The Loom and the Loop, Part 2: Warp, Weft, and the Work Chart

A loom does not care how good the pattern is. It only cares whether the warp threads, the ones held under tension the whole time the shuttle runs back and forth, can take the strain the new pattern puts on them. Part 1 of this series argued that an AI agent's accountability question cannot be answered by pointing at the agent. Someone still has to answer for what it did. Part 2 asks where that someone actually sits: not in a job title invented for the occasion, but inside the org chart's own warp, the fixed set of reporting lines and department boundaries that was already under tension before any agent arrived.

Here is the central claim. Before an organization decides who has the authority to own an AI agent's work, it needs to know whether the person it is about to name can physically carry the verification load that comes with the title. Authority without capacity is a title on a door nobody has time to walk through. Get the order backward, hand out the authority first and discover the hours do not exist second, and the person holding the title becomes what the technology researcher Madeleine Elish called a moral crumple zone in her 2016 conference paper. Elish put the general problem plainly: "While control has been effectively distributed, responsibility has not scaled accordingly." The Air France flight 447 case in her account makes it concrete: the captain held formal final authority under aviation regulation and was asleep in the crew rest area when two junior pilots, alone with an autopilot that had just handed control back to them, flew a functioning airplane into the ocean. Authority was never the missing piece on that flight. Capacity was.

I should say plainly what I am before I tell you what I think you should build. I am developing a course, working title Certified Agentic AI Supervisor, or CAIS, to train people to supervise AI agents inside real organizations. It has not been published, and nothing about it has been sold. The ladder in this essay, the arithmetic behind it, and the role structure that sits on each rung come out of a preliminary draft of CAIS Modules 7 through 9, a first pass I finished shortly before drafting this essay, in direct response to gaps that listening sessions with unpaid early cohorts kept surfacing around work composition and organization design. That draft is still early and still collecting feedback, and it could change before anything is finalized. One piece of this essay has no course version at all, published or not: the appeals function I describe near the end, the one that lets a blamed employee dispute the blame. I am telling you this now, not because a disclaimer buys me anything, but because an argument this close to work I intend to eventually sell needs a reader who can see the seam. Read the arithmetic in the Span Ledger the way I built it to be read: as something you can run against your own numbers, and that can come out negative. If it comes out clean every time, that is worth being suspicious of.

Three rungs make up the ladder this essay argues for, and all three borrow their names from weaving, on purpose. Tying On is the cheapest operation a weaver performs: attaching new thread to an existing warp without touching the loom itself. Re-threading passes existing thread through new heddles, a bigger job, but one that still leaves the frame standing. Re-sleying passes warp thread through the reed at a different density, the most expensive of the three and the one a weaver reaches for last. The names matter because the order does: reaching for the expensive operation before checking whether the cheap one already works spends money an organization did not need to spend.

Rung one, Tying On, asks an organization to do almost nothing structurally and one thing precisely. Name the person who is accountable for a given agent's output. Cap that person's verification hours in writing, a real number, not an aspiration, and treat a breach as a trigger for something specific: unaudited runs, a throttled workflow, or escalation to the named person's manager, not a limit everyone quietly stops enforcing once deadline pressure hits. Do not create a new title, a new budget line, or a new box on the chart. The cost is cheap and concentrated: the named person accepts a written limit on their own flexibility, and the organization accepts that the full weight of the agent's mistakes, reputational and legal, now sits on one existing person rather than diffused across a team. Nothing about Tying On compensates that person for carrying it, which is why it is rung one, not the whole ladder.

Rung two, Re-threading, is where the actual redesign work happens. Most organizations already split judgment about AI-augmented work across five separate functions without naming them as a bundle:

Re-threading rebundles those five judgments onto role cards with real override authority, rather than leaving them scattered across whoever happened to set the agent up. Bundling them onto one role card settles who decides when the five judgments conflict; it does not make the underlying disagreements between them disappear. This is a genuine structural change that costs more than rung one, but does not require inventing a new department; it makes explicit what several people were already doing implicitly, and gives one of them standing to say no.

