Innovation Matters

The expertise inside the investment

Imagine a company asking an experienced team to help build an AI system that will eventually allow the same work to be done with fewer people. Employees explain the exceptions, correct the confident mistakes and supply examples of good judgment. The company pays for the technology, carries implementation risk and remains accountable for what it deploys. Both contributions matter. The difficult question is what happens to the employees’ contribution after the system becomes useful.

An AI transformation should address that question before expertise changes hands. Existing compensation may cover the work of contributing, but it does not automatically settle the fairness of a capability that continues producing value after a contributor’s role disappears. Sharing some sustained gains through compensation, time or ownership deserves consideration as part of the investment itself.

What workplace evidence establishes

The mechanism is more than speculation. In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of an assistant among 5,172 customer-support agents. Successful workers’ conversations helped train the system, and less-experienced workers benefited particularly from its assistance. The authors explicitly raise the question of compensation for workers’ training contributions.

That finding supports a discussion about recognition without deciding its terms. The system assisted agents who retained discretion over responses; the study did not demonstrate their replacement. It also could not observe changes in wages or overall labor demand. A finding that expertise can be transmitted through AI cannot establish how much an individual should receive, how long payments should last or whether a particular business has distributable gains. Those remain decisions for the people designing the transformation.

A dividend can take different forms

Money is only one possible benefit. A survey preprint by Donghyun Suh and Samil Oh, based on 5,512 employed Koreans surveyed in 2025, interprets reported efficiency gains partly as reduced work intensity. Workers could complete tasks with less active effort, even where output did not rise. Because the evidence is self-reported and comes from a particular employment setting, it cannot establish the cash savings available in another company. It does show why the value of AI should be examined beyond output alone.

Workers also differ in what they want automated and how much human involvement they prefer. WORKBank collected views from 1,500 workers across 104 occupations during 2025, using online recruitment and targeted outreach. Its varied responses challenge the idea of a single worker preference. It did not ask respondents to endorse continuing royalties.

The practical implication is to give affected employees a voice in defining the benefit. A team might value a gain-sharing payment, protected development time or a reduction in workload. Someone facing the end of a role may place more weight on transition support. A scheme that offers only future workplace benefits would miss people who no longer have a workplace there.

The strongest case against continuing payments

The strongest objection starts with the employment bargain: salaries already compensate expertise, including improvements that remain valuable after the employee leaves. Companies also supply capital, software, customers, management and the conditions in which expertise develops. Treating every useful contribution as a continuing individual claim could make administration costly and discourage investment.

Attribution presents a further problem. A system may combine thousands of examples, years of accumulated practice and later improvements by other people. A highly visible contributor is not necessarily the person responsible for the largest share of value. A promise of perpetual individual royalties could therefore create a false precision about contribution.

These objections favor a workable collective arrangement over a claim that every model output can be traced to an employee. They also require supporters of a dividend to distinguish a fairness principle from a proven incentive effect. The cited studies do not establish that sharing gains causes better cooperation or makes an AI transformation succeed.

Agree on the bargain before the gain

A useful starting proposal would identify the work being transferred, the people contributing and the forms of benefit under consideration. It would define a baseline for judging improvement, account for ongoing operating costs and explain how gains will be reviewed. The aim should be a rule people can understand and administer, with a stated duration and treatment of employees whose roles end.

The arrangement should also acknowledge uncertainty. If promised savings do not materialize, both the company and its employees need to know what follows. If the system creates value in a different way than expected, the agreement should provide a way to revisit its terms.

The central choice is whether contributors have a defined stake in a transformation they help make possible. Continuing value does not imply an unlimited claim. It does make an explicit bargain more defensible than asking people to transfer their expertise first and discover the distribution rules later.

What the purchase really buys

Consider a hypothetical electronic warfare platform that performs exceptionally well but requires its original supplier’s participation in every meaningful modification. The initial demonstration may settle whether it meets today’s requirement. It cannot settle whether the buyer will be able to adapt it when that requirement changes. A purchase can deliver impressive capability while leaving crucial choices in someone else’s hands.

I would accept a higher upfront price for meaningful freedom to maintain and adapt such a system. That preference has limits: independence must be usable, and its benefits must justify what the buyer gives up in performance, affordability or delivery time. The task is to price the dependency deliberately, alongside the equipment.

Ownership does not confer independence

The concern has a concrete basis in acquisition experience. GAO’s 2025 sustainment review examined five selected U.S. weapon programs and found dependence on single suppliers associated with data-rights shortfalls. Programs also struggled to review delivered technical information for completeness and accuracy. Officials described potential effects on costs and repair times.

Those findings distinguish possession from practical control. Receiving equipment does not necessarily mean receiving usable documentation or the permissions needed to use it for maintenance and modification. The audit concerns selected programs, so it does not measure the prevalence of dependence across all defense procurement. It nevertheless shows that rights and information can become consequential constraints after delivery.

For a buyer, the relevant question is therefore specific: which future actions remain available without the original supplier’s agreement? A broad promise of ownership is less useful than a clear account of the maintenance, replacement and adaptation work the purchase enables.

Freedom has a technical and commercial structure

The official modular open systems approach treats adaptability as an integrated business and technical strategy. Its intended benefits include competition, technology refresh and interoperability. Modular architecture and verifiable interfaces are paired with business arrangements that permit components to be added, replaced or changed during the system’s life.

That combination matters. Accessible interfaces are of limited value if the necessary permissions are unavailable. Broad permissions are equally limited if documentation is incomplete or a replacement component cannot be integrated responsibly. The approach offers a way to organize those considerations; its stated benefits are not a guarantee that every implementation will be cheaper or easier.

