, the pattern here is uncomfortably familiar β and it is the same pattern that decides whether on-chain systems inherit the trust they promise, or merely relocate it.
Context
The event, as reported, is a cluster of three competitive announcements compressed into a 48-hour window. Kyndryl is the most legible: a named customer, Toyota Finance, and a described technical stack β graph-based orchestration, integration with the Kyndryl Bridge ITSM and observability platform, and grounding in customer-specific rules, knowledge, IT environments, and runbooks. What Kyndryl did not provide is a single outcome metric. No accuracy figure. No cost reduction. No incident count. The value proposition is carried entirely by the credibility of the customer's name.
DXC took the opposite route. No named customer, but a dense set of numbers: 500 use cases, 95% MTTR reduction, 92%+ diagnostic accuracy. Numbers without an anchor are auditable only in theory. You cannot call the reference. You cannot inspect the baseline. You can only decide whether to believe the vendor.
HCLTech took a third route, and it is the most revealing. No outcome data at all. Instead, a scale metric β 25,000 engineers trained on Anthropic's Claude β presented as evidence of seriousness. This is the classic substitution of a process metric for a result metric. Training volume measures input. It says nothing about whether the resulting agents do anything correctly. HCLTech also did what neither of its peers would: it named its model dependency. Claude. The other two left their model layer dark.
The industry context matters more than the announcements. Global IT services is roughly a $1.5 trillion market built on a human pyramid. At the base sit large numbers of junior engineers performing standardizable work β L1 and L2 support, change execution, monitoring and alert triage. That base is precisely what agentic AI targets first. When DXC advertises 92% diagnostic accuracy, it is advertising a capability aimed directly at the bottom of its own pyramid. The three announcements are not independent. They are a narrative arms race inside a sector whose core asset is the labor that the technology threatens to replace.
A reader whose world is blockchains might ask why this matters. It matters because the same failure is being rehearsed here that the crypto industry has spent a decade trying to solve β the gap between a claim and its proof. Enterprise AI is now generating machine decisions inside production systems with no verifiable trail, no independent audit, and no shared standard for what counts as evidence. We already have a name for systems that assert their own correctness without external verification. We call them trusted, and we have learned, expensively, what trusted means when it is wrong.
Core
The Architecture Is Not the Innovation
The first thing an auditor does is strip the marketing language down to a schema. Kyndryl's description reduces cleanly: graph-based orchestration, retrieval over enterprise knowledge, tool invocation, and change-risk analysis. That is an Agent plus RAG plus tool-calling stack. Graph orchestration maps to the LangGraph family of frameworks. This is not an architectural novelty. It is a combination-level application of a paradigm that was validated in 2023 and 2024. The engineering is real; the invention is not.
The genuinely difficult question is not whether an agent can perform change-risk analysis. Rule engines and classical machine learning have done a version of this for years. The incremental value of an LLM here is narrower than the press releases imply: natural-language comprehension of runbooks and semantic correlation across systems. That is a meaningful gain. It is not a new capability class.
The unresolved question β the one the coverage does not answer β is whether the agent reaches production-grade reliability in the highest-risk surface of IT operations. Change management is where errors become outages. A misapplied change can take down a payment system. Yet neither Kyndryl nor its peers published a misjudgment rate, a human fallback mechanism, or a responsibility assignment for when the agent recommends the wrong change. In a domain where the cost of a false negative is a production incident, the omission of error statistics is not a detail. It is the whole risk profile.
The most telling signal is what both Kyndryl and HCLTech chose to emphasize: governability. Kyndryl says explainable and auditable. HCLTech says governed agentic AI. When two competitors independently lead with governance as their differentiator, they are confessing that the core technical bottleneck is the structural conflict between LLM opacity and enterprise compliance. Explainability is being sold as a feature. It is actually an unsolved problem wearing a feature's clothes. Graph orchestration provides process traceability β you can see the sequence of steps. It does not provide reasoning explainability β you cannot see why the model weighted one risk above another. The vendors are packaging the former and implying the latter.
