The Agent Delegation Canvas: human capacity as a condition of AI delegation

MBA Final Work · Entrepreneurial venture · Innovation

Abstract

AI delegation can reduce human work while leaving duties that qualified people must complete at specific times. This paper proposes the Agent Delegation Canvas to connect that work to managerial permission. Under its circuit-breaker rule, managers withhold or suspend affected authority when qualified people cannot perform the required duties in time. Resumption requires checking that the required support has been restored. The Canvas draws on allocation and governance research. Two constructed cases show the decision process. A staffing cut fits total available hours but misses a specialist response deadline. A family-office team first limits the number of ideas to preserve full investment analysis, then allocates spare time to further reasoning. Five ideas fit the original three-idea preparation window; six exceed it. The cases distinguish which arrangements can be authorized from which are worth choosing. For Mosaic, the author's commercial venture developing a valuation workspace, the paper recommends that customer teams control analytical scope, reserve qualified assessment time and choose how to use research savings. Practical evaluation should test whether the Canvas improves decisions enough to justify its use.

1. The managerial problem

Consider a service director who has introduced an AI agent to handle routine enquiries. With less work left for the team, moving one of two specialists to another assignment seems reasonable. The director adds up the remaining work and checks it against the hours of the specialist who would stay. It fits. Then the director looks at when that work is needed. Six cases may require ninety minutes of specialist attention, all between nine and ten in the morning. The specialist has enough time across the morning, but cannot give those customers ninety minutes within one hour. Free time later cannot help them before ten. Total hours fit; the response promise does not.

The director's decision raises the research question: How should managers authorize AI delegation when automation changes the amount, composition and timing of the human work required to sustain it? This paper proposes that permission should depend on whether qualified people can perform the required human work when it is needed. A manager must check that support before authorizing the workflow. If a later change makes the required work infeasible, the manager must suspend the affected permission until support is checked and restored. The paper calls this continuing condition the circuit-breaker rule.

A referral still needs a specialist who can respond in time. A pause control still needs someone able to intervene. The paper proposes the Agent Delegation Canvas as a way for managers to examine these dependencies when deciding whether to authorize an AI workflow. It records the human duties required by specific actions and controls, the reasons for the decision and what must happen if support is lost.

An agent has discretion to select or compose workflow actions. Capability is what it can technically do; authority is what an entitled decision maker permits. Required human duties are work on which the service, controls or recourse depend. Recourse lets affected people seek review or correction.

This question arose during my development of Mosaic, a commercial venture developing a hosted valuation workspace for asset managers and analysts of publicly traded companies. I am its builder and have a commercial interest. The minimum viable product (MVP) brings company models, scenarios, assumptions, source evidence and saved versions into one workspace. Customers' chosen agents prepare model changes and request calculations from Mosaic's engine. Users then inspect the results, adjust assumptions and save reviewed versions in the browser. They retain judgment over assumptions and interpretation.

In October 2026, I spoke briefly with Alejandro Vidal Crespo at IE University's Asset & Wealth Management Day in Madrid. Vidal, Head Investment Manager at Deutsche Bank Spain, expressed his personal view that AI could offer less biased investment analysis because it has no personal stake in an investment. This made him more open to delegating analytical work to AI. The conversation prompted me to consider how delegation to AI agents changes workflows and the work people retain.

The customer-service case starts with a manager considering whether automation allows a specialist to move to other work. The family-office case starts with a fixed team opening more investment memos after agents reduce research effort. One reduces the people available; the other increases the ideas they must examine. The first exposes the timing problem in the opening example. The second shows how developing more ideas can increase the human work required after delegation.

Three SDA Bocconi MBA courses shaped my approach: Professor Gianmario Verona's Technology and Innovation Strategy, Professor Marco Ottaviani's Managerial Economics, and Professor Giada Di Stefano's Corporate Strategy. I draw on these courses to examine competence change, opportunity cost, and venture scope and growth. The applications to AI delegation and Mosaic are my own.