My proposed buying test would begin with the changes the customer reasonably expects to need. It would then ask what information, permissions, expertise and validation those changes require. This keeps the negotiation tied to practical freedom rather than an abstract demand for complete independence.

The cost of doing more yourself

A supplier can provide valuable integration knowledge and ongoing support. The buyer that assumes more responsibility must fund people who understand the system, maintain that knowledge and evaluate changes. An unused right to modify something may have option value, but it should not be mistaken for an existing organizational capability.

Delivery is a serious counterweight. GAO’s 2025 annual acquisition assessment found continued difficulty delivering innovative systems promptly and within budget. It described delays associated with immature technologies and recommended stronger product-development practices early in programs. The audit does not show that independence caused those delays. It does demonstrate why a procurement argument focused on future flexibility must still confront maturity and time to delivery.

A more adaptable option could be the worse choice if it cannot provide sufficient capability when needed. Likewise, paying for extensive rights may be poor stewardship if the buyer has no realistic plan to exercise them. An independence premium consumes resources that could support additional equipment, training or maintenance today.

Negotiate the dependency deliberately

The useful comparison is between complete support arrangements. One option might offer stronger immediate performance and substantial reliance on the supplier. Another might cost more initially but permit more maintenance and adaptation by the buyer or alternative providers. Each should be judged against the same expected changes, staffing assumptions and delivery needs.

A practical decision record would identify which dependencies are acceptable, which could prevent an essential future change and what would trigger a reassessment. It would assign responsibility for reviewing the information delivered, keeping it usable and sustaining the skills needed to act. Promised flexibility should be examined as part of acceptance, alongside promised performance.

The supplier deserves a return on its invention. The customer deserves to understand the freedom it has purchased. A deliberate agreement can recognize both, including a choice to remain dependent where the advantages justify it. The mistake is allowing today’s performance and price to answer a different question: whether the capability can continue evolving on terms the buyer can live with.

Make disagreement testable

A research proposal can be unlikely to succeed and still be worth testing. If a consequential assumption is wrong, a rigorous experiment may change the direction of an entire program. If the assumption survives, a decisive result may close an expensive line of inquiry. The value lies in the decision the evidence enables, rather than the pleasure of backing an outsider.

I would protect some research funding for credible challenges that qualified reviewers expect to fail. Reserving 10% is a useful proposal to examine, not an established optimum. Every project admitted to that reserve would need a discriminating experiment, a reason the answer matters and a route to independent confirmation before a larger commitment.

Expert doubt is useful information

There is institutional room for unconventional research. The NIH High-Risk, High-Reward program supports innovative approaches, including work that could challenge existing paradigms. That is evidence that a major funder deliberately makes space for ambitious ideas. It does not endorse a fixed allocation to proposals reviewers expect to fail.

A protected reserve would need a sharper distinction between doubt about a hypothesis and doubt about the quality of its test. Reviewers may have sound reasons to expect a negative result. Those reasons should improve the experiment: identify the assumption at issue, explain what observation would be surprising and expose alternative interpretations before money is committed.

DARPA’s Heilmeier Catechism offers a disciplined model for asking about purpose, novelty, significance, risk, cost, timing and tests of progress. Applied to this proposal, those questions would require proponents to show why settling the uncertainty matters. Disagreement alone would confer no special claim on the budget.

A negative result must resolve something

The strongest methodological objection is that a failed experiment may teach very little. A prototype missing its target, an effect shown to be too small to matter and measurements too imprecise to support a conclusion are different outcomes. Grouping them together as useful failure would make the reserve easy to defend and difficult to evaluate.

Samuel Pawel and colleagues demonstrate why non-significance in an original study and its replication can still leave the evidence inconclusive. An absence of a statistically significant result does not, by itself, establish that an effect is absent. Their analysis gives substance to the concern that repeated negative findings can reproduce uncertainty.

A proposal should therefore state what a meaningful effect would be and how precisely the experiment can distinguish it from alternatives. It should explain what would justify continuing, stopping or redesigning the work. An inconclusive result may be honestly reported, but it should not receive the same credit as evidence that materially changes a decision.

Protect the result as well as the experiment

Funding a test is only part of the commitment. If an organization welcomes unconventional proposals but publishes only positive outcomes, the promised value of negative evidence can disappear. The Registered Reports model addresses that problem by reviewing the question and methods before outcomes are known, with publication contingent on the agreed quality and execution criteria. It provides a practical precedent for judging a study separately from whether its result is exciting.

A reserve could adopt that principle even outside a journal process: agree on the protocol, retain documented changes and make properly conducted outcomes available to the people allocating future resources. Exploratory findings can remain valuable when readers can distinguish them from the original test.

Replication also needs resources of its own. The Reproducibility Project: Cancer Biology encountered incomplete procedural information, difficulties obtaining necessary materials and experiments that required modifications. Those findings support budgeting for the practical work of checking a claim. They do not supply a universal replication rate for science or make one unsuccessful replication the final word.

Make the reserve answerable for its cost

The opportunity-cost objection remains. A dollar assigned to an unlikely challenge cannot simultaneously support a promising incremental advance. A fixed reserve can also become an entitlement: teams may learn to describe ordinary proposals as radical, while managers feel obliged to spend the allocation regardless of quality.

That is a reason to test the funding arrangement itself. A pilot should have a defined duration, explicit selection criteria and a review of which results changed subsequent research decisions. It should examine the cost of replication, the usefulness of negative findings and the projects displaced. If too few proposals meet the standard, the allocation should be reconsidered.