The Model Layer Is the Arms Dealer
Kyndryl never disclosed its underlying model. HCLTech bound itself publicly to Claude. This asymmetry exposes something structural about the IT services layer: it is hollow at the model level. These firms are integrators, not model builders. Their technical autonomy has a ceiling, and the ceiling is set by suppliers they do not control.
This is the arms-dealer dynamic that anyone who has watched infrastructure markets will recognize. Anthropic sells to HCLTech and, by extension, to HCLTech's competitors. It sells the same capability to every side of the competition. The service providers cannot build a moat at the model layer because they do not own the layer. Every deployment they win strengthens their supplier's bargaining position. The value migrates upward, away from the labor-intensive service layer and toward the model and platform layer.
I learned this pattern the hard way, from the other side. In 2017 I spent three months manually auditing the initial CryptoKitties smart contracts during the ICO boom. My applied mathematics background let me find a critical integer overflow in the breeding logic that others missed. I reported it privately to the core developers rather than publicizing it, because network stability mattered more than my recognition. That overflow could have detonated during the December 2017 traffic peak. What I took from it was not a trophy. It was a method: establish the foundational truth before you discuss the sentiment. When I later wrote about the DeFi oracle delays in 2020, I embedded the mathematical proof rather than the narrative, because the proof was the only part that could not be spun. The same discipline applies here. Strip the vendor language and you find that the integrators have no model of their own β and a system that depends on a supplier it cannot audit is a system with a single point of failure. Fragility hides in the single point of failure.
The governance claims make this worse, not better. An integrator that promises explainability while outsourcing the reasoning to a third-party model is promising something it cannot deliver at the layer where it matters. It can explain its orchestration. It cannot explain the model. And the model is where the judgment lives.
The Billable-Hour Paradox
Here is the commercial core that the coverage gestures at but does not finish. Traditional IT services are billed by headcount and by hour. Agentic AI, by its own value proposition, replaces human hours. Every hour an agent removes is an hour of revenue removed. This is not a subtle conflict. It is a structural contradiction at the center of the business model, and it explains the disclosure deficit better than any claim of missing evidence infrastructure.
The reason 85% of deployments do not publish financial impact is not that they lack the tools to measure. It is that they lack the motive to disclose. To prove a 95% efficiency gain is to prove that the customer no longer needs as many billable engineers. A rational service provider, facing that arithmetic, chooses ambiguity. Ambiguity preserves the option. Disclosure forecloses it.
Watch how each firm navigated the contradiction. Kyndryl led with a named customer and withheld numbers β brand endorsement over proof, borrowing Toyota's manufacturing credibility to substitute for evidence it would rather not produce. DXC led with numbers and withheld the customer β data as market signal, but data that cannot be anchored to anyone who can confirm or deny it. HCLTech led with scale and withheld both β 25,000 trained engineers as a proxy for impact. Three different credibility proxies, one shared absence. None of them published price.
That last omission is the deepest. In a seat-based or hour-based model, the pricing of agentic AI is genuinely unresolved. Do you charge for the hours the agent replaces, which is self-cannibalizing? Do you charge for outcomes, which requires proving outcomes you have no incentive to prove? Do you charge a platform subscription, which decouples price from value? This is not a marketing gap. It is an unsolved business-model question, and it is the real reason the numbers can wait β the numbers would force a decision the incumbents are not ready to make.
There is a specific trap inside DXC's disclosure that deserves naming. If the 95% MTTR reduction is real, DXC's own managed-services revenue model is disrupted β the efficiency it advertises is the efficiency that erodes its billing base. If it is not real, the claim is potentially misleading. Both directions are bad for DXC. This is why it disclosed numbers (to buy market voice from a position of relative weakness) but not a customer (to avoid a verifiable anchor). The asymmetry is the tell.
Change Management Is the Highest-Risk Surface
Most AI-safety discussion lives in abstract harm. This deployment lives in concrete operational risk, and the coverage treats it as routine. It is not routine. Change management is the highest-consequence surface in IT operations because changes are what break production. An agent that misjudges a change risk does not produce an awkward sentence. It produces an outage.