1.1. What the Canvas records

The Canvas has four sections, each answering a managerial question:

• Canvas Section A. Mandate and resources: who is authorizing what, for how long and with which people and resources?
• Canvas Section B. Actions and human support: which actions are permitted, under what limits, and what human work do they require?
• Canvas Section C. Capabilities and controls: what can the agent actually do, and how are actions outside its permission blocked or their consequences contained?
• Canvas Section D. Decision and reconsideration: what is the decision, and what would require a pause or a fresh assessment?

The manager uses the four sections to record the proposed arrangement, applies four checks and records the outcome. The first decision is whether the arrangement can be authorized. The manager then compares its value with other feasible choices. Sections 2 and 3 explain the foundations and procedure. Section 4 works through the two cases, and §5 applies the reasoning to management and Mosaic.

2. Foundations and existing approaches

2.1. Automation changes how work is shared

A person can be available in name yet unprepared to intervene in time. Bainbridge (1983) identifies maintained skills, current process knowledge and response time as constraints on human takeover. Feigh and Pritchett (2014) require the full allocation of functions, including teamwork, to be feasible under realistic workloads and time constraints. Expertise, workload peaks and interactions all matter. When several tasks need the same person, checking each task separately cannot establish whether that person can perform them together.

IJtsma et al. (2019) include monitoring, intervention, communication timing and work dependencies in allocating work and interaction in human–robot teams. The Canvas draws on these concerns for its joint feasibility test. Can the assigned work be completed together, in the required sequence, while qualified people are available?

Separating delegated from retained work also matters when interpreting reported outcomes. In a randomized deployment involving 647 Alibaba service workers, AI-eligible chats became faster, but their recorded ratings fell. Ratings on human-retained chats rose. The all-chat rating effect was not statistically significant (Wang et al., 2026, pp. 8, 24–25, Tables 2–4). Only 17% of chats were rated, limiting what the ratings can establish. Elapsed chat duration is also different from worker time saved (Wang et al., 2026, p. 9). The contrast invites managers to examine the work delegated to AI, the duties people retain and the support those duties require.

For its constructed cases, this paper assumes that tasks may all require their full upper-bound effort, unless evidence supports a tighter joint limit. In the staffing case, each of the six specialist cases may take fifteen minutes. The manager must therefore allow for all ninety minutes, even though some cases might finish sooner. Reserving that capacity may leave time unused and forgo other work.

2.2. Assessment effort and evidence selection

Gu et al. (2026) jointly model routing, or how tasks are assigned, and review engagement under finite effort and performance constraints. Their effort budgets are not measured schedules of staff hours. The model motivates checking engagement as well as availability. A name on a roster does not establish that someone will perform the assessment.

Assessment also depends on which evidence the reviewer can examine. Höppner (2026) models the cost of checking selectively presented evidence. In an audit using AI models as reviewers, access to the full, fixed source set improved detection of directional briefs, which steer interpretation through evidence selection. Detection remained imperfect. This working paper motivates retaining the family-office case's full source packets for assessment. It establishes neither human review times nor whether staffed review is adequate. The case stipulates both the assessment duty and its cost.

2.3. Authority, supervision and the Canvas design

Agency theory helps explain who may act for whom. Humberd and Latham (2026) apply it to AI through delegated decision rights and theorize how monitoring changes as capabilities change. Their account informs the Canvas's treatment of authority. The particular checks and recourse requirements are design choices in this paper.

The closest governance comparator is the Agent Capability and Authorization Profile, or ACAP. The World Economic Forum and Capgemini (2026) combine developer technical information with enterprise delegation policy. Their playbook requires a register of consequential actions and an approved authority matrix specifying who may authorize them. Technical enforcement must support those limits, with checks against bypass. The playbook includes signed production authorization, validation, monitoring and reauthorization. It also treats review volume, complexity, expertise, fatigue and agent speed as supervision constraints, and requires operational supervision and tested intervention for deployment readiness.