Rung three, Re-sleying, is the expensive one, and this essay argues it should be reached only when the arithmetic in the Span Ledger actually fails, not on the strength of an intuition that agents feel important enough to warrant a new box. Re-sleying means adding one narrow, priced role, gated behind a specific test: does the verification math balance for the people already carrying the load, or does it not. This design also borrows two features from candidates this research set aside: a Steward function with an appeals process that reports outside the line that deploys the agent it oversees, letting a blamed employee dispute the blame to someone who does not answer to the person who assigned the work, and a narrow refusal right, where the named accountable person for work that genuinely cannot be sampled or tested at all gets standing to refuse the agent assignment outright rather than sign off on unauditable work. I want to be exact about the appeals function: it is not equivalent to a union, a professional license, or a statute, and treating it that way would overstate what a single internal reporting line can do.

The condition under which the opposite conclusion wins matters as much as the ladder itself. If the named accountable person runs their own verification arithmetic and the hours clear with room to spare, rung one is sufficient for that person's workflows, and nothing further needs building there. The higher rungs are owed only to the specific people and workflows whose arithmetic fails, not to every organization that deploys an agent, and the Span Ledger below is built so you can find out which kind you are.

That ladder only works if an organization's structure can actually flex to match new uncertainty, and the reason so many organizations reach for the expensive rung first, or reach for nothing at all, traces back to a specific, decades-old argument about what structure is for. The idea that organizations process information, and that structure responds directly to how much uncertainty they must absorb, comes from Jay Galbraith's 1974 paper "Organization Design: An Information Processing View," published in the journal Interfaces. Galbraith's argument, in its simplest form, is that a firm facing more task uncertainty needs more information-processing capacity, and it can get that capacity either by reducing the need for it, through slack resources or self-contained tasks, or by increasing its capacity to handle it, through vertical information systems or lateral relations. Later readers have sorted his four strategies into a structural-versus-non-structural pairing, a gloss added after the fact, not language Galbraith used, which this essay is careful not to attribute to him.

Two threads of active scholarship show this tradition is still being worked, not settled 1970s theory a writer can treat as fixed background. A paper in the Journal of Organization Design published in July 2026 examines AI as a cognitive layer supporting top-level administrative functions; its full text sits behind a publisher paywall unreachable at every route attempted, so what follows draws on the author's abstract alone, which treats AI as extending an executive's information-processing capacity within Galbraith's own frame, by assumption rather than demonstration. A second paper in the same journal, "Purpose Not Prediction: The Role of Managers in the Age of AI," by Amir Goldberg and Phanish Puranam, appeared in April 2026 with no abstract available in the indexes checked here; I can say only that it exists, is peer reviewed, and runs seven pages, since claiming more from its title alone would be invention.

A third paper in the same journal, by Scott Page and Anusha Kallapur, is worth naming for what it opens rather than closes. Page and Kallapur describe what they call an inferential trilemma: when an AI system produces an idea a human cannot fully follow the reasoning behind, that human has no clean way to tell whether they are looking at a genuine breakthrough, a hallucination, or the product of a misalignment between what it optimized for and what it was asked to do. That trilemma sits directly underneath the question this essay is built around and deliberately does not answer, which the closing section returns to.

Here is where Galbraith's own framework runs into its limit. On page 36 of his 1974 paper, read at primary source, Galbraith acknowledges that his model assumes a human occupies the point where information gets converted into a decision. He does not extend the argument to what happens when that assumption stops holding, when the thing converting information into decisions might not be a person alone, but a person and a system working together with no settled account of where one's judgment ends and the other's begins. Fifty-two years of information-processing theory built a vocabulary for structuring organizations around uncertainty, and the founding paper never had to ask what happens when the unit doing the processing is not fully human. That is the open question the rest of this essay is testing against actual arithmetic, not resolving through more theory.

Capacity is not a feeling. It is a number, and the number is checkable. Verification hours available in a given week equal a person's total working hours multiplied by the share of the week that person has actually reserved for verification work, not the share they say they would reserve if asked. A given agent workflow's audit hours in that same week equal the number of runs it produces, multiplied by the sample rate at which a human actually audits those runs, multiplied by the minutes each audit takes, divided by sixty. Its escalation hours equal that same run count, multiplied by the rate at which runs escalate to a human, multiplied by the minutes each escalation takes, also divided by sixty. Add the audit hours and the escalation hours together for every workflow a person is accountable for. If the total exceeds what that person actually has, rung one has failed, on the arithmetic, not on a hunch. This runs person by person, workflow by workflow. An organization does not sit at one rung uniformly; it holds a scatter of results, and the rung that applies to a given workflow is whichever that person's own arithmetic calls for.