This is where the blockchain analogy sharpens rather than decorates. A production change is structurally similar to a smart contract upgrade. Both are irreversible in practice. Both execute against live state. Both carry consequences that propagate far beyond the point of the decision. In the on-chain world we learned β through The DAO, through every bridge exploit, through every upgrade that drained a pool β that the moment of highest risk is not steady state. It is the transition. Change management is the transition layer of enterprise IT, and agentic AI is being deployed into it without published failure rates.
I spent the 2020 DeFi Summer building a Python framework to model price-manipulation risk in early Compound Finance. I found that the oracle delay in specific liquidity pools could be exploited by well-funded actors during high volatility. I published the data-backed warning to a community of 5,000 and told them to hedge. Most ignored the math. Those who followed avoided significant losses when the wETH oracle glitch hit weeks later. The lesson was not that I was right. The lesson was that the failure mode was knowable in advance β and that the people who ignored it did so because the warning was inconvenient, not because the evidence was weak. The same dynamic governs agentic change management. The failure mode is knowable. Nobody is publishing it. And the people deploying the agents are the people with the least incentive to publish the ways the agents fail.
The Oracle Problem Arrives in IT Operations
Truth is an oracle, not a price feed. I have written that line for years in the context of DeFi, where the manipulation of a single data source can cascade into liquidations across a protocol. The meaning generalizes. An oracle is any external input that a system trusts to be correct without being able to verify it internally. DeFi oracles fail because they import a fact the chain cannot check. Agentic IT operations fail for the same reason: the agent imports a judgment β this change is safe β that the surrounding system cannot independently verify.
A named customer like Toyota Finance is a brand oracle. It imports credibility the vendor cannot generate itself. A number like 92% diagnostic accuracy is a data oracle. It imports a measurement the market cannot recompute. In both cases, the trust is placed in a source that is not verifiable at the point of use. This is the oracle problem in a suit. It is not exotic. It is the oldest failure pattern in systems that outsource their truth.
This matters for the crypto reader because it reframes a debate we have been having internally. We have spent years arguing about whether decentralized oracles are better than centralized ones. The enterprise case shows that the question is not centralization versus decentralization. The question is verifiability at the point of decision. A centralized oracle that publishes its methodology and its data can be audited. A decentralized oracle that hides its inputs is just a committee with a token. The enterprise vendors here are running the committee-without-a-token version: they ask for trust, they decline to provide verification, and they wrap the whole thing in the language of governance.
What Blockchain Actually Offers
It would be easy, and lazy, to conclude that blockchains solve this. They do not, automatically. But they offer three primitives that the enterprise agent stack conspicuously lacks, and it is worth being precise about which is which.
The first is an immutable audit trail. An on-chain record of every agent decision, every input, every human override, written append-only and independently verifiable. The enterprise stack can produce logs, but logs are mutable and controlled by the party being audited. An immutable trail changes the power relationship: the auditor no longer depends on the audited party's goodwill.
The second is verifiable computation. Trusted execution environments and zero-knowledge proofs let a system prove that a computation was performed correctly without revealing the inputs. This is the primitive that directly attacks the explainability gap. It does not require the model to explain itself. It requires the model to prove it ran the committed process on the committed inputs. For a change-risk agent, that is enough to establish accountability without demanding interpretability the model cannot provide.
The third is provenance. In 2021 I founded a curated community around the philosophical implications of on-chain provenance rather than speculative flipping. I spent weeks analyzing the transaction history of early Art Blocks projects, documenting how immutable ledger entries create a new form of artistic history. I wrote a series arguing that the value lies in the verifiable, tamper-proof narrative of creation, not the image. We do not buy pixels, we buy history. That principle transfers directly here. The value of an agentic IT operation is not the agent's output. It is the verifiable, tamper-proof record of how that output was reached and who authorized it. An agent without provenance is a black box with a signature line. An agent with provenance is an accountable participant.