ACAP already connects human support with permission. The Canvas brings named duties, time windows and their links to individual permissions and controls into one decision record. If required human support falls short, for example because a required person is unavailable, the findings change for every permission and control that depends on that support. The record then shows why authorization must be withheld. Competent ACAP use may supply the same dependencies and reach the same decision.

2.4. Competence change and continuing permission

Anderson and Tushman (1990, p. 609) distinguish competence-destroying change, which makes existing expertise obsolete, from competence-enhancing change, which builds on existing know-how. Professor Verona's course introduced this distinction. I apply it conditionally to one delegation, without assigning AI an industry stage.

A manager considering a staffing reduction must ask which expertise the changed workflow still needs. Delegation can remove the need for some human competences while making others more critical and time-bound. Before treating automated production as a staff saving, the manager must identify and preserve that retained expertise, consistent with Bainbridge's concern about maintained skills and timely intervention (§2.1). In the customer-service case, specialist competence remains necessary within the response window.

In the family-office case, the agent prepares investment ideas, calculations and interpretations. The analyst must then judge the evidence selected and omitted, check the calculations and challenge the interpretation before release. The manager must assign this retained assessment work. Assessing an argument or calculation may still require the expertise used to produce it. The competence lens identifies expertise to retain; the schedule tests whether it is available when required. The case assumes qualified assessment rather than measuring a change in skills.

Delegation may leave fewer production tasks while concentrating demanding judgment in the remaining human work. If teams start more projects, the additional reasoning may consume more time than agents save in research. Filling the day with that work may also displace less demanding activity. The paper calls this theoretical possibility intellectual overload: sustained judgment could become harder as opportunities to turn to lighter work shrink. The family-office case examines this pressure without measuring cognitive overload or its effects. Managers should examine the kinds of judgment retained and the mix of work alongside total hours.

The agent's technical capabilities can stay the same while its human support changes. The work accepted may increase, different expertise may become necessary, or required people may become unavailable. Such changes call for reconsidering the affected permission. Scheduled expiry prompts reassessment for a different reason: assumptions may have become stale even when no change has been detected.

3. Developing the Canvas and demonstrating its use

3.1. Design method and selection boundaries

I developed the Canvas by comparing theory and methods, then applying the proposed rules to two constructed cases. The review began with decision rights and agent governance and expanded to automation, collective allocation, assessment effort and coordination. It focused on English-language primary and methodological work, using selected starting sources and one expansion through allocation references. ACAP was the closest governance comparator. As the review developed, I revisited allocation, innovation, authority and customer-service AI, adding Alibaba's reported contrast between delegated and retained work without changing the allocation requirements. News, product promotion and empirical staffing estimation were excluded. Selection was purposeful, not systematic or exhaustive. A later practitioner-source check informed the fictional investment-committee setting in §4.2.

Canvas Sections A–D organize the information needed for the decision. Four checks assess valid authority, acceptable consequences, effective restrictions and operational support. Table 1 shows how the research concerns inform these checks. Human capacity matters wherever a permission or control depends on human work, so a shortfall can affect several checks.

Table 1 Requirements and their proposed placement

Note. Sources establish concerns; the grouping and cross-check links are the author's design. Human dependencies can affect several findings.

The two constructed cases apply the same authorization rules to different managerial choices: whether to reassign a specialist and how many investment-idea memos a fixed team can assess for its committee. The appendices retain the assumptions and timed assignments needed to examine each decision. The completed sheets state their permissions, support and findings directly.

I used AI to assist literature discovery, drafting and internal review. I remain responsible for the final text, sources and judgments.

3.2. Canvas Section A: mandate and resources

The manager begins by stating the service purpose, scope and available resources. The mandate identifies the principal (the party delegating work) and the ratifier (the person entitled to approve it). In the staffing case, the service director holds both roles. Figure 1 shows the completed Canvas for the staffing request, and Table 2 supplies the timing calculation. Appendices A and B provide the blank sheet and completion guide.

