Agents Don’t Create Value. Authority Does

The popular story says this is a temporary adoption gap. People will learn. Prompts will improve. Workflows will mature. Returns will appear. That belief sounds reasonable, which is why it persists. What would have to be true for this outcome to keep repeating. It would have to be true that the bottleneck is not intelligence. It is legitimacy. It would have to be true that AI has collapsed the cost of knowing while leaving the cost of deciding intact. It would have to be true that most organizations still manufacture permission socially, through meetings, rituals, and reputational risk sharing, because legitimacy is not portable. And it would have to be true that this is not a culture issue. It is an architecture issue. Insight is instant now. Permission is still slow.

Wall Street is not overreacting. It is pricing controllability. The market is doing something that looks emotional and is actually rational. It is repricing the application layer because software creation and integration costs are falling, and because the profits of “AI software” were often built on scarcity that no longer holds. When the unit cost of code falls, what gets priced is not cleverness. It is control. In early February 2026, Reuters described a steep decline in software and services stocks tied to fears that fast advancing AI could disrupt traditional software revenues, with market value losses measured in the hundreds of billions and then roughly a trillion as the selloff extended. Financial Times reported a JPMorgan index tracking U.S. software stocks down materially year to date during the same period. Those numbers matter less as headlines and more as signals. Markets discount uncertainty. They pay for systems that can hold under pressure. They do not price your internal dashboards. They price whether your enterprise can still bind itself to action faster than competitors, without breaking the rules that keep it alive. This is why boards have become more skeptical even as AI spending forecasts rise. Gartner’s January 2026 forecast put worldwide AI spending in the trillions for 2026. At the same time, PwC’s 2026 Global CEO Survey said only a small minority of CEOs reported both cost and revenue benefits so far, while most reported neither. EY’s research on responsible AI governance found financial losses from AI related risks were widespread, with a meaningful share of organizations reporting losses above a million dollars. Capital is not going away. Proof standards are rising. That is the context leaders keep missing. AI is not only a capability wave. It is a proof wave. It makes claiming easy and it makes earning hard. In that environment, a strategy that sounds like “we will use AI to improve decisions” reads like theater unless it specifies who can act, under what constraints, with what evidence, and with what audit trail.

Most enterprises bought seeing and called it control

For twenty years, enterprise technology investment rewarded a particular posture. Instrument more. Centralize more. Visualize more. Add another dashboard layer, another data product, another “single pane of glass,” another set of alerts and scores. The enterprise became better at noticing. It did not become better at intervening. In the one degree world, that becomes a trap. Connectivity and access make inference cheap. They also expand attack surface and error surface. So the enterprise responds rationally by tightening gates. Cybersecurity, safety, quality, finance, contracts, and credit policy all impose enforceable cost when action is wrong. Those functions are not enemies of speed. They are the operating constitution. When a decision touches them and legitimacy cannot be proven quickly, permission must be negotiated. Negotiation produces delay. Delay produces value leakage. Then leaders misread the leakage as “we need more insight.” That loop can run forever because it feels responsible. It has all the right words. Governance. Alignment. Diligence. Risk control. Those words are not lies. They become a problem only when they become permanent architecture. When legitimacy is not portable, the meeting becomes the operating system. Trust that scales is not explainability. It is auditability. This is why so many AI programs produce impressive demos and weak earnings. They accelerate artifacts. They do not accelerate commitments. They add speed to language and they create more decision prompts, then the same permission staircase turns each prompt into coordination. In the enterprise, that looks like productivity. In the ledger, it looks like overhead. If you want a clean diagnostic that does not depend on slogans, watch what happens after an AI system makes a recommendation that matters. Does someone with authority accept responsibility quickly. Does the organization translate it into a bounded intervention quickly. Does the intervention reach the edge where work happens quickly. Does the system learn whether it worked. If those times are long, it does not matter how brilliant the model is. You built a company optimized for seeing, not shaping.

The word “agent” is being used to sell comfort. Real agents raise the accountability bar

A chatbot answers. A real agent acts within bounded authority. That is not semantics. Definitions become architecture. An agent must be able to shape an outcome. Otherwise it is only a tool. The reason this matters is not philosophical. It is financial and operational. When leaders call something an agent, people begin to outsource responsibility. When the intervention fails, humans still own the bill. They owned it the whole time. The label only blurred it.