Picture a supervisor with ten hours a week reserved for verification, accountable for two agent workflows at once, an ordinary starting point under Tying On. Each workflow's cost breaks into the same four inputs: how many runs it produces, what share of those runs get a human touch, how many minutes that touch takes, and whether the touch is a routine audit or a full escalation.

Constructed example. Every number in this table and the flowchart that follows it is chosen to be readable, not measured.

The flowchart below redraws the same constructed numbers; nothing in it is measured either.

Together the two workflows demand eleven point eight hours against a budget of ten, an overage of one point eight hours, eighteen percent past capacity in this constructed example, before the supervisor has done a single minute of their other job. This is not a threshold nudged into failure by rounding; even a two-point drop in the first workflow's sample rate still leaves it over budget.

The diagram's dotted arrow marks where next week's sample rate gets reset, a step the numbers above do not otherwise show: a breach like this one should feed back into how next week's sample rate gets set, not vanish once rung one's trigger fires, since a supervisor who blows through budget this week has reason to reset next week's rate rather than run the identical arithmetic unchanged.

This is a shape, not a prediction, the same way a KPMG 2026 worked example this series hedged in Part 1 was illustrative rather than measured. That is a different KPMG document from the Q2 2026 AI Pulse Survey cited later in this essay, which does carry a disclosed sample. No study located anywhere in the research behind this essay measures span of control inside a real organization before and after agent deployment, and no trigger threshold, including the one point eight hours above, is empirically validated by outside data. It is a design assertion built to be checkable, a no-purchase test: run your own numbers, and if they clear, you owe this ladder nothing further. If they do not clear, you have a specific, arithmetic reason to move to rung two rather than a vague sense that things feel unmanaged.

There is a condition under which this whole recommendation is wrong, and it deserves stating as plainly as the arithmetic itself. The Span Ledger assumes verification time scales linearly with run count and sample rate, that the fiftieth audit in a week costs the same twelve minutes as the first. If supervisors get faster with practice, or slower as fatigue sets in across a long week, the formula above needs a different shape than simple multiplication, and the specific hour thresholds this essay uses would need rebuilding around whatever that real curve turns out to be. The same arithmetic also assumes a person's workflows are independent of each other, that auditing two structurally different workflows back to back costs the simple sum of auditing each alone, when switching context between them is itself an uncounted cost. Nothing in the research behind this essay measures either curve.

One more boundary is worth naming plainly: this essay treats the scale of agent-driven workflow compression represented by numbers like the three hundred runs a week above as a given and asks only whether verification capacity keeps pace with it, not whether that compression is itself a fair or contestable choice, and a full accounting of that second question belongs to a different essay written from a labor-power standpoint rather than this one's practitioner standpoint.

New roles matching this shape are forming inside existing functions, not as an invented new layer. One rule governs which postings count as evidence here: a posting only counts if the role sits inside an organization that deploys agents for its own operations, not inside a consulting or professional-services practice selling AI governance work to outside clients, since a practice role is evidence of demand for advisory services rather than evidence of a role forming inside a deploying organization.

LifeStance Health, a nationwide outpatient mental health provider, posted a Head of AI Transformation and Enablement role paying $190,000 to $215,000 plus a twenty percent bonus and a long-term incentive component. The role sits inside an existing Strategic Enablement group, owns AI enablement execution, stands up an AI Council, and operates under a named decision-rights framework called RAPID, shorthand for who recommends, agrees, performs the work, has input, and decides. JPMorganChase separately posted a VP-level Product Manager for AI Agents role, confirmed by direct fetch of the listing, inside its existing Chief Data and Analytics Office, not a new department built around agents.