3.3. Canvas Section B: actions and human support

Each consequential action needs an explicit permission: authorized, prohibited or pilot only. Workflow approval does not itself permit every action. Figure 1 summarizes the staffing request; permissions remain provisional until the ratifier signs the assessed scope.

Identify the human work each action needs to keep service promises, respect consequence limits, operate controls or provide recourse, then assign it feasibly. Figure 1 links actions to support. A duty may be omitted if unnecessary or an adequate substitute is established; individual human review is not always required.

A named owner must explain the basis for the volume accepted into the workflow, bounds on required effort, expertise, dependencies and deadlines. The ratifier challenges exclusions and accepts remaining risks. Calling a known shortage or ordinary demand an 'exception' cannot remove it from assessment.

The joint test (§2.1) checks qualified availability, competing commitments, exclusive attention and simultaneous coverage. Specify each reserve's contingencies, capacity, expertise and available windows; count it once. A compact roster and timing calculation suffice if they demonstrate feasible assignments. Upper bounds must fit together unless a tighter joint limit has support.

The manager also needs grounds for expecting people to perform their duties adequately. Check that they have accepted responsibility, have the competence and protected time required, and will receive any necessary supervision. Protected time is reserved for the assigned duties without competing commitments. Pausing the workflow must leave outstanding work and recourse assigned, with arrangements for notice and handoff and any unavoidable loss accepted by the ratifier.

3.4. Canvas Section C: capabilities and controls

The manager next checks what the agent can actually do. This inspection must be independent of the permission list and include alternate routes and uncertainty. Each consequential capability needs a disposition explaining how it is handled. Authorized means it matches explicit permission. Blocked means checked controls prevent the effect within the assessed boundary. Contained means the effect remains possible but checked measures keep consequences within accepted limits. Unresolved means that support is missing and prevents permission.

Section 4.2 illustrates containment through investment-idea memos whose drafts may contain errors but cannot circulate to the committee before assessment and approval.

3.5. Canvas Section D: decision, suspension and restoration

The decision records four checks on the proposed scope. Each starts with a practical question:

1. Who may authorize this work, and is its scope clear? Legitimacy and specification require an entitled decision maker and explicit purpose, actions, destinations, service conditions and dependencies.
2. Are the consequences acceptable, and can affected people obtain correction? Consequence and recourse require justified limits and supported response arrangements, including a route to request review, a responder and a response period.
3. Are the agent's possible actions adequately controlled? Capability-control requires a supported disposition for each known consequential capability. Alternate routes and controls that depend on people must also stay within accepted consequence limits.
4. Can people carry out their duties together, and when must permission change? Operational responsibility and reconsideration require named owners, accepted duties, compatible assignments and checks of competence, protected time and any necessary supervision. Permission needs a justified period and conditions for reconsideration, pause and restoration.

Each check receives a reasoned finding of met or not met; missing support means not met. One shortage may invalidate several checks, and one repair may restore them. The findings are not additive: enough staff cannot confer authority or prove that a restriction works.

Each alternative needs findings for all four checks, either assessed afresh or justified by support that still applies. Passing them makes an option eligible for authorization. The manager must then compare its expected benefits, costs and opportunity costs with those of other feasible arrangements.

The ratifier records one of four outcomes:

• Authorize the stated scope: all four checks must be met for ordinary authorization.
• Pilot only: all four checks must be met within the pilot's own scope. The pilot must define a question it will answer to inform the decision, support the learning duties and set limits on exposure and continuation. Appendix B gives the full requirements.
• Remediate: identify a repair and withhold execution until the repaired scope passes.
• Do not delegate: close the proposal with reasons.

Failed checks prevent permission. The ratifier decides whether to pursue repair or close the proposal.

