AI-Native Enterprises

AI and the organizational operating model

An operating model arranges work, people, authority, information and technology to deliver outcomes. The Operating Model Canvas makes these arrangements explicit, connecting strategy to everyday operating choices. [1]

AI-native describes deliberate redesign around AI capabilities; it is not a certification or maturity rank. A sociotechnical design approach addresses technology, management practices and human work together. Its principles support discussion among designers, operators and managers, but principles alone do not supply a working organizational design. [2]

Accountability is the obligation to answer for decisions and results. An agent can investigate, prepare changes and return evidence while an accountable owner decides whether the work is justified and sufficiently supported. Automating execution does not remove that obligation. [3]

Connected work, bottlenecks and transferred labor

Here, a value chain means connected activities contributing to a customer or organizational outcome. Lean value-stream mapping more specifically follows the actions and information required for a particular flow, including waste. Both perspectives widen attention beyond one task. Workflow Automation explains why decisions, waiting and organizational exchanges belong inside the process boundary. [4]

A customer-order design example spans sales, fulfillment, billing and support. A requested change affects delivery and charges. Completion requires an accepted decision, correct authoritative records and fulfilled downstream obligations. Faster correspondence alone leaves those dependencies intact. [5]

Business process reengineering reconsiders the process itself. Hammer’s historical Ford account describes replacing invoice reconciliation with shared purchase-order information checked when goods arrived. Purchasing, receiving, payment and supplier practices changed together. This was not an AI deployment: its relevance is the removal of reconciliation work rather than faster execution of the same task. [6]

A bottleneck is a stage whose effective capacity limits the connected process. The Theory of Constraints directs improvement toward that limitation. Faster preparation can simply deliver more work to already-busy reviewers. Effective capacity depends on case difficulty, correction and coordination, so equal case counts need not represent equal workloads. [7][8]

Regulating work release around the limiting stage can prevent upstream acceleration from creating excessive unfinished work. A queue buffers temporary imbalance; sustained arrivals above processing capacity keep increasing backlog unless intake, workload or capacity changes. Once one constraint is relieved, another stage may determine completion. [9][10]

Establish the baseline before changing the flow: eligible requests received, accepted completions, incorrect or reopened changes, intake-to-completion time, handling effort across all teams, and unresolved review work. Keep elapsed time separate from labor: two people working simultaneously contribute two people’s effort without necessarily doubling customer waiting. [11][4]

Organizational knowledge and its owners

A system of record is the designated authoritative source for a particular business fact. An assistant’s statement that an order changed cannot replace the order record. Integration into existing work establishes this distinction. [14]

Tacit knowledge means practical judgment people have not fully written down. Experienced workers may recognize an unusual request that documented rules omit. Their contribution includes explaining that judgment, identifying acceptable outcomes and reviewing proposed guidance—not merely supplying documents to an engineering team. [15]

A data owner answers for meaning and permitted use; a data steward manages everyday quality and correction work. The Data Ownership Model provides this distinction. Local job titles may differ, but both responsibilities must be assigned. [16]

Retrieval finds relevant information; it does not confer authority on the result. Search and Retrieval covers selection mechanisms. Organizational stewardship determines whether a retrieved source is current, applicable and suitable for the decision. [18]

A useful correction procedure preserves conflicting assertions with their sources, routes the disagreement to the authorized owner, records the resolution and identifies dependent guidance or cases requiring review. Publication time, retrieval time and business-effective time remain distinct. A newly retrieved document may describe an old rule. [17]

Publication readiness is a managed judgment. Knowledge-Centered Service, or KCS, distinguishes unresolved, unvalidated and validated service knowledge, with publication privileges tied to demonstrated competence. This supports a review path from recurring exception to reusable guidance without automatically converting an AI answer or local workaround into policy. [19]

Knowledge maintenance can sit inside the operating workflow. Workday Help’s documented capabilities combine drafting and translation with author review, version control, approval workflows and case management. These features illustrate integration between maintained guidance and receiving teams; their availability does not establish a customer’s configuration or improved outcomes. [20]

Purposeful information access

Data governance assigns decision authority, handling rules and evidence obligations. Purpose and accountable data use develops that foundation. Cross-functional sharing should identify the business purpose, information owner, intended user and permitted operation before expanding access. [21]

Authentication establishes identity; authorization determines permitted actions on particular resources. Reading an order does not grant authority to disclose internal notes, approve a fee exception or modify billing. Authorization boundaries explain enforcement. Tool availability and model-generated arguments do not supply permission. [22]

A common knowledge service can span repositories while preserving different permissions. Unauthorized documents must be excluded before entering generation context. Source permissions must survive indexing and document splitting; stale permission copies require correction. Central discovery neither requires unrestricted visibility nor removes source ownership. [23]

Assign an access-request route: workers state purpose; owners authorize use; stewards investigate delays and stale permissions; service operators verify propagation. An unresolved access dispute goes to the designated authority rather than being bypassed by an assistant. [16][23]