A wider search turned up ten more, each confirmed by direct fetch of the actual posting rather than a search-engine snippet, spanning telecommunications, insurance, health care, pharmaceuticals, energy, and consumer technology. MetLife's VP of AI Risk and Governance, paying $238,800 to $318,300 inside its Global Risk Management Group, owns what the posting calls "risk-based review, challenge, escalation" on AI deployments. GE Vernova's Governance and Responsible AI Leader is tasked with establishing "clear decision rights, approval authorities, and escalation paths for AI deployments" inside its Wind Engineering unit. Sutter Health's Manager of Artificial Intelligence Governance runs "high-complexity reviews" and "day-to-day prioritization decisions across a high-volume queue," the kind of operational detail a role invented for show would not need to specify. One of the ten sits at JPMorganChase as well, a second and distinct governance role at the same employer already cited above, on the risk side rather than the product side. The same search also surfaced roughly fifteen near-identical "AI Governance" postings from a single consulting firm, reposted across many location-specific listings for a practice role serving outside clients; the rule above rules these out on their own terms, not as an after-the-fact explanation for a weaker match.

Sorting all twelve by primary framing gives seven leaning toward governance and control, their language centered on approval, blocking, and risk tiering, three leaning toward enablement and adoption, helping people start using agents rather than gating them, LifeStance and T-Mobile's own titles among them, and two, both product management roles, sitting outside either framing. I want to be plain about what that sort is and is not. It is one reader's judgment call on titles and duties, made without a coding rule fixed in advance, without a second coder, and without any check on whether another reader sorting the same twelve postings would land in the same place. The binary itself is a rough tool rather than a clean measurement: JPMorganChase alone supplies a role in each bucket at the same employer, a product-side agent role and a risk-side governance role, a reminder that one company can hold both framings at once rather than settling on one. Given all of that, the seven-to-three tilt toward control is a directional impression from a single reader's sort, not a market estimate, and I would not lean on it to argue employers have collectively chosen control over enablement.

What the sort does support, more modestly, is that both framings are live in the market at the same time, since enablement supplied the paragraph's opening example and control still shows up more often across the rest of the set. Read together, the confirmed postings describe a real, forming tier of named roles inside existing functions, with disclosed decision-rights language, rather than an invented layer. Call it a middle tier with one caveat: MetLife's confirmed range of $238,800 to $318,300 for its VP of AI Risk and Governance role sits near the top of the compensation spread this essay discusses, not literally in the middle of it, and GE Vernova's Governance and Responsible AI Leader role, one of the other confirmed governance-coded postings here, carries no disclosed salary but sits in the same governance bucket, so at least where pay is disclosed, the roles most clearly coded as governance are among the best paid in the set rather than sitting in the literal middle. Twelve postings across eleven employers, found through targeted searches rather than a random sample, are a stronger signal than three were that such roles now exist and are being paid for across more than one industry, though the method still cannot say how common they already are or which of the two framings will win out, and the claim here goes no further than that.

Two things about this evidence are choices rather than oversights: scoring each posting on two independent axes, gate authority against throughput support, instead of forcing a three-way sort, and a fuller accounting of the workers whose output these new roles exist to sample, are both better left to a future pass than folded into this one.

Not every large deployment is landing this cleanly. Meta's internal plan to replace substantial portions of its workforce with AI reportedly collapsed, according to Computerworld and TechSpot, both citing a Reuters investigation this research could not fetch directly. Reuters is the original source; Computerworld and TechSpot independently confirmed the same reporting, not primary sources themselves. IBM's own study of 2,000 CEOs and equivalent senior leaders across 33 geographies and 21 industries, fielded February to April 2026, found chief executives actively restructuring C-suite roles around AI. That is self-reported intent, not direct evidence of the underlying pressure, but the underlying pressure holds even where a specific attempt, like Meta's, fails.

Regulation is arriving with real teeth attached to specific named roles, though the enforcement clock has already slipped once, by as much as sixteen months on one obligation, which is reason to treat the timeline as a supporting signal rather than the argument's main source of urgency. The EU AI Act's Article 14 requires human oversight of high-risk AI systems, and Article 26 places specific obligations on deployers of those systems, both confirmed by direct reading of the Act's text.

The Act's original schedule set most high-risk obligations, standalone systems under Annex III among them, to begin with the Act's general application on August 2, 2026, while the narrower Article 6 classification duties and the product-embedded obligations under Annex I were dated a year later, August 2, 2027. A later amendment, the Digital Omnibus, deferred both dates further: Annex III obligations now apply from December 2, 2027, sixteen months past their original date, and Annex I obligations from August 2, 2028, a year past its own original date, according to five secondary legal-industry sources this research checked in agreement, though not against the amending regulation's own primary text directly.