Permission requires the ratifier's signature on the exact assessed scope, including any necessary attachments. A complete decision may fit on the sheet; complex arrangements may need additional support. An assessment or signed negative decision grants none. Expiry requires fresh assessment and signature; protective denial controls remain in place.

The circuit-breaker rule applies before work starts and throughout the permission period. Before work enters the assessed workflow, the manager must withhold an action if qualified people cannot perform its required human duties in time. During the permission period, suspend affected authority when a change invalidates that support. A separable scope may continue if its own conditions remain met. Adequate human capacity remains one necessary condition wherever actions depend on it; all four checks still apply.

The record must identify what reveals a shortfall, which permissions and controls depend on the lost support, who responds and how quickly. People and technical controls must carry out the restriction or pause before consequences exceed accepted limits. The circuit-breaker is a decision rule, not an automatic enforcement mechanism; detection and response take time.

Suspension leaves work behind. The named owner must preserve notice, handoff, correction and recourse, while identifying any performance that recovery cannot restore. Before resumption, the owner checks the affected support and restoration conditions. A changed or expired scope needs fresh ratification. Stopping further action cannot recover a missed deadline or discharge an existing obligation.

4. Two worked decisions

These constructed cases apply the same procedure to a staffing reduction and an increase in assessment demand. The quantities, controls, qualifications and performance of assigned duties are stipulated for each case. Each assessment covers one batch or committee-preparation cycle.

4.1. Customer-service case: retaining coverage or redeploying staff

The service director must decide whether to keep two equally qualified specialists for a Monday batch or reassign one to another planned delegation during 09:00–10:00. The batch contains up to thirty routine enquiries and six specialist cases. All are present at 09:00, and no later arrivals enter. Keeping both specialists means withdrawing the competing assignment. Resource owners confirm that this leaves no other external service promise unmet. The director gives priority to the batch's response commitment.

All six specialist cases need a response by 10:00. Each requires a qualified specialist's exclusive attention for ten to fifteen active minutes, and all six may take fifteen minutes.

Under the staffing cut, the remaining specialist has 130 minutes reserved for this batch, from 08:50 to 11:00, with no competing work. This is protected time. Preparation takes ten minutes before 09:00. Assurance and coordination, including checks of routine responses and referral outcomes, take up to twenty-five minutes after 10:00, followed by five minutes to close and hand over outstanding work. The full workload fits those 130 minutes. Table 2 checks whether it also fits the deadlines.

Table 2 Customer-service case: aggregate capacity and the specialist response window

Note. The values are assumed limits on active work. Operations and later recourse have separate allocations. Appendix C gives the supporting allocation and completed retained-roster Canvas.

The staffing cut fits total available hours but fails the response window. Its outcome is remediate: retain the second specialist, as Figure 1 records. With two specialists, each handles three cases and finishes by 09:45. Later free time cannot repair the one-person schedule. Nor can the operations owner's spare minutes: she lacks the qualification to decide specialist cases.

The remaining checks concern how the service operates. Checks of both the service interface and direct tool access confirm that messages stay within assigned customer cases and ambiguous enquiries go to a specialist. Routine mistakes are low-consequence and correctable within accepted limits, supporting release without individual human review. Assurance and recourse remain assigned. Customer receipts direct requests for review or correction to the service portal. A separate qualified staff member has time reserved to handle them. That reservation remains available after permission expires and cannot be used for morning specialist work.

With these controls and assignments, the retained roster meets all four checks. Given the director's priority, the assessment recommends authorizing it for this batch. Retaining coverage carries the opportunity cost of the forgone redeployment even with unchanged payroll (Baye & Prince, 2025, p. 5). Comparing financial value would require a separate valuation of the alternatives.

The circuit-breaker rule governs both admission and later disruption. The batch cannot start without required coverage. Unexpected specialist loss after it starts triggers pause, notice and recovery. A response may still be late because stopping further sends cannot replace the missing specialist time; the director accepts that exceptional risk within the recovery arrangements in Appendix C.