The case for funding doubt is conditional. Expert judgment should sharpen the test and make weak proposals harder to fund. It should also leave a path for a well-designed experiment whose answer could overturn an important assumption. Ten percent is open to challenge; the stronger commitment is to buy useful evidence even when the expected answer is no.

Authority should come with scrutiny

A company that records employees’ performance in increasing detail should be able to explain how it evaluates the leadership decisions shaping that performance. Consider an executive who sponsors an AI-based assessment system on the promise of better decisions, lower costs and fairer treatment. Those promises, the assumptions behind them and the results after deployment should become part of the executive’s own accountability.

Comparable scrutiny would not require a single scoring system for every job. It would require leaders to accept the same expectation of explanation, correction and review that they impose elsewhere. Greater authority should make consequential decisions more answerable, with procedural fairness extending throughout the organization.

The appeal and the concern share a source

The case for algorithmic management includes real perceived benefits. An OECD working paper published in 2025, drawing on a survey of more than 6,000 firms in six countries, reports that managers often perceived improvements in decision quality and their own job satisfaction. Those responses deserve consideration when judging why organizations introduce these tools.

The same managers also raised concerns about unclear accountability and difficulty following the tools’ logic. The category studied was algorithmic management, which includes software that may or may not use AI. These are employer perceptions, rather than an experiment proving that the tools improved decisions or a poll of employee support.

Together, the findings make a stronger case for examining the deployment decision than for assuming either success or harm. Leaders should be able to explain what benefit they expected, what evidence justified the expectation and how they would discover that the system was performing poorly.

Record the decision before the outcome

A useful leadership review would begin before deployment. Its record would include the problem being addressed, the alternatives considered, the intended beneficiaries and the conditions under which the system should be revised or withdrawn. Forecasts would preserve their uncertainty and time horizon. Known limitations would remain visible when later results are discussed.

This proposal extends an established governance principle. In its voluntary AI Risk Management Framework 1.0, NIST assigns executive leadership responsibility for decisions about risks arising from AI development and deployment under GOVERN 2.3. The framework also calls for clear roles and responsibilities. It supplies a governance reference, without demonstrating that organizations follow it or endorsing automated executive scoring.

For an assessment system, the record should explain what its measures represent and what they leave out. An AI tool could help organize that record and compare commitments with observed results. Any conclusions it produces would need the same scrutiny as the employee assessments: traceable evidence, attention to missing context and a person responsible for deciding what follows.

Reciprocity cannot repair a weak metric

The strongest objection is that strategic work differs from bounded operational tasks. Its consequences can emerge slowly, involve many contributors and depend on conditions outside a leader’s control. A simple numerical comparison could reward fortunate outcomes and punish reasonable decisions made under uncertainty. Applying such a measure to senior executives would not make it fair for everyone else.

A credible review should separate the quality of the decision from the outcome alone. Did the leader use the available evidence responsibly? Were important uncertainties acknowledged? When evidence changed, did the organization respond? A failed forecast may deserve a different judgment from a forecast that was never supported, even if the final numbers look alike.

Those distinctions must extend to employees. Context should not become a privilege reserved for leadership, while lower-level staff are reduced to scores. Comparable scrutiny means equivalent access to reasons, relevant evidence and correction, adapted to the work each person actually controls.

Make review capable of changing something

A review has little force if it produces a report without a consequence. The organization should identify who can require a flawed measure to be changed, pause its use or correct decisions made from unreliable information. People affected by an assessment need a usable way to challenge factual errors and add missing context. Leaders need the same opportunity to respond to criticism of their own decisions.

One practical starting point is to review the deployment itself at a stated interval. Compare the promised benefits with the evidence available, examine serious corrections and explain any changes to the system. Where findings are contested, preserve the disagreement rather than allowing an AI summary to erase it.

The standard is straightforward: scrutiny should reach the decisions that created the assessment regime. An executive willing to defend the criteria, acknowledge their limits and answer for the results provides a more credible basis for evaluation than authority alone. The organization then has a process for learning whether its measures deserve to be used at any level.

What a sign-off establishes

Consider a hypothetical AI-assisted electronic warfare system operating in conditions that change faster than its reviewer can understand. A person receives a recommendation and authorizes it. The procedure can identify who approved the action. It cannot, by itself, show that the person had enough information or time to make an informed choice.

Human approval remains consequential. Removing it could allow actions to proceed without anyone specifically authorizing them. The harder task is to design an arrangement in which responsibility corresponds to an actual ability to judge and intervene. A sign-off should be evidence of that capability, supported by understandable boundaries and tested limits.

Human agency has a substantive meaning

The ICRC’s May 2025 statement on autonomous weapons calls for binding prohibitions and restrictions intended to preserve human control over force. It expresses a humanitarian position about retaining human agency in consequential decisions. It does not establish that any particular approval mechanism achieves that aim.

The distinction in scope matters. Lethal autonomous weapons are not equivalent to every use of AI in electronic warfare. Evidence and arguments concerning lethal force cannot simply be transferred to every EW application. They do, however, bring the underlying governance question into focus: what must remain within human judgment for an authorization to carry the meaning assigned to it?

My proposed standard would ask what the person understands about the permitted activity, the evidence relevant to a decision and the limits of the system. It would also ask what the person can actually change. Formal responsibility without the practical means to act on it is an incomplete account of control.

