<p><strong>TL;DR:</strong> AI can accelerate parts of knowledge work without changing the meetings, queues, and authority structures around them. That does not make review unnecessary. It makes vague, one-size-fits-all approval harder to defend. Organizations should separate quality, risk, strategic coherence, and accountable ownership; move repeatable checks closer to creation; grant bounded decision rights for lower-consequence work; and reserve named human scrutiny for the decisions that genuinely require it.</p><p>The draft takes six minutes.</p><p>Then it spends three days waiting for someone senior enough to feel nervous about approving it.</p><p>During those three days, the team explains the work in two meetings, adds comments for people who will never read the source material, and moves one adjective back and forth until the document acquires the unmistakable scent of governance.</p><p>Nothing unusual has happened.</p><p>That is the problem.</p><p>When production itself was slow, a final review at the end could feel proportionate. The work took time to research, assemble, write, analyse, design, and revise.</p><p>Now production time can shrink sharply in some kinds of knowledge work. Approval routines often do not.</p><p><strong>When production accelerates but decision routines stay fixed, undifferentiated approval can become the visible constraint.</strong> That has not happened everywhere, and speed does not mean the work is ready. The mismatch is simply harder to miss: creation gets faster while authority stays exactly where it was.</p><h2>The machine got faster. The meeting kept Tuesday at 3:00.</h2><p>The evidence for this timing mismatch is suggestive, not universal.</p><p>In a six-month randomized field experiment involving 6,000 knowledge workers, access to generative AI reduced time spent on email by about three hours per week and appeared to help people finish documents moderately faster. Meeting time did not change significantly. The <a href="https://www.microsoft.com/en-us/research/publication/shifting-work-patterns-with-generative-ai/" target="blank" rel="noopener noreferrer">Microsoft Research study</a> did not measure approval queues, so it cannot tell us that approval is the dominant bottleneck. It does show something more useful: individually adjustable work can change before coordination-heavy routines do.</p><p>That matters because an organization can install new capability inside old choreography. A team can generate a credible first draft before the calendar invitation for the kick-off meeting has found everyone. The technology arrives at machine speed. Authority continues by Outlook consensus.</p><p>The gains also vary by task and context. In a field study of 5,179 customer-support agents, AI assistance increased issues resolved per hour by 14% on average. Less experienced workers gained more; the most experienced saw little benefit. The <a href="https://www.nber.org/papers/w31161" target="blank" rel="noopener noreferrer">research</a> is strong evidence for that deployment, not a promise for every team, function, or tool.</p><p>The point is not that every task now takes minutes. It is that <em>some production cycles have compressed enough to expose every unchanged wait around them</em>.</p><h2>Fast output is not finished work</h2><p>There is an obvious objection to any argument against final review: sometimes the draft is fast because it is wrong.</p><p>Quite right.</p><p>In an experiment with 758 BCG consultants, people using GPT-4 completed 12.2% more tasks and worked 25.1% faster on tasks judged to be within the model’s capability frontier. On one complex task outside that frontier, however, AI users were 19 percentage points less likely to reach the correct answer. The <a href="https://doi.org/10.1287/orsc.2025.21838" target="blank" rel="noopener noreferrer">peer-reviewed study</a> is a useful warning against confusing production speed with release readiness.</p><p>Removing every review would be a category error. Legal exposure, safety, financial consequences, reputational risk, effects on people, and hard-to-reverse decisions can justify deliberate human scrutiny. Some work should slow down.</p><p>But that does not rescue the default ritual in which everything, regardless of consequence, waits for the same final blessing.</p><p>So do not ask whether approval should disappear. Ask:</p><p><strong>What job is this approval supposed to perform?</strong></p><h2>Four jobs hiding inside one word</h2><p>"Approval" often compresses several jobs into one vague request. This four-part model is a practical decomposition, not a universal or exhaustive taxonomy.</p><ul><li><strong>Quality and correctness.</strong> Does the work meet the standard? Are the facts, calculations, logic, and execution sound?</li><li><strong>Risk and compliance.</strong> Could it harm someone, violate a duty, create legal or regulatory exposure, breach a policy, or cross the organization’s risk tolerance?</li><li><strong>Strategic coherence.</strong> Does it fit the intent, priorities, customer reality, and trade-offs the organization has chosen?</li><li><strong>Accountable decision ownership.