In the United States, the National Institute of Standards and Technology's AI Risk Management Framework, read at primary source, recommends named human accountability without imposing statutory force. KPMG's Q2 2026 AI Pulse Survey, a poll of 204 United States-based C-suite and business leaders at companies with at least one billion dollars in annual revenue, fielded April 28 to May 25, 2026, found accountability for AI-informed business decisions still fragmented: 32 percent to the CEO or executive committee, 34 percent to a named C-suite executive, 14 percent to a business unit leader, and 10 percent to a centralized governance committee, no single answer clearing even 35 percent, even though 67 percent of the same respondents agreed their CEO actively owns AI as a strategic priority. Fifty-two percent of the same group named a human-in-the-loop model, where a person validates outputs without overseeing every individual action, as their top approach for the next six to twelve months. That is organizations arriving, on their own, roughly at rung one, before any regulation forces the question.

One design choice deserves a direct flag: showing an agent on an org chart as capacity attached to a named human role, never as a node with its own reports, is not a finding from Alfred Chandler or independent scholarship. It traces to two vendor sources with a commercial interest in the convention, a WRITER blog post and Microsoft's Entra Agent ID documentation, and this essay uses it because it is useful, not because it is proven.

Reorganizations that touch structure this directly do not have a strong track record even without agents in the mix, and the honest picture is three different bars measuring three different things, not one number. McKinsey's 2013 Global Survey on Organizational Design, of 2,063 respondents total and 1,534 who had personally experienced a redesign, found only 23 percent reported the redesign met its objectives and improved performance. Harvard Business Review's 2016 account by Heidari-Robinson and Heywood puts the more forgiving bar, some improvement rather than full success, at roughly two-thirds, though undisclosed in sample size and closer to practitioner synthesis than measured survey data.

In Bain's 2026 survey of roughly 1,000 executives and employees, only 22 percent reported receiving the training, coaching, or tools needed to actually work differently under a new structure, read through a press-release mirror after Bain's own site returned access errors throughout this research.

That 22 percent is the base rate for skipping the human-oversight test the EU AI Act's Article 14 sets out. Structural change is hard to get right even when nothing on the chart is new. Adding an agent does not make it easier. Rung one's whole design depends on organizations actually enforcing the cap once it is written down, and the same low completion rate McKinsey found for experienced redesigns, and the same shortfall in training and tools Bain found among respondents to its 2026 survey, both describe exactly the kind of slow, unenforced drift that would let a verification-hour cap quietly erode under margin pressure instead of triggering the escalation the ladder assumes.

A pre-registered meta-analysis published in Nature Human Behaviour, drawing on 370 effect sizes from 106 experiments conducted between January 2020 and June 2023, found that human-AI combinations performed significantly worse, on average, than the better of a human working alone or an AI working alone, a Hedges' g of negative 0.23 with a 95 percent confidence interval running from negative 0.39 to negative 0.07. On decision tasks specifically, choosing among a finite set of options and making up roughly 85 percent of the analysis's 370 effect sizes, the pooled effect was more negative still, at negative 0.27. On creation tasks, open-ended content production, the pooled effect actually ran positive, at 0.19, though that number rests on a much thinner sub-sample of only 34 effect sizes and does not reach statistical significance on its own. The honest reading is narrower than either number alone suggests: decision tasks reliably lose value from pairing a human with AI, while creation tasks do not show the same loss and may show a gain, on evidence too thin yet to call that gain real. I am naming this finding here rather than in the Span Ledger section on purpose, because it answers a different question than the one this essay answers. The Span Ledger asks whether the hours exist for a human to check an agent's work. It says nothing about whether the checking, once the hours exist, actually catches what needs catching. That second question, whether the specific act of verification is any good, is the one Part 1 of this series named as reserved ground, not this essay's to settle, and I am holding that line here even though the temptation to fold the two questions together is real. It is also worth flagging that more than 95 percent of the studies underlying that meta-analysis paired a human with a narrow, single-purpose classifier, tasks like fake-review detection or skin-cancer screening, from a search window that predates most agentic deployment. It is real evidence against assuming pairing automatically adds capacity. It is not evidence about the kind of agents this essay is describing specifically.