People do not all respond to automation alike

Behavioral evidence challenges a blanket assumption that people inevitably defer to automated recommendations. A 2026 preprint by Lauren Kahn, Michael C. Horowitz and Laura Resnick Samotin compared 236 West Point cadets with a demographically similar public sample. In a controlled identification task, participants received advice from an algorithm or human analyst and could reconsider an initial answer. The cadets showed better-calibrated responses.

The finding leaves room for effective human-machine arrangements. It also has clear limits: the cadet experiment took place in May 2025, the comparison used earlier public data, and the task was controlled. The study does not establish a causal effect of training or demonstrate reliable oversight in rapidly changing EW conditions.

The useful implication is to examine human performance in the proposed arrangement instead of assuming it. A reviewer’s presence is insufficient evidence of control, but evidence of poor review in one setting would also be insufficient grounds for declaring human oversight futile everywhere.

Public attitudes distinguish among uses

A separate survey preprint covering nine countries provides another reason to avoid a simple choice between human approval and unrestricted autonomy. Based on 9,000 respondents surveyed in early 2026, it found greater acceptance of non-lethal and human-supervised applications, with support declining for fully autonomous lethal force.

The online-panel design and selected countries constrain how broadly those results can be applied. They show that the application matters to respondents; they do not measure current opinion among EW practitioners or establish that a supervised system is safe. Public acceptance and demonstrated control are different questions.

For an organization evaluating a system, this suggests a need to describe its actual role clearly. The relevant discussion concerns which decisions people make, which decisions software makes and how those responsibilities interact. Calling the arrangement supervised supplies too little information to judge it.

Make control demonstrable

A practical review should test the human role alongside the system. Can the reviewer recognize a meaningful departure from the conditions that justified authorization? Is the information needed to assess that departure available and understandable? Does the person have authority to withhold approval or intervene, with enough time for that choice to matter?

The standard should be realistic. No reviewer can understand every internal computation or eliminate every uncertainty. Meaningful understanding should concern the action being authorized, the relevant limits and the basis for trusting the arrangement in the circumstances at hand. Where those circumstances materially change, an earlier approval should not silently become permission for an unexamined new situation.

Responsibility also continues after authorization. The organization should be able to reconstruct the basis for a consequential decision, investigate failures and revise the arrangement when its assumptions no longer hold. These are proposed governance criteria, not a certification method. Their purpose is to make the claim of human control answerable to evidence about what a person can understand and influence.

When a tradeoff becomes a habit

A security requirement can acquire a permanent place in the performance budget long before anyone has established which part of the delay is necessary. Once slower service becomes the accepted price of protection, the organization has less reason to examine the machinery that produces it. The useful challenge is to separate the protection objective from the particular method used to meet it.

Cloudflare’s September 8, 2026 engineering report offers a concrete example. The company describes testing origin servers outside production requests, learning which key agreements they support, and using that information when opening connections. It favors compatible post-quantum key exchange and reports fewer handshake retries during its rollout. Those findings suggest that some delay can be reduced by changing when compatibility is established. They remain company-reported deployment results. Cloudflare engineering report

Understand what the retry protects

The TLS 1.3 specification explains why a retry exists. When the client has not supplied a suitable key share, the server can request another attempt through a HelloRetryRequest. If the two sides cannot agree on common cryptographic parameters, the handshake must abort. The extra exchange therefore serves a legitimate compatibility function. Eliminating it indiscriminately would miss the purpose of the protocol. RFC 8446, section 2.1

Cloudflare’s approach changes the information available at the start. Its documentation limits the feature to new TLS 1.3 origin connections in specified encryption modes and excludes connections through Cloudflare Tunnel. Requests using an existing connection do not repeat the key exchange. The documentation also separates key-share selection from the encryption mode that governs origin certificate validation. A faster handshake is consequently a specific improvement, with specific boundaries. Cloudflare documentation

For a team considering a similar design question, these boundaries determine the business case. Measure how often users encounter the affected operation and how much it contributes to their delay. An improvement at connection establishment may matter greatly to one service and barely register in another. The relevant benefit is the improvement users receive across the actual workload.

Complexity belongs on the same ledger

The strongest objection is that smarter coordination creates something else to maintain. A mechanism that learns capabilities, retains preferences and changes behavior needs dependable updates, understandable failures and a recovery path. Those requirements may outweigh the saved delay. They are questions for evaluation, rather than evidence that Cloudflare’s implementation is unsafe.

NIST’s crypto-agility white paper provides the counterweight. It warns that unnecessary options and complicated negotiation can enlarge implementations and create opportunities for defects or downgrade attacks. Adaptability has to be balanced with simplicity. That caution applies directly to the investment decision: a performance graph cannot establish the total security or operating cost of a redesign. NIST, Considerations for Achieving Crypto Agility

A credible proposal should therefore say which protection requirements remain fixed, what new state the design introduces, and who will diagnose its failures. Maintenance and recovery effort deserve the same visibility as latency. Otherwise the apparent saving may simply move from the user’s wait time into the on-call team’s workload.

Fund a question that can be answered

RERUT’s Paradox Shift principle is useful here as an investigative habit: examine the operating rule behind a tradeoff before treating the tradeoff as unavoidable. In this case, the question is whether compatibility knowledge can be gathered and governed earlier. The principle supplies a direction for inquiry; testing must establish whether the change works.

Start with one service and a documented baseline. Compare the existing design with an alternative under normal operation, changed server configurations and failed updates. Record connection failures, user-visible delay, recovery behavior and engineering effort. Set stopping conditions before the experiment, including any unauthorized reduction in protection, unacceptable reliability loss or operating burden beyond the agreed budget.