The same allocation discipline applies to the operations owner: assigning her to another simultaneous intervention would require a substitute or supported joint arrangement. The retained roster's findings depend on protecting her monitoring time as well as specialist coverage. Appendix C: customer-service staffing gives the supporting allocation and completed Canvas.

4.2. Family-office case: fit the ideas, then deepen the reasoning

The research director expects agents to help the team develop more investment ideas. Instead of starting three memos, the team considers six at once, intending to submit each when it can defend the analysis. Source summaries and initial calculations arrive quickly, and the expansion appears manageable. The analyst still has to resolve conflicting evidence, challenge assumptions and decide whether each investment thesis holds. The director must decide how many ideas the research savings can support while preserving that analysis, then how to use any time left over.

The committee setting draws on WE Family Offices (2026), which describes advance preparation for committee meetings and illustrates deferral pending deeper analysis. Here, each memo is at most two pages, with a retained source packet and calculation record, and supports preliminary discussion of a publicly traded company. Memos develop before agenda submission. Only when the analyst and portfolio manager are confident they can defend the analysis is a memo submitted for circulation and discussion. For this committee, approval takes place on Friday at 16:00, with circulation by 17:00 before a Tuesday 10:00 meeting. The committee retains investment-strategy decisions. Unready or revised memos are deferred; Tuesday follow-up can clarify or withdraw a memo, not approve a late replacement.

For accounting, the comparison separates human research from investment reasoning. Research includes gathering and organizing sources, reading relevant material and checking drafts. Reasoning tests the evidence, assumptions, calculations and downside risks supporting the investment thesis. At the upper bounds, research takes 120 minutes per idea without agents and forty with them, one third as much. Full assessment still requires up to sixty minutes of investment reasoning per memo. The director can reserve additional reasoning time once that requirement is covered.

The same analyst has 540 minutes for research and reasoning across Thursday afternoon and Friday, before approval at 16:00. Setup, circulation approval and immediate correction each add thirty minutes, giving 630 investment-work minutes on Thursday and Friday. A separate sixty-minute period of lighter work remains unavailable for investment analysis; Tuesday's fifteen-minute follow-up is also reserved separately. There are no competing duties, added staff or overtime. All upper bounds may coincide. Appendix D gives the supporting allocation and completed Canvas.

Table 3 Human research and investment reasoning at the upper bounds

Note. The five-idea standard-reasoning row is the intermediate calculation before reallocating spare time. Setup, approval and immediate correction add ninety investment-work minutes to each arrangement. Background lighter-work and Tuesday follow-up remain separate. The human-only row is a workload comparator, not an agent permission decision.

Table 3 shows that six ideas exceed preparation capacity by sixty minutes. Allocating fifty minutes of reasoning per memo would fit the calendar but leave ten required minutes unsupported when the full allowance is needed. Memos must remain defensible; intellectual overload remains a theoretical possibility (§2.4).

The director first scopes five ideas with full assessment, then reserves the forty spare minutes for further challenge, eight per memo. Compared with three agent-assisted ideas, the additional ideas use 200 minutes. Holding other costs fixed, their expected benefits must justify the work displaced; further reasoning must justify its own forty-minute opportunity cost. If those benefits are insufficient, the director should preserve the corresponding time for other work, subject to assessing the revised scope. The director prefers this allocation, without establishing a financial or quality gain.

The supported assignment finishes Friday at 16:00. Its five named ideas are supplied before agent preparation; setup confirms scope and records the sixth idea's deferral, without ranking six researched ideas. After researching all six, completing five memos with sixty-eight reasoning minutes each would require 580 preparation minutes in total. Deferral cannot erase that work.