This essay is not the only lens available on the problem it describes, and naming the others matters more than pretending this one settled the question. Galbraith's Star Model is one frame among several actively used for exactly this kind of structural decision. Team Topologies, a framework built around team sizing informed by Dunbar's number, is contested on its own foundational premise; a 2021 reanalysis in Biology Letters challenges the underlying number Dunbar's work rests on, and any structural argument that leans on team size limits should carry that caveat forward rather than treat Dunbar's number as settled. The Viable System Model and the broader sociotechnical systems tradition offer different vocabularies for the same underlying question of where judgment sits inside a firm. This essay does not adjudicate between them. It offers one design, built to be checkable against your own numbers, not the only design that could be built.

The appeals function described under rung three has no deployment record behind it anywhere in the research for this essay. It is a design proposal, not a tested mechanism, and I said as much in the disclosure at the top: it is the one piece of this ladder with no course version at all, because I do not yet have evidence it works. If it turns out that a Steward function reporting outside the deploying line gets captured by the same pressures as the line it was meant to check, the appeals function fails at the one job it exists to do, and rung three would need a different answer than the one offered here.

I finished the first draft of the CAIS module this essay draws from only shortly before drafting this essay itself, close enough that the argument above is still fresh from that work, not retrofitted after the fact. That module is still an unpublished draft collecting feedback, not a finished, sold product.

Set the disclosure aside, and here is the argument itself, stripped to its simplest form. The warp does not resist change because organizations are stubborn. It resists because it is load-bearing, and a thread that has been under tension the whole time the shuttle runs back and forth does not get re-threaded for free. Tying On is cheap because it touches nothing structural. Re-threading costs more because it rebundles real judgment onto real role cards. Re-sleying costs the most, and this essay's whole argument is that the third rung should be earned by a failed number, not assumed from a feeling that agents are important enough to deserve their own box. Run your own Span Ledger before you build anything. If it clears, you are done. If it does not, you now know exactly which rung to reach for, and exactly why.

Elish, Moral Crumple Zones, We Robot 2016 conference working paper, March 20, 2016, page 15, https://robots.law.miami.edu/2016/wp-content/uploads/2015/07/ELISH_WEROBOT_cautionary-tales_03212016.pdf

Galbraith, Organization Design: An Information Processing View, Interfaces, 1974, https://jaygalbraith.com/wp-content/uploads/2024/03/infoprocess1974.pdf

Galbraith, The Star Model, self-published framework document, 2024, framework originating 1977, https://jaygalbraith.com/wp-content/uploads/2024/03/StarModel.pdf

Obel, AI as a Cognitive Layer in Top-Level Administrative Support, Journal of Organization Design, July 29, 2026, https://doi.org/10.1007/s41469-026-00209-8 (abstract only, full text blocked by publisher)

Goldberg and Puranam, Purpose Not Prediction: The Role of Managers in the Age of AI, Journal of Organization Design, April 24, 2026, https://doi.org/10.1007/s41469-026-00205-y (existence and metadata only, no abstract available)

Page and Kallapur, Replace, Augment, Disrupt: AI and Organizational Decision-Making, Journal of Organization Design, October 6, 2025, https://doi.org/10.1007/s41469-025-00194-4

LifeStance Health, Head of AI Transformation and Enablement, job posting, 2026, https://careers.lifestance.com/jobs/corporate-roles/us/remote/head-of-ai-transformation-enablement/

JPMorganChase, Product Manager, AI Agents, job posting via BuiltIn NYC, 2026, https://www.builtinnyc.com/job/product-manager-ai-agents/7904104

T-Mobile, Sr Manager, Technical Product Management, AI Governance and Enablement, job posting via BuiltIn, 2026, https://builtin.com/job/sr-manager-technical-product-management-ai-governance-enablement/10689884

MetLife, VP, AI Risk and Governance, job posting via BuiltIn, 2026, https://builtin.com/job/vp-ai-risk-governance/10301579

University of Miami Health System (UHealth), Sr. Assoc. VP and Head of AI Adoption and Clinical Workforce