Cryptographic discovery and controlled interoperability testing already have distinct roles in NIST’s migration project. That separation offers useful background for preparing an evaluation, although the project does not validate this particular proposal. NIST NCCoE migration project

Accept the result, including the limit

The people responsible for the service should interpret the result together. The application owner can assess whether customers notice a benefit; the security reviewer can identify protection assumptions left untested; the on-call engineer can explain what becomes harder to operate. These perspectives turn a local optimization into a decision about the whole service.

A modest speed improvement may justify investment when configuration and recovery also become easier. A larger improvement may fail the case when its maintenance demands are disproportionate. Either conclusion is useful. Retiring the security tax means refusing to budget indefinitely for an unexamined implementation cost, while remaining willing to accept a measured tradeoff when the alternative does not earn its place.

The improvement that stops at the connection

Imagine a water-reuse project with a cheaper treatment process and an installation payment that some intended households still cannot afford. The process has improved. The households remain disconnected. This hypothetical project exposes a gap between a component’s performance and the outcome that justified building it: reliable access to a useful service.

Lower treatment cost is worth pursuing. The problem begins when that measure becomes sufficient evidence that the customer’s problem has been solved. A project can improve its average economics while leaving the hardest barrier untouched. If connection arrangements determine who receives the benefit, those arrangements belong in the definition of success, even when the engineering team cannot decide every funding question alone.

What an aggregate measure leaves out

A July 2026 study in Earth’s Future gives this measurement question a wider context. Gi Joo Kim and colleagues examine conditional scenarios for future resource security, differentiating food and residential-energy burdens by income group while measuring physical water scarcity regionally. Their model explores plausible futures rather than selecting one most likely outcome. The results show why differences across populations and regions deserve attention. They do not establish household connection costs or evaluate water-reuse financing. Kim and colleagues

The lesson for a local project is analytical: choose measures that match the outcome being claimed. Cost per unit answers a production question. The timing of household payments, eligibility, reliability and sustained participation answer different questions. Combining them into a single average can hide the distinction between making a service cheaper to produce and making it possible to use.

EPA’s Water Affordability Needs Assessment similarly cautions that affordability has no single measurement approach, and that household measures may be unsuitable for broader community planning. Its rate data also have geographic coverage limitations. A national assessment can guide the questions a project asks; it cannot substitute for evidence about the households the project intends to serve. EPA assessment and FAQs

The financing objection deserves an answer

Access cannot become an unlimited obligation attached to a finite budget. EPA recognizes that necessary capital investment and operating costs can increase household burdens, while smaller utilities can lack the administrative capacity to run assistance programs. A proposal that overlooks those constraints may undermine the service it is trying to extend. EPA assessment and FAQs

The National League of Cities made a related argument in its March 2023 critique of EPA’s financial capability guidance. Carolyn Berndt challenged reliance on median household income and advocated investment flexibility that better reflects vulnerable residents. This was a municipal association’s policy position, with concerns about both household burdens and workable investment schedules. It illustrates why the financing perspective cannot be reduced to resistance to inclusion. National League of Cities commentary

These concerns make the design problem more demanding. An alternative must identify who pays, when funds are needed and which obligations continue after installation. Moving a payment away from the connection date may help a household join, but the project still needs a credible way to cover the cost. Shared assets may change the economics while introducing ownership or maintenance responsibilities that require agreement.

Follow the household through the system

RERUT’s Gemba & Perspective Analysis starts with the journey from intended beneficiary to reliably connected customer. Households, field technicians, customer-service staff and finance personnel should help examine that journey. The aim is to establish whether the barrier is an installation payment, eligibility, ownership, maintenance or unreliable service. Each calls for a different response.

Then compare the original connection design with one financially bounded alternative, such as shared infrastructure or phased connections. Agree on service requirements, available funding and cost responsibilities before assessing the options. Compare total lifecycle cost, the timing of household payments and sustained access. Keep intended households that remain unconnected in the evaluation; otherwise the project can quietly define success around those already able to participate.

A phased proposal deserves particular scrutiny. Connecting some households sooner can be useful, but a later phase needs a credible funding and completion path. The assessment should make waiting visible alongside the benefits delivered to early participants. A promise of eventual inclusion is weaker evidence than a plan with assigned responsibilities and committed resources.

Make the boundary of responsibility visible

Engineering teams should be responsible for exposing the gap between technical performance and usable service, proposing a credible alternative and showing its consequences. Funding authorities and service owners must make the choices that exceed the team’s remit. That division works only when someone owns the unresolved barrier and the decision is recorded.

The alternative may prove too expensive, delay useful service or transfer burdens to another group. If the original design delivers better sustained access within the available resources, retain it and describe the remaining gap honestly. A better definition of success does not guarantee that every barrier can be removed. It ensures that an impressive average cannot make the people still waiting disappear.

Choice begins before deployment

A patient’s refusal becomes difficult to accommodate when the service has already been built around a single pathway. By that point, the discussion can narrow to explaining the technology and securing acceptance. A better starting point is to examine choice while the service is still being designed: what happens when an informed patient declines a particular AI-assisted intervention?

Consider a hypothetical clinic where an AI-assisted screening pathway is safer than an established alternative. Assume the alternative remains clinically acceptable and the patient understands the difference. My position is that the service should examine how to offer that alternative. The assumption of greater safety makes the choice harder, but it does not make the patient’s preference irrelevant. This scenario tests a service-design principle; it describes no actual screening product or clinical result.

