What the Discourse Misses About AI

An institution can adopt an impressive AI system while leaving the people it serves with an unimpressive process. A form may be handled faster but remain difficult to correct. A worker may produce more while spending extra time checking suggestions. These possibilities deserve attention alongside the larger debate about capability and risk. My argument is not that frontier questions are irrelevant. It is that institutional progress needs its own evidence: what work was completed, what errors remained and who carried the burden of repairing them.

Brynjolfsson, Li and Raymond’s 2025 study offers one concrete result. Using a staggered rollout and comparisons over time among 5,172 customer-support agents at one firm, it estimates a 15 percent average increase in issues resolved per hour. Effects vary: less experienced and lower-skilled workers gain more, while the most skilled show small speed gains and small quality declines. This is a workplace study using a comparison design, not a randomized trial of all institutions. The authors do not estimate aggregate employment or wage effects from this record.

A result has a task and a date

METR’s July 10, 2025 report supplies a contrasting setting. Sixteen experienced open-source developers completed 246 real repository tasks, randomly assigned to allow or disallow AI. The allowed condition took 19 percent longer, with a reported confidence interval from 2 to 39 percent longer. The tools and observations belong to early 2025. This is evidence about those participants, tasks and conditions; it does not establish that AI slows most developers, has the same effect in other occupations or must remain ineffective as tools change.

The follow-up changes how that result should enter a current argument. METR’s February 24, 2026 update reports that newer estimates are unreliable because participation and task selection changed, pay was reduced and concurrent agents complicated time measurement. The authors believe speedups probably increased, but describe weak evidence for their size. This is not a clean reversal proving a universal benefit. It is a documented measurement problem that prevents a confident current percentage. Repeating only the earlier slowdown would conceal a material qualification.

The two workplace settings need not cancel each other. Assistance can fit one task and create additional checking in another. The support study and developer trial also use different designs and outcome measures. My judgment is that an institution should learn from this specificity rather than select the percentage that flatters its position. A procurement proposal needs evidence tied to its own work and a comparison that keeps service quality visible. A benchmark score or a popular anecdote cannot supply that comparison by itself.

Count the correction work

Consider a constructed workflow handling 1,000 cases. Its old process takes 20 minutes per case, for 20,000 minutes of staff time. A new system reduces initial handling to 12 minutes per case, or 12,000 minutes. Suppose 100 cases then need an additional 30-minute correction, creating 3,000 more minutes. Total staff time is 15,000 minutes, a saving of 5,000 minutes or 25 percent. All counts and durations are invented; this is an accounting example, not a measurement of a benefits agency, court or water authority.

Change only the correction requirement to 300 cases at the same 30 minutes each. Total time becomes 21,000 minutes, five percent above the old process. The initial handling improvement is unchanged. The overall comparison turns on work that an intake-speed dashboard could omit. This does not forecast an error rate or show that automation necessarily increases it. It demonstrates why the chosen boundary matters and why initial speed, final accuracy and repair effort should be reported separately.

A process improves for people when its completed service and correction path improve together.

Even that calculation has a limit. It counts staff minutes but omits time a person might spend discovering a problem, supplying another document or requesting review. A lower operator cost can coexist with a larger burden elsewhere. My proposed evaluation would record those burdens separately and identify which are measured, estimated or unknown. It would not collapse them into a single invented dignity score. The relevant unit might be a correctly completed case, with a time boundary that includes reopening and correction.

Governance needs an operating record

NIST’s January 2023 AI Risk Management Framework 1.0 supplies a voluntary structure organized around governing, mapping, measuring and managing risk. It is guidance, not a legal compliance certificate. This essay cites that identified version; NIST’s current site says it is being revised. Naming an adopted framework therefore cannot prove that an institution has implemented its practices or that a person can obtain a correction. Those claims require an operating record and an identified authority responsible for the process.

My proposed minimum starts with a named decision owner and a clear account of what the system assists or decides. Keep enough version and case information to investigate a disputed result. Identify who can change it and how the affected person can reach that reviewer. Evaluate the path with realistic cases, including one the system handles incorrectly. These are editorial criteria for usable accountability. The legal requirements depend on the jurisdiction, decision and institution; this review does not declare a universal appeal right or duty.

The complication is that a human reviewer can exist in an organizational chart without having time, information or authority to repair a decision. My test would ask whether review can actually change the outcome and how long that takes. It would also examine whether the person must repeatedly explain the same problem. Those are proposed observations, not findings about an unnamed bureaucracy. Confidentiality may constrain public release of case data, but an institution can still state review scope and the limits of its evidence.

A useful institutional account is particular

A complete procurement account would also explain what happens when a vendor changes a model, withdraws a service or makes the previous result difficult to reproduce. I would look for a way to preserve relevant records, compare versions and reassess performance after a change. Those arrangements can be costly and imperfect. The question is whether the institution has considered them in the decision, rather than treating adoption as the end of oversight. This essay has not examined a particular contract or established a universal lock-in outcome.

The inherited litigation claims do not meet this standard. This review found no matching evidence for a sixty-percent discovery-cost reduction or a reduction from forty review hours to four, so both are withdrawn. A general institutional report cannot verify a specific workflow result merely because it concerns AI. The municipal and court scenes likewise do not become observed cases without a record. Removing those assertions makes room for examples whose assumptions and evidentiary limits a reader can inspect.

What the discourse misses is often the distance between a capability and a service that someone can rely on. The studies here show why an answer needs a setting, a date and a method; the constructed example shows why it needs a complete work boundary. The public judgment then asks what the institution does with that evidence. Progress would mean useful assistance, a credible correction path and responsibilities that remain visible when the system changes. That is a concrete ambition, and one that an operating institution can be asked to demonstrate.