# A Capability Demonstration Is Not an Operating Model ## Claim Showing that a system can perform a task establishes nothing about whether it can hold the job that task belongs to, because a job is a set of obligations that persist between tasks. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/From Autonomous AI Threats to Workplace Consequences Analysis by Forrester, ACM, and Reuters|From Autonomous AI Threats to Workplace Consequences]] — Robert Jacobi, diagnosing the reversed-layoff pattern Forrester and Reuters report > [!quote] > "They treated a capability demonstration like an operating model. A guy can do the task. Therefore AI can replace the job. Those are not the same" The inference fails on what a demonstration is allowed to exclude. A demo picks its input, runs once, and is graded by a person who wants it to work. A job absorbs the inputs it is given, runs continuously, and carries the parts nobody demonstrates: knowing when the task should not be done, noticing the case that does not fit, holding the relationships the work travels through, and being accountable when it goes wrong. None of those appear in a task, so none of them can fail in a demo. What makes this more than a category error is who pays for it. The reversal is not symmetric: > [!quote] > "So the company gets to reverse its AI experiment. The worker gets to absorb the discount." The organisation treats the failed substitution as information — a pilot that did not pan out, reversible at the cost of rehiring. The person who was removed does not get the interval back, and typically returns, if at all, on worse terms. So the experiment is cheap to run precisely because its cost was externalised onto the party with no say in whether it ran. ## Scope & conditions - **Holds** wherever a role's value includes judgment about *whether* to act, exception handling, or accountability — which is most roles and almost all coordination roles. - **Weakens** for jobs that genuinely are one repeated task with stable inputs and a clean success test. Those exist, and the substitution argument is sound there. - **The demo/job gap widens with input variance.** Low-variance work makes the demonstration nearly representative; high-variance work makes it nearly meaningless, and nothing in the demo reveals which regime you are in. - **Distinct from** doubting the capability. The claim concedes the demonstration entirely and rejects only the inference drawn from it, which is what makes it hard to argue with. - **Does not follow** that AI cannot replace jobs. It follows that a demonstration is not the evidence that would establish it, and that running the substitution live is a test whose failure cost lands on someone who did not choose to take it. ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — the evidence standard that separates a working demo from a staffing decision - [[3. Resources/MOCs/AI Agent Safety MOC]] — deploying on demonstrated capability alone is the same error one layer down, where the demo stands in for operating limits - [[3. Resources/Permanent Notes/A Prototype's Polish Signals Nothing About Its Process Stage]] — extends: that note says polish carries no information about readiness; this one says a successful run carries no information about durability, and both are cases of a demo's chosen conditions being read as general - [[3. Resources/Permanent Notes/A Creator Demo That Lets the Product Describe Itself Is Not Independent Evidence]] — because these share a defect in who controls the conditions; there the vendor picks the framing, here the pilot picks the task - [[3. Resources/Permanent Notes/Bounded Agents Earn Broader Authority Through Verified Runs]] — because contrasts: this is the constructive alternative, where authority expands on accumulated verified operation rather than on one demonstration - [[3. Resources/Permanent Notes/Prevented Work Is Invisible to Velocity Metrics]] — evidence-for: the obligations a demo omits are largely non-events, which is exactly the category no measurement system records ## Tests / falsifiers - For a role proposed for substitution, list what happens between its tasks — the escalations, the refusals, the relationships. If that list is empty, the substitution argument survives and this claim does not apply. - Track reversals. If organisations that piloted substitution rehired at comparable rates to those that did not, the demo/job gap is smaller than claimed. ## Maintenance - Last tended: 2026-08-28 - Next review: 2026-11-28 ## Source [[3. Resources/Readwise/Articles/From Autonomous AI Threats to Workplace Consequences Analysis by Forrester, ACM, and Reuters|From Autonomous AI Threats to Workplace Consequences]]