What the commission actually recommends

The UK National Commission into the Regulation of AI in Healthcare addresses this tension in its September 10, 2026 report. Recommendation 35 calls for proportionate transparency and the ability to opt out “where possible or appropriate.” The surrounding discussion also warns that seeking explicit consent for every supporting use of AI could be impractical and detrimental to equitable care. It proposes general awareness alongside more active discussion where the context warrants it. Commission report, section 3.1

The qualification matters. Refusing one screening tool is a different service-design question from requiring a separate authorization for every technology supporting care. A useful policy must define the use under discussion and the circumstances in which a choice can be offered. A blanket promise would obscure those differences.

These are advisory recommendations. In its announcement, the MHRA said the government and regulator would consider them and issue a formal response. The report itself does not create a universal legal entitlement to a separate pathway without AI. MHRA announcement

Conditional support leaves room for refusal

The Health Foundation’s research summary describes participants’ support for healthcare AI as conditional on safeguards, including accuracy and human oversight. That finding supports taking expectations seriously, while leaving the funding of parallel services unresolved. The Foundation was the commission’s research partner, so these documents are connected parts of the evidence base. Health Foundation report summary

For service design, the practical implication is to understand what a patient is declining. The concern may involve a particular tool, its use in a particular decision, or the role of human judgment. A conversation that explains only the technology’s benefits may leave the actual disagreement untouched. A conversation that promises an equivalent alternative without examining the differences would be equally inadequate.

A meaningful option requires an accurate explanation of what changes when the patient chooses it: the clinical process, the responsible clinician, any expected delay and the uncertainties. Clinical acceptability remains a condition of the proposal. Respecting a preference does not justify presenting an unsafe workaround as a reasonable alternative or describing unequal options as identical.

Capacity is part of the ethical question

The strongest objection is the effect on other patients. Maintaining a separate route could consume staff time, create additional handoffs or reduce capacity elsewhere. The commission’s warning about disproportionate consent burdens reinforces the need to assess how a policy affects the service as a whole. It does not quantify the cost of any particular alternative. Commission report, section 3.1

That objection deserves a concrete comparison. Start with a planning exercise involving patients, clinicians and operations staff. Map one proposed route from the initial explanation through scheduling, screening, interpretation and follow-up. Identify who decides clinical eligibility, who handles a refusal and what happens when the preferred route is unavailable. RERUT’s Gemba & Perspective Analysis makes those encounters the starting point for deciding what the service can responsibly promise.

Compare the route with the current plan using documented staffing assumptions, projected waiting times and explicit clinical criteria. Ask patient participants whether the explanation makes the differences understandable. Treat forecasts as estimates until operational evidence supports them. Any subsequent clinical evaluation belongs within the appropriate clinical and governance processes.

A defensible decision can change

The decision should turn on whether the alternative is clinically acceptable, understandable and supportable without an unreasonable loss of care elsewhere. If it fails those conditions, the service should explain the constraint and the available next steps. If scheduling or referral changes could make it workable, that possibility should remain open for review.

Funding an alternative indefinitely is a larger commitment than examining one before deployment. The latter should be a normal part of accountable innovation. It gives patients a meaningful place in the decision, gives clinicians and service leaders a way to articulate limits, and prevents a procurement choice from silently settling a question about care. Patient refusal deserves a designed response, even when the final answer has boundaries.

A successful plan can leave capability unexplored

A team can meet every objective and still leave its most consequential possibilities unexplored. If every measure rewards delivering the plan or slightly exceeding it, the scorecard answers a limited question: did the team achieve what it was asked to achieve? It says much less about what a different method, sequence or allocation of work might make possible.

That is the ceiling a strategy can create. The risk arises when attainment becomes the outer boundary of leadership’s curiosity. Employees have little reason to investigate alternatives if discovery receives no time, resources or recognition. The remedy begins by distinguishing the outcome worth imagining, the result the organization can responsibly promise, and the evidence that could justify changing that promise.

Dependable commitments still matter. Customers need delivery, and teams need protection from arbitrary demands. A prudent forecast may reflect uncertainty rather than low ambition. Equally, an ambitious possibility has not earned a place in the forecast simply because it is exciting. Strategy should make room for both responsibility and investigation.

Goals direct attention, including away from discovery

Edwin Locke and Gary Latham’s 2002 synthesis explains how goals direct attention and effort toward relevant activities. It also emphasizes commitment, knowledge and feedback, and discusses the value of learning goals when unfamiliar work requires people to discover effective strategies. That makes the kind of goal consequential: a demand for more output cannot by itself supply the knowledge needed to produce it. Locke and Latham

The implication for leadership is to ask whether learning is part of the assigned work. A delivery objective might specify the result due this quarter. A related learning objective could examine whether a different method reduces rework or removes a recurring delay. The second needs a defined question and usable evidence, rather than a vague expectation that everyone will somehow become more innovative.

The research supports examining this mechanism. It does not prove that conventional planning invariably suppresses innovation or that every team has unused capacity waiting to be released. The ceiling is a leadership metaphor for a question left unanswered.

A higher quota can reinforce the same problem

The evidence gives little reason to replace an unexamined ceiling with an unsupported stretch target. In two business-simulation experiments, Michael Shayne Gary and colleagues found no positive main performance effect from stretch goals and lower risk-adjusted performance compared with moderate goals. Some participants performed better, but that outcome did not generalize across the experiment. Gary and colleagues

Saeedeh Ahmadi and colleagues also found important differences among employees in one company’s idea-generation program. The relationship between stretch goals and useful ideas varied with experience, previous success and seniority. Increased participation was therefore insufficient evidence of equal benefit. This observational study cannot establish one ideal approach for every organization, but it challenges the assumption that the same pressure helps everyone innovate. Ahmadi and colleagues

Leaders should be alert to a simple substitution: asking for more while leaving the method and workload unchanged. That can turn the language of possibility into an additional delivery burden. A credible discovery effort instead creates room to test how work could change, with permission to conclude that a proposed method does not help.