</strong> Who is authorized to release the work, own the consequences, monitor the outcome, and correct it if reality disagrees?</li></ul><p>Those jobs require different information, different expertise, and different timing. Treating them as one final review usually leaves everyone unsure which job they are doing.</p><p>A senior leader may be asked to verify a claim they are not equipped to check, confirm a policy another function owns, assess alignment against context they never made explicit, and provide the reassuring sensation that an important person has touched the document.</p><p>The comment appears, and civilization survives.</p><h2>Move judgment closer to the work</h2><p>Better approval design does not remove judgment. It puts each kind of judgment where it can do useful work.</p><p>Repeatable quality checks should happen during creation, supported by clear standards, capable peer review, and automation where it is reliable. Risk should be classified early enough to change the approach, not discovered as a red comment twenty minutes before release. Strategic coherence should come from shared context before the work begins. Decision ownership should be explicit, with defined authority and a named path for exceptions.</p><p>Software delivery offers a concrete example. DORA reports that heavyweight approval by an external change-advisory board or senior manager is associated with worse delivery performance, without evidence that more formal external review reduces change-failure rates. Its <a href="https://dora.dev/capabilities/streamlining-change-approval/" target="blank" rel="noopener noreferrer">guidance</a> recommends peer review during development, automated checks, early identification of high-risk changes, and extra scrutiny for those changes.</p><p>That is an association from software delivery, not a causal result or cross-industry proof. Still, the design principle travels well: <strong>put routine checks near the work and reserve escalation for meaningful risk</strong>.</p><p>This changes what leaders must provide. If people can act only after a leader inspects every artifact, the organization has not distributed judgment. It has distributed production while keeping responsibility in a queue.</p><p>This is where the argument connects to <a href="https://www.linkedin.com/pulse/delegation-trap-ai-can-do-your-work-cant-own-outcome-maciejewski-123re" target="blank" rel="noopener noreferrer">The Delegation Trap</a>. AI can help produce the work, but organizations still have to decide who owns the outcome. Making a senior leader approve every sentence does not solve that. Decision rights have to match consequence, reversibility, competence, policy, and risk.</p><h2>A final review should be a designed interruption</h2><p>NIST’s voluntary <a href="https://airc.nist.gov/airmf-resources/airmf/5-sec-core/" target="blank" rel="noopener noreferrer">AI Risk Management Framework</a> treats governance as continuous across the lifecycle. It calls for defined roles and responsibilities, documented human oversight, and risk-management activity proportionate to impact, likelihood, and organizational tolerance. It is guidance, not law, and it does not prescribe abolishing human approval. It asks organizations to be specific about how governance works.</p><p>That specificity matters because a human signature is not the same as meaningful oversight. <a href="https://airc.nist.gov/airmf-resources/airmf/appendices/app-c-ai-risk-management-and-human-ai-interaction/" target="blank" rel="noopener noreferrer">NIST notes</a> that human roles differ and human-AI outcomes vary. In a <a href="https://doi.org/10.3233/FAIA240598" target="blank" rel="noopener noreferrer">2024 experiment</a> involving 1,411 HR and banking professionals in Europe, generic human oversight did not prevent discriminatory choices supported by a biased AI system. Participants wanted clearer guidance about when to override it. The result applies to those tested sensitive-decision scenarios, not to human review in general. It shows what generic oversight can lack: a clear job and usable override rules.</p><p>Before sending another item into the approval queue, ask:</p><ul><li>What risk or quality problem is this review meant to control?</li><li>Who has the competence and context to judge it?</li><li>Could the check happen earlier or automatically?</li><li>What consequence, uncertainty, or exception should trigger escalation?</li><li>Who may decide, and who owns monitoring and repair afterward?</li></ul><p>If those questions have no clear answers, the queue is not governance. It is institutional anxiety with a workflow status.</p><p>Leadership work here is not endless permission. It is providing enough purpose, context, boundaries, and decision rights for people to act without guessing what will be punished later.</p><p><strong>A final review should be a designed interruption, not a ceremonial pause.</strong></p><p>Keep it where the consequence demands it. Move it earlier where the work allows it. Remove it where it controls nothing.</p><p>Then let authority catch up with the speed we have already bought.</p><p>Rob</p><p><em>P.S. Where does work wait longest in your organization: quality, risk, strategic coherence, or someone’s permission to decide?</em></p>