The enterprise does not need AI to “replace governance.” It needs governance that can run at the speed AI now forces into the system. Governance is not a brake. It is the mechanism that makes action legitimate. The missing artifact is what makes legitimacy portable across trust boundaries. Call it a decision packet. It is not a dashboard. It is not a chat transcript. It is a structured, replayable record that carries the minimum information required to commit responsibly. Evidence and provenance. Identity and integrity checks. Uncertainty bounds and failure conditions. The causal rung being used, so the system can justify intervention rather than association. Constraints and invariants that cannot be violated. Policy checks that mattered. The authority scope being invoked. The proposed action. Monitoring triggers. Rollback criteria. The accountable human owner. When legitimacy travels with that packet, relitigation falls. When it does not, legitimacy must be rebuilt socially, and that is why meetings multiply. If it cannot shape an outcome, it is not an agent. This is where many leaders get trapped by fluent explanations. Large language models can produce narratives that feel like clarity. Narrative can help humans in low consequence contexts where reversibility is cheap. But in complex, coupled operations, narrative is not legitimacy. A fluent explanation does not survive confrontation. It cannot be replayed. It cannot be audited. It cannot carry authority across functions whose job is to impose cost when mistakes become catastrophic. If you want to know whether a vendor is selling you language or selling you control, ask a question that sounds almost rude because it is so operational. When your system recommends an intervention, where exactly does it stop. At recommendation, or at action. If it acts, what invariants bound its authority, what evidence thresholds trigger action, what audit trail is written automatically, and who can override in seconds when context is wrong. Ask another question that is harder to dodge because it cuts through the slideware. When your system makes a causal claim, what would change its mind. What evidence would force the mechanism diagram to update. What evidence would leave it intact. If the answer is “the model will learn over time,” what stops it from learning a story that sounds right and fails under intervention. Those paragraphs are not rhetorical. They separate help from theater in minutes.

The alphabet soup is not what is breaking you. The burdened interface is

Executives keep pointing at the stack. ERP. MES. CMMS. LIMS. CRM. PLM. And now another layer. Connected worker apps. Copilots. Context graphs. Agent platforms. The instinct is to rank the tools and decide which one survives. That is the wrong frame. Most of the stack is not vulnerable because it is old. It is vulnerable because it asks the enterprise to pay the subscription in humans.

In a recent conversation, one operator ranked connected worker platforms as among the most exposed because they “present the news” but do not convert that news into outcomes without forcing the organization to change itself. That is not a condemnation of dashboards. It is a statement about where value leaks. If your product increases awareness and then routes action into the same fragmented permission chain, you did not remove work. You moved it. The enterprise still has to interpret, infer, seek approval, and negotiate. You just made the signals arrive earlier, which means you can now escalate earlier, and call it progress. This is why boards hesitate on major ERP and MES spend unless there is an outcome strategy tied to how AI changes decision and permission. They sense that a new front end menu, no matter how modern, will not restore control. They are right. The next wave does not replace systems of record first. It replaces the human middleware layer that lives between them. Menu driven interfaces lose because they externalize integration cost onto workers. Taskless surfaces win because they preassemble transactions, check constraints, and propose the next legitimate action without forcing a person to hunt through seventeen screens. That does not mean the enterprise becomes a pile of disposable apps. It means the edge becomes disposable and the core becomes more explicit. Generate. Operate. Discard. Repeat can be a winning option only if the durable core holds intent, semantics, decision rights, evidence standards, and accountability. Disposable software without that core is faster chaos. It is institutional risk that arrives at the speed of code generation. The graph is not the point. The point is legitimate action. This is also where PLM and CAD and CAM finally get understood in the correct frame. Product Lifecycle Management is not just a repository. It is the discipline of product definition, revision control, change governance, and the digital thread that ties requirements to designs to manufacturing to service. CAD is how the geometry is authored. CAM is how manufacturing turns designs into toolpaths and process plans. That stack is not merely “software.” It is how the firm decides what the product is, what version is real, and what changes are allowed. AI will touch that world, but not in the way most leaders assume. It will not make physics negotiable. It will not remove the need for disciplined change control. It will not eliminate the reasons quality and safety impose enforceable cost on bad releases. What it will do is collapse the time between question and next test. It will draft variants, generate documentation, summarize change impact, compare requirements to design intent, and find inconsistencies humans miss. It will also generate more plausible changes, which increases the volume of decision prompts. That is where many PLM programs will stall if leaders treat AI as a feature and not as a permission redesign. In design and manufacturing, AI’s best work will look like compression. Shorter loops from concept to simulation to prototype to process plan. Faster detection of conflicts across BOM,

routing, and spec. Faster translation between engineering language and manufacturing language. But the value will still only appear when someone can commit. Release the design. Approve the deviation. Change the setpoint. Qualify the supplier. Block the shipment. Those commitments are the enterprise. If AI increases proposal volume without increasing commit capacity, it will not create advantage. It will create noise with better grammar.