Decision rights, redesigned roles and human capacity

Decision rights specify authority to make particular decisions. Advice and execution are different responsibilities. RACI expands to Responsible, doing the work; Accountable, answering for the outcome; Consulted, contributing input; and Informed, receiving updates. The assignment only becomes operational when people accept it and receive the necessary authority and resources. [16]

Naming an agent a manager does not transfer accountability to software. Nor does naming a human approver establish meaningful oversight. Accepted handoffs require receiving capability; substantive review requires evidence, expertise, time and authority. Separation of duties keeps construction and acceptance from relying entirely on the same builder’s judgment. [3][24]

Jobs combine activities, relationships and responsibilities. Automating preparation can expand customer explanation, supervision and exception handling. In Decagon’s account, a combined deployment role split into agent builders handling configuration and agent software engineers turning enterprise needs into product capabilities. That is one specialization pattern, not a staffing ratio or evidence that fewer employees are required. [25]

Automation can leave people with a harder residual task mix while removing the routine practice that sustains their expertise. Bainbridge’s analysis of industrial automation exposes this tension. Applied to AI operations, training should include practical diagnosis and recovery exercises, with experienced substitutes available. Course completion alone does not establish readiness for unusual failures. [28]

Escalate conflicts that exceed the process owner’s mandate: competing departmental priorities, staffing commitments or disputed risk authority. Leadership must return a decision, resource commitment or explicit constraint. AI Engineering Leadership addresses portfolio, staffing and governance choices beyond this operating boundary. [29]

Shared capabilities and authoritative operations

An internal platform supplies shared capabilities with explicit service boundaries. Teams may share model access, information connections and operating controls while retaining different policies. Common capability and differing policy explains this distinction; AI Platform Engineering covers implementation architecture. [30]

Platform owners remain responsible for maintaining shared services, including contributions made easier by agents. Enforced policies should protect hard boundaries; instructions can guide preferred conventions. A platform outage belongs with shared support, while an incorrect cancellation rule belongs with its domain owner. Consumers need an escalation route for failures spanning both. [31][30]

Shared ownership can emerge as recurring needs become clear. Bloomberg’s account favors integrated delivery during early uncertainty, followed by horizontal capabilities such as common policy controls. Centralizing too early can constrain discovery; leaving every team to reinterpret shared obligations creates duplication and inconsistency. [32]

A data contract records a producer–consumer agreement broader than a schema. For the order service, specify shared identifiers, field meanings, quality, acceptable freshness, ownership and support. The Open Data Contract Standard includes these organizational concerns. Integration into existing work explains why valid JSON alone cannot establish agreement. [33]

Receiving teams must accept the work into their own procedures. Billing needs enough evidence to apply the correct charge; fulfillment needs an actionable delivery instruction; support needs reliable completion information. In Workday Help’s product description, contextual case information and configurable service commitments illustrate this connection between knowledge and operational handling. [20]

Operational transition and continuity

Operational readiness means the receiving organization can actually run the process. Agree supported hours, incident and exception owners, trained substitutes, documentation and operating exercises. Sustainable operating ownership describes accepted transfer. Google’s SRE handoff account includes practical instruction and continued development-team support; transferring documents alone is insufficient. [34]

A working prototype leaves this transition work unfinished. In Sanja Grbic’s design-system tracker account, fixes, hosting and deployment took longer than the initial prototype. The example distinguishes getting a tool working from making it available in the environments where people depend on it. [35]

A cutover is the planned switch to the new operating process. Temporary parallel operation needs an explicit end condition. The proposed decision map separates acceptance of new work from responsibility for historical cases, allowing redundant handling to end without abandoning unfinished obligations. [36]

Pausing new activity does not undo committed effects. A delivered shipment or posted charge may need a business-specific correction, not a software rollback. Record unfinished obligations and assign recovery owners. Partial completion and business recovery covers the mechanics; some effects cannot be reversed. [37]

Name who may suspend the AI-assisted route, who directs continuity handling and who authorizes resumption. Resume only after the relevant failure has been addressed and changed behavior evaluated. A permanent duplicate process is an operating cost and responsibility arrangement, not a free precaution. [38][36]

Adoption in everyday work

Change management is structured work that prepares, equips and supports people whose everyday responsibilities change. Prosci’s definition distinguishes this from delivering the technical solution. Adoption means sustained, appropriate use in the intended process. [39]

Access, trial, repeated use and effective integration are different observations. A worker may repeatedly use a tool on unsuitable tasks, or appropriately avoid it when checking costs exceed its help. Investigate relevance, competence, manager support, available assistance and time to inspect consequential outputs. [40]

Protected practice makes changed responsibilities possible. Automattic’s reported experiment paused roadmap work for participating teams, gave small groups ownership and required a real deliverable. Role-specific training and development documentation supported participation. AI use was optional, and training overlapped the experiment, so its outcomes cannot be attributed solely to AI. [41][42]