Make room without assuming that every constraint is waste

Time matters, but more time does not guarantee better ideas. Ajay Agrawal and colleagues linked university breaks with additional crowdfunding projects, including increases at both high and low ends of measured quality. The setting offers evidence about available time and experimentation, rather than an optimal utilization rate for a corporate team. Agrawal and colleagues

An integrative review by Oguz Acar and colleagues likewise describes constraints as capable of helping or hindering creativity, depending on their type and context. The leadership task is to distinguish a protective limit, a useful creative boundary and an inherited assumption. Treating all three alike makes both planning and experimentation less informative. Acar and colleagues

RERUT’s Constraint-Free Divergent Exploration offers a practical exercise: ask the people doing or observing the work to suspend one consequential assumption during imagination. Explore the outcome a different method might allow. Then bring cost, quality, safety, workload and dependencies back into the discussion before approving a test. Freedom in the exercise should produce a hypothesis that can survive real conditions.

Give discovery a place in the review

A performance review can preserve delivery measures while asking what assumption the team tested, what it learned and what changed as a result. The experiment needs a baseline, bounded resources and stopping conditions. Evaluate repeated attempts rather than a single promising demonstration. Include total staff hours, rework and displaced work so the apparent gain cannot depend on hidden overtime or weaker quality.

Recognize a well-run test that rules out an idea. Its value lies in a better decision, even when the existing method remains preferable. When a repeatable improvement does emerge, use the evidence to revise the strategy and resource the change. Reliable delivery deserves recognition; demonstrated capability deserves a route into the next commitment. The plan becomes a ceiling when that route is missing.

Permission after disappointment

Permission to experiment has little meaning if a leader silently withdraws it when the outcome disappoints. The position in this article is a standard for judging responsible exploration: consider what the person reasonably knew, the authority they had, the safeguards they respected and their candor afterward. A poor result can matter without supplying a complete verdict on the person. This is a leadership judgment, not a claim that every unsuccessful effort deserves continued funding.

What psychological safety can tell us

Amy Edmondson’s June 1999 field study examined 51 teams in one manufacturing company and found psychological safety associated with learning behavior. The paper treats psychological safety as confidence around interpersonal risks such as asking for help or discussing mistakes, rather than unrestricted permission to act. Its observational design does not establish causation, and it cannot establish the employee climate of Tesla, SpaceX or any other company. Edmondson, 1999.

Two different ways to learn

The two company examples illuminate different kinds of learning. Tesla’s May 2, 2018 update acknowledged introducing too much automation too quickly and described temporarily restoring selected manual or semi-automated processes. That is a company-reported correction of an operating approach. Tesla, 2018. NASA’s January 19, 2020 account describes a completed Crew Dragon launch-escape demonstration in which Falcon 9 broke apart as planned. That was successful testing designed to include hardware loss, not an accidental failure relabeled as success. NASA, 2020. Neither source measures whether employees felt safe disagreeing, reporting mistakes or proposing another experiment.

Accountability still matters

The strongest competing concern is that authorization can be mistaken for immunity. An approved experiment may still expose inadequate preparation, ignored warnings or conduct outside agreed limits. Conversely, a conscientious person can pursue a reasonable idea that does not work. My position is that those circumstances deserve individual consideration rather than an automatic reward or penalty based on the final result. People affected by the consequences also belong in that judgment; support for the experimenter must not erase their experience.

Making room for human judgment

The AI connection is deliberately conditional. If automation takes on more routine processing, leaders still choose what room to make for human curiosity, interpretation and critical thought. This article does not assert that AI only handles routine work, that standardized work never requires judgment, or that particular jobs will disappear. RERUT’s Bulletproof Psychological Safety principle contributes the expectation that candid learning remains possible after disappointment; Gemba & Perspective Analysis adds attention to how the leader’s response is experienced by the people doing and affected by the work.

The circumstances change the judgment

Different contexts may justify different boundaries. A tightly contained prototype and an experiment affecting patients or customers cannot be judged solely by the shared label of innovation. One non-exclusive question for a review is whether the assessment of preparation changes simply because the outcome is now known; another is what evidence of conduct, disclosure or preventable harm should change that assessment. These are considerations for critical thought, not a universal procedure. The public question remains open: how much should an authorized experiment’s outcome change our opinion of its proposer?

Listen before the label becomes the decision

Calling someone a “good idea fairy” is a poor substitute for understanding what they’re trying to fix and a dated idea we need to talk about. The colleague pitching an unfinished idea and the one saying “we’ve tried that” both have experience worth hearing, and more importantly signal a cultural rift that needs attention. I want to understand what one sees as possible and what has led the other to expect failure.

I use “cultural rift” here as a leadership interpretation: a reason to investigate the distance between what people believe is possible and what they expect the organization to support. Disagreement alone does not establish that a culture is broken. The idea may be impractical, the objection may be outdated, or both people may know something the other has not yet considered.

Different contributions deserve a hearing

Liang, Farh and Farh’s two-wave study of 239 employees distinguished suggestions for improvement from warnings about harmful practices. Different psychological conditions were most strongly associated with each form of voice. These are behaviors, not permanent categories of people, and the associations do not establish causation.

Satterstrom, Kerrissey and DiBenigno followed one healthcare team over 31 months. Of 208 ideas voiced upward, 49 were implemented after apparent initial rejection. Colleagues helped proposals develop, including by raising specific problems. This illustrates a process in one setting; it does not establish a universal success rate or prove improved performance.