Source packets remain fixed and complete for each specified preliminary discussion question. Every selected memo needs full qualified reasoning against the retained evidence. Deferral cannot remove analysis necessary for a circulated memo. Release requires assessment and approval of the exact version; edits invalidate both, and a material defect withdraws the approved version. These controls cannot establish that the analyst actually devoted adequate thought to the analysis. The research director accepts the remaining risk after the portfolio manager's review. Separate controls block official changes, client sends, holdings access, trading and production writes; regeneration cannot reset the limits.

The six-idea request fails operational responsibility and reconsideration. Its other three checks remain met under the supported gate, correction and deferral arrangements. Its outcome is remediate: scope five ideas before research, then reserve sixty-eight minutes of reasoning per memo. Both the three-idea agent option and the revised five-idea scope meet all four checks on their own supported assignments. Counts remain ceilings, not agenda quotas, and using the full window is this director's choice. The sixth idea retains an owner and needs fresh checks before later development. Lost required support triggers reconsideration and suspension of affected circulation; correction, recourse and restoration duties continue as specified in Appendix D: investment-committee preparation.

5. Managerial recommendations and conclusion

5.1. Recommendations for managers

1. Specify the service and tolerances before expanding delegation. Identify the promised result, affected people and acceptable consequences. Examine delegated and retained work separately when aggregate outcomes can hide opposing effects, as the Alibaba study illustrates. Specify whose experience counts, how incomplete outcome coverage will be handled and what would stop expansion.

2. Reserve qualified time for the work being accepted, including assessment. Identify the expertise and duties the workflow retains. Distinguish assessment from final approval, record deadlines and competing commitments, and name who owns the workload estimates. Test whether the work fits when ordinary demand reaches its supported bounds. An optimistic average cannot replace that test. If conservative bounds leave capacity unused, compare the value forgone with an alternative whose limits have a sound basis. Any revised service promise needs a fresh assessment, including correction and recourse.

3. Compare feasible alternatives and state what would reverse the choice. Record why the chosen result is worth its resource commitment, including Canvas completion and maintenance. Valuable redeployment cannot make an incompatible schedule feasible. Once alternatives pass the checks, compare the benefits being traded and identify the change in their relative value that would select a different option.

4. Apply the circuit-breaker rule when required support is lost. Name who detects the change, suspends affected authority and checks restoration. Reassess changed volumes, deadlines, evidence mixes, staffing and shared reserves before relying on the old permission. Preserve outstanding duties and identify what pause cannot repair. For repeated workflow use, set admission limits and reserve shared capacity for each scope. Arrangements for recovering interrupted work, recourse and protective controls must remain supported after pause or expiry. Any required periods of lighter work must remain protected. Resumption requires checked support; changed or expired permission requires fresh ratification.

5.2. Mosaic's product and service boundary

The family-office case makes the use of research savings a decision for the customer team. The director first preserves full investment analysis when choosing how many ideas to develop, then assigns the spare time to further reasoning. For Mosaic, I recommend applying that sequence to proposed model changes and valuation questions. Teams should agree on the analytical scope and reserve qualified assessment time before accepting more agent-prepared work. The time released can support more ideas, deeper examination or other work, according to the team's priorities.

Mosaic's product direction keeps analytical judgment with customers. Agents can prepare and evaluate changes, but their own access does not let them save a version. Users review a fixed proposal in the browser. Choosing Save while signed in records their approval and creates a version. Changes to the proposal or its calculation basis require a fresh review before the changed work can be saved.

Customers retain responsibility for deferred work and must reconsider the allocation when its support changes.

Supplying assessment would be a separate service-scope decision, informed by customer research. Drawing on Professor Di Stefano's Corporate Strategy course, I would ask whether combining software and assessment creates value beyond separate provision and whether that benefit requires ownership. Capron and Mitchell (2012) distinguish the strategic rationale for acquisition from the need for control and the ability to integrate a provider. They caution against invoking synergy without a credible integration plan.