What AI can do, and what it cannot do, is already visible if you stop pretending the bottleneck is adoption

Most enterprises are using AI as a language layer, not a control layer. They draft. Summarize. Rewrite. Search. Translate. Generate code snippets. They get faster at producing artifacts. That can remove real friction, and it is worth doing. But it does not change the speed of the enterprise where it matters. It does not create legitimate action. A control layer is different. A control layer holds bounded authority. It can act inside explicit thresholds and audit trails. It can execute, log, verify, and escalate. It can carry decision packets. It can be contested. It can be replayed. It can be rolled back. It converts inference into legitimate action under guardrails. Leaders often assume that if they automate enough tasks, autonomy will emerge. The opposite is more common. Autonomy fails because accountability remains human and legitimacy remains social. When a system cannot prove it will act safely, it will negotiate, delay, and normalize. It will become a recommendation engine with a modern label. This is why the most useful boundary condition is also the least comforting. AI will not grant decision rights. Leaders must design them and delegate them. AI will not supply purpose or values. Leaders choose the ends. AI will not remove accountability. Humans remain responsible for outcomes. AI will not create truth without evidence. It will amplify bad inputs and bad incentives. AI will not solve culture. It will expose fragmentation, misalignment, and fear. That last point is not a moral judgment. It is a mechanical one. Fragmentation is what forces the meeting to exist. When data is scattered, semantics are inconsistent, workflows are not instrumented for proof, and decision rights are unclear, the organization cannot let action move down. It will centralize. It will ask for alignment. It will convert accountability into committee, because committee is where reputational risk gets shared. So the question is not whether your leaders are brave enough to “embrace AI.” The question is whether they are willing to do the work that makes action legitimate at speed. Compile permission. Standardize the decision packet. Decide which repeatable operational decisions should run with bounded autonomy, and which must remain escalatory. Record outcomes. Defend counterfactuals. Update permissions with evidence, or you will keep paying the meeting tax forever.

The counterexample is real. It does not save the old model

There are decisions where permission should not move down far, even with excellent evidence. Irreversible actions with catastrophic downside. Rare actions where the organization has limited experience. Actions that cross legal or ethical boundaries with high exposure. Actions in adversarial environments where signals cannot be trusted. In those contexts, centralization is not pathology. It is a boundary condition. Regulated industries will also move more slowly on true autonomy, not because leaders are timid, but because auditability and rollback must be strong enough to carry legitimacy across trust boundaries. If regulation clamps down hard, or if early autonomous failures create political backlash, the future will be a slower version of the same idea. Generation will be allowed. Operation will be tightly bounded. Evidence standards will rise. That does not contradict the thesis. It sharpens it. Even in that world, the scarce resource is not code. It is legitimacy. Here is a prediction that will be embarrassing if wrong, because it is concrete enough to audit. Within the next 18 months, the majority of “agent” deployments inside large enterprises will stall not because models fail, but because permission architecture cannot name what the agent is allowed to do, under what invariants, with what evidence, and with what rollback authority. Those stalled systems will still be celebrated as pilots and capability building. The value will not show up where the business case said it would. The reason will be called adoption. The reason will be permission. There is another prediction hiding behind the economics of disposable software. Within three years, many large enterprises will run more disposable edge workflows than durable applications. If that fails, it will fail because risk proves too high, or because regulation makes disposable edge politically impossible. But if it holds, the winners will not be the ones with the best generation models. They will be the ones that can discard edge workflows without losing coherence, because semantics, permission, and evidence are treated as first class operating assets. The market is already acting like it believes this. Not because markets are wise, but because markets are unforgiving. They do not wait for your committees. They do not pay premiums for your vocabulary. They pay for controllability. The last trap is the simplest. Leaders keep treating AI as a tool purchase. It is not. It is a cost collapse that turns insight into an ambient utility, then charges rent in expectations. When insight becomes assumed, every other bottleneck becomes competitively unacceptable. The firms that still govern like insight is scarce will feel like a factory trying to run a modern line by lantern light. The question is not whether you can see. The question is whether you can authorize action fast enough to stay legitimate. References:

This article is anchored in the supplied operational and market material that frames permission latency as the binding constraint, including “AI Makes Insight Instant Now. Then Why is Performance Still Paying the Bill” , “Your Data Got Cheaper. Permission Got More Expensive” , alongside the BOFA investor deck’s decision packet, transcript, and governance constructs , the 2026 predictions draft on menu burden and taskless surfaces , and the transcript’s discussion of software vulnerability and the limits of connected worker layers. Time sensitive market and spend claims were verified against Reuters reporting on the February 2026 software selloff , Financial Times reporting on the JPMorgan software index drawdown , Gartner’s January 2026 AI spending forecast , PwC’s 2026 Global CEO Survey on realized AI benefits , and Reuters and EY’s October 2025 findings on AI risk related losses . The mechanism claims rest on the durable canon of coordination and legitimacy in firms, including Coase’s “The Nature of the Firm” in 1937, Simon’s bounded rationality beginning in 1947, Williamson’s transaction cost economics, Jensen and Meckling’s agency costs in 1976, Deming’s systems critique in 1986, and Pearl’s causal hierarchy from 2000 through 2018, plus governance ballast from NIST’s AI Risk Management Framework 1.0 in 2023 and Zero Trust Architecture guidance, all used here to separate association from intervention, and explanation from auditable legitimacy.