The hard part β and the reason I am not writing a triumphant essay about on-chain AI β is that none of these primitives are being used in the deployments under discussion. Not because they are impossible. Because they are inconvenient. Provenance is expensive. Verifiable computation has overhead. An immutable trail is discoverable, and discoverable trails are the last thing a vendor with a disclosure deficit wants. The technology to close this gap exists today. The incentive to close it does not. And that is the more honest finding.
The Disappearing Junior and the Provenance of Skill
The labor dimension deserves its own audit, because the coverage skips it entirely. HCLTech training 25,000 engineers on Claude admits two readings, and only one of them is comfortable. The optimistic reading is reskilling β engineers move up to supervise agents. The pessimistic reading is that a single engineer can supervise more agents, so the total number of engineers required falls. The article does not touch this. That omission is significant, because the second reading is the one consistent with the efficiency numbers DXC is advertising.
The deeper risk is structural, and it is the one nobody is pricing. IT operations is a pyramid with a broad junior base. That base is where senior engineers are made. Junior engineers learn by doing the standardizable work β L1 and L2 support, alert triage, change execution. If agents absorb that work, the industry removes the entry point through which it has always produced its senior talent. A field that cannot train juniors cannot produce seniors. This is a capability cliff that appears slowly and then all at once, and it has no owner.
I watched a version of this in the 2022 bear market, when I advised my community to exit 80% of volatile altcoins and hold stablecoins. I published a stark, emotionless report on the failure of lending protocols like Celsius and used game theory to explain the inevitable collapse. Many left. The core stayed, because they valued survival over speculative gains. What that period taught me is that structural truths are most valuable precisely when they are most unwelcome. The disappearing junior is such a truth. The industry is optimizing the visible quarter and mortgaging the invisible decade, and the audit that would reveal it is the audit nobody commissions.
Contrarian
The comfortable interpretation of this episode is that the disclosure deficit is a failure of evidence infrastructure β that vendors simply have not built the tools to measure what they claim, and will do so once the tooling matures. I do not accept that. The disclosure deficit is not an infrastructure gap. It is a competitive equilibrium.
Consider the game. If every vendor withholds outcomes, no vendor bears the cost of an unfavorable number. The moment one vendor discloses a real, unremarkable result β say, a 30% efficiency gain instead of a 95% one β it hands its competitors a comparison they can exploit. In a market where the alternative to a spectacular number is a mediocre one, disclosure is a dominated strategy. Collective silence is the stable outcome. The 85% who do not publish are not failing to measure. They are winning a coordination game whose payoff is ambiguity.
This is the contrarian point that most analyses miss, and it carries a warning for my own side of the industry. The crypto world has spent a decade arguing that transparency solves trust. But on-chain transparency is not a cure for the disclosure problem; it is a different surface for the same disease. A transaction hash proves that something was written. It does not prove that the something means what its author says it means. A wallet can broadcast a worthless token and the chain will faithfully record it forever. Immutability guarantees the persistence of a claim, not the veracity of one. The chain is an oracle for what happened, not for what it is worth. We are not immune to the disclosure deficit. We are running a faster, more permanent version of it, and calling the permanence proof.
So the honest position is uncomfortable for both camps. The enterprise vendors are managing a rational silence. The on-chain evangelists are managing a rational overclaim. Both are dodging the same question: who verifies the verifier, and what does verification actually establish?
Takeaway
Code is law, but audits are conscience. The law of this market is now written β agents are entering production systems, vendors are competing on narrative, and the numbers can wait because the numbers are inconvenient. The conscience has not been written yet. It will be the first firm that publishes a real, mediocre, verifiable result and survives it β the first vendor willing to say 30% instead of 95% and stake its credibility on being checkable rather than being loud. Proof precedes value; provenance is the only art. The question for the next 48 hours, and the next 48 months, is not which vendor has the better agent. It is which vendor will let you audit it. Until one does, the only number that matters remains 14.7% β the fraction of this industry willing to be measured. Alpha is quiet, noise is just noise.