Capron (2012) distinguishes build, borrow and buy: develop internally, contract or partner, or acquire a provider. Mosaic would need evidence of demand and workable quality and staffing arrangements before choosing among them. Contracting remains an option where quality and timely capacity can be specified, checked and enforced; internal provision would need a case for development or ownership and its feasibility. Supplying assessment in any mode would make its quality and timely, qualified staffing part of Mosaic's own service obligation.

5.3. Conclusion and further work

The Agent Delegation Canvas makes timely, qualified human support a continuing condition of permission. Its constructed cases show how staffing reductions or workload expansion can invalidate that support; feasible alternatives still require a value judgment. Suspension leaves outstanding duties and cannot restore missed performance.

For Mosaic, the proposed application surrounds customer review and saved-version approval with scope control, qualified assessment time and supported deferral. Neither the cases nor this application establish effectiveness in practice. Evaluation should test whether the Canvas reveals missed dependencies and improves decisions enough to justify its effort compared with competent existing practice. Mosaic's customer research and a service trial could examine analytical scope and the use of research savings.

References

Anderson, P., & Tushman, M. L. (1990). Technological discontinuities and dominant designs: A cyclical model of technological change. Administrative Science Quarterly, 35(4), 604–633. https://doi.org/10.2307/2393511

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8

Baye, M., & Prince, J. (2025). Managerial economics and business strategy (2025 release). McGraw Hill.

Capron, L. (2012, May 3). The corporate growth dilemma: Build, borrow or buy? INSEAD Knowledge. https://knowledge.insead.edu/strategy/corporate-growth-dilemma-build-borrow-or-buy

Capron, L., & Mitchell, W. (2012, December 17). Warning against binge buying. When is M&A the right answer to your growth challenges? INSEAD Knowledge. https://knowledge.insead.edu/strategy/warning-against-binge-buying-when-ma-right-answer-your-growth-challenges

Feigh, K. M., & Pritchett, A. R. (2014). Requirements for effective function allocation: A critical review. Journal of Cognitive Engineering and Decision Making, 8(1), 23–32. https://doi.org/10.1177/1555343413490945

Gu, W., Li, M. L., & Zhu, S. (2026). Should humans be in the loop? Human–AI collaboration paradox and automation cliffs [Working paper]. SSRN. https://doi.org/10.2139/ssrn.6417798

Höppner, M. (2026). Better information, cheaper capture: Personal AI and selective evidence [Working paper]. SSRN. https://doi.org/10.2139/ssrn.7375940

Humberd, B. K., & Latham, S. F. (2026). When AI becomes an agent of the firm: Examining the evolution of AI in organizations through an agency theory lens. Journal of Management Studies, 63(2), 668–694. https://doi.org/10.1111/joms.13274

IJtsma, M., Ma, L. M., Pritchett, A. R., & Feigh, K. M. (2019). Computational methodology for the allocation of work and interaction in human-robot teams. Journal of Cognitive Engineering and Decision Making, 13(4), 221–241. https://doi.org/10.1177/1555343419869484

Mavračić, J. (2025). Policy Cards: Machine-readable runtime governance for autonomous AI agents [Preprint]. arXiv. https://doi.org/10.5281/zenodo.17464706

Wang, Y., Zhu, C., Feng, T., Lu, L. X., & Jia, B. (2026). Agentic AI and human-in-the-loop interventions: Field experimental evidence from Alibaba's customer service operations (Version 2) [Working paper]. arXiv. https://arxiv.org/abs/2605.14830v2

WE Family Offices. (2026, May 14). How do investment committees work in family offices? https://www.wefamilyoffices.com/resource/how-do-investment-committees-work-in-family-offices/

World Economic Forum, & Capgemini. (2026). AI agents in action: A playbook for trusted adoption, authorization and scaling. World Economic Forum. https://www.weforum.org/publications/ai-agents-in-action-a-playbook-for-trusted-adoption-authorization-and-scaling/

Appendices

Appendices A–D contain the blank Canvas, a short completion guide and the assumptions and calculations supporting both cases. The completed canvases state each decision without coded references.