Participation must include receiving teams, not only tool users. Their extra checking or correction can explain why a locally attractive tool fails to integrate. In the cross-government Copilot trial, existing workloads constrained some training; temporary access and uneven rollout also complicated interpretation. Supplying licenses did not establish workable learning conditions. [43][44]

Incentives and cross-team consequences

Incentives include rewards, penalties, targets and workload expectations. Goodhart’s law names the risk that optimizing a measure weakens its relationship to the intended outcome. A target can remain useful, but its interpretation must account for the behavior it encourages. [45]

Usage mandates can reward superficial compliance. A trivial daily AI task may satisfy an activity target without improving delivery. Similarly, recognizing preparation volume while ignoring receiving-team corrections can reward work transfer. The useful unit is completed service with acceptable quality and total effort. [46][44]

Psychological safety concerns whether people can ask questions, admit mistakes and seek help without threatening their standing. Reporting an AI failure may expose the reporter’s own earlier acceptance of it. Fewer reports therefore need not mean fewer failures; they may reflect a less usable reporting environment. [47]

Recognize knowledge maintenance, responsible escalation and receiving-team work alongside visible output. KCS recommends outcome goals while using activity counts to understand trends. This supports shared service measures without prescribing one compensation system for every organization. [48]

Evidence of organizational improvement

Task correctness, adoption and organizational improvement establish different claims. A correct proposal may remain unused; a frequently used assistant may add checking; a faster workflow may merely move labor elsewhere. Connect system measures to downstream objectives rather than treating activity as the outcome. [49][50]

A counterfactual describes the outcome under an alternative intervention, including no change. Confounding occurs when shared causes influence both intervention exposure and outcomes. [51]

For the order service, compare defined populations and observation windows. Record eligible arrivals, completed obligations, corrections, elapsed time, total handling effort, customer effects and worker burden. Include the work awaiting completion at the end of the window. Missing measurements remain unknown; they cannot establish zero downstream cost. [11][52]

Comparison design must account for changing workload, self-selection and learning time. Spillovers are effects reaching people outside the nominal intervention group: reviewers may absorb extra work, or colleagues may reuse improved guidance. These dependencies complicate team comparisons. Live evidence and causal improvement develops the methods. [53][54]

When tools, training and work allocation change together, evaluate that combined intervention unless the comparison can separate their contributions. Track customer outcomes long enough to capture delayed correction or reopening. A short observation window may capture preparation speed while missing the work needed to make the result useful. [41][55]

Benefits realization and continuing ownership

Benefits realization converts improvement into sustained organizational benefit. Saved minutes, usable capacity, avoided future expenditure and actual spending reductions are distinct. Assign a benefit owner to verify the baseline, costs, allocation decisions and persistence with operations and finance. [52]

Released time becomes usable capacity only when it can be allocated to needed work. Small fragments across many jobs may not cover a specialist queue. More preparation capacity has limited throughput value when approval remains constrained. Reallocation therefore requires both suitable demand and a workable staffing decision. [9][7]

Total cost of ownership includes acquisition and continuing operation, including support and labor. For AI-assisted work, count integration, knowledge maintenance, training, review and transition—even when another team pays. Cost attribution explains allocation. Avoid counting the same shared expense twice. [56]

Inference price is only one cost. Dan Bjornn reports a rebuilt application with higher per-message expense but lower overall maintenance effort. Without a normalized cost breakdown, that account supplies a mechanism to investigate—not a savings estimate transferable to another organization. [57]

Complementary investments are the skills, procedures and organizational capabilities needed alongside technology. Building them consumes resources before their benefits necessarily appear. The Productivity J-Curve explains a possible measurement pattern, not a promise that an unsuccessful project will eventually recover its costs. [55]

Continuing ownership includes changing course. Reassess when service quality deteriorates, costs move elsewhere, the intended work changes or the benefit fails to persist. Revise the design, narrow its scope or retire it while preserving outstanding obligations. AI Engineering Leadership addresses the resulting investment and portfolio decisions. [38]

Open questions

1. 01

   Preserving intervention expertise remains difficult when routine practice disappears. Progress would include repeated operating exercises showing that designated substitutes can diagnose failures and resolve exceptions after sustained automation, rather than relying on training attendance. [28]

2. 02

   Measuring work transferred across departments remains difficult when reviews, corrections and informal coordination use different records. Progress would connect sampled cases across participating teams and reconcile handling effort with accepted completion, including work missing from normal dashboards. [54]

3. 03

   Keeping corrected guidance and permissions consistent across organizational copies remains difficult because source owners and service operators control different stages. Progress would demonstrate a correction or revocation reaching dependent information services and affected cases within an agreed interval. [23][17]

4. 04

   Separating technology effects from training and organizational redesign remains difficult when these changes arrive together. Progress would come from comparisons that identify the combined intervention clearly and test which benefits persist after transition costs and learning periods are included. [55]