For me, the implication is a question of attention. A polished presentation does not settle whether an idea is useful, and an awkward first contribution should not settle whether its originator deserves to be heard. Understanding the contribution comes before deciding what to do with it.

The history behind an objection matters

Wanous, Reichers and Austin examined change cynicism as pessimism about future change coupled with beliefs about who was responsible for earlier failures. Their six-sample study gives leaders a reason to explore the history behind a response. It does not establish that a cynical employee has correctly diagnosed a problem, and skepticism should not automatically be called cynicism.

Someone who expects failure may be protecting colleagues from work that has previously gone nowhere. They may also be applying an old lesson to circumstances that have changed. I want to understand that distinction before treating their position as wisdom or resistance. The same care belongs to the person whose enthusiasm has outrun the detail.

Development requires judgment about support

Listening does not require funding every suggestion or giving every objection a veto. It requires judgment about who needs time, expertise and direction to develop an idea, and how a skeptical colleague’s specific concerns could help that idea become workable.

Baer’s research examined data from 216 employees and their supervisors. Implementation of creative ideas depended on more than creativity; motivation and supportive relationships also mattered. I take that as a reason to examine the conditions around a proposal rather than expecting its originator to arrive with every skill and relationship already in place. It is not evidence that funding every suggestion would pay off.

Direction without dismissing the person

Within RERUT, the Superpower Funnel asks whether an apparent weakness or threat can become a useful capability. In this context, unstructured enthusiasm might need expertise and time; a specific objection might contain a design requirement someone else missed. The Resource Rule matters equally: low-yield work can be documented and retired without treating the person who raised it as disposable. These are RERUT design principles, not a scientifically validated promise of results.

There are real limits. An oversimplified proposal can transfer substantial work to colleagues. Listening cannot become indefinite delay, a skeptical veto, or privileged access for the loudest people. Leaders still have to consider feasibility, safety, workload, competing commitments and the people who would carry the work.

What happened after you listened

The relevant judgment depends on the contribution and its context. An idea may need clarification before money; a concern may need evidence before authority. People also deserve to understand why a decision was made and to retain credit for what they contributed. The aim is to make useful thinking easier to recognize and develop while remaining willing to decide that a particular proposal should not proceed.

When has an unlikely idea or an uncomfortable objection changed your mind about what your team could achieve? What did you do about it?

Different visions of improvement

A promise of greater productivity or an initial improvement in profit margins can sound very different to someone imagining better work and someone wondering whether they will still have a job. Organizations should always pursue improving value, but better margins alone do not tell us whether we have done that. I think some of what we call resistance to change is a disagreement about the future we are building—and whether the people helping build it will still have a meaningful place in it.

The distinction matters because a financial improvement and a human concern can both be real. Reducing the effort required for a task may be useful, but that result does not decide what the organization should do with the capacity released. Leadership makes that choice, and employees experience its consequences.

The capabilities behind a brand

In The AI Dividend, I argued that employees should share in AI’s gains, and in Permission Must Survive Failure, I argued for making room for human curiosity and critical thought. Hemorrhaging human capital can also mean losing the accumulated knowledge, creative judgment and relationships that built a brand and its capabilities. I believe treating those people as expendable risks undermining faith in the brand while weakening its ability to create what comes next.

This deliberately connects [The AI Dividend](/discussions#the-ai-dividend) with [Permission Must Survive Failure](/discussions#permission-must-survive-failure). The first considered employees’ continuing stake in value they help create; the second considered the conditions for human curiosity and responsible experimentation. Here I am bringing those arguments together around reinvestment and the capabilities an organization risks losing.

What the evidence supports

Schmitt, Borzillo and Probst’s knowledge-retention paper, first published online in June 2011 and appearing in a 2012 journal issue, develops a framework for understanding knowledge loss during employee downsizing. Its accessible abstract identifies risks to critical skills, experience and capabilities. It supports taking those risks seriously; it does not establish that every workforce reduction destroys a brand.

Brynjolfsson, Li and Raymond’s revised November 2024 manuscript reports a 15% average increase in issues resolved per hour among 5,172 customer-support agents given AI assistance, with effects varying across workers. This is evidence of a specific productivity gain, not a general answer about employment levels or where every organization should reinvest.

Pew’s February 2025 report found that 52% of U.S. workers were worried about future workplace AI use. Those are historical perceptions, not a current displacement estimate or proof that anxiety causes resistance. My argument about faith in a brand is a leadership judgment about risk, not a measured universal effect.

Reinvesting in human possibility

RERUT’s reinvestment strategy brings these ideas together: use AI to relieve burdens that consume employees’ psychological and emotional capital, giving them room to ideate, imagine and accelerate the creation of value. That means strengthening the human judgment, care and relationships that make delivery valuable, with people empowered to shape better work—not simply expected to produce more or be eliminated.

Psychological and emotional capital are used here to describe the human capacity to engage, care, think and imagine, not a quantity measured by these studies. Removing routine work does not automatically restore that capacity if the organization replaces it with greater pressure or insecurity. Whether reinvestment improves value depends on what people can actually do with the room it creates.

This position does not establish that every role must remain unchanged or that every proposed reinvestment will succeed. Financial viability, quality, customer needs and the consequences for people all remain part of the judgment. The question is whether an efficiency decision preserves and develops the capabilities on which future value depends.

How should leaders judge whether an AI transformation is creating lasting value when its efficiency gains may come at the expense of the people and capabilities that made the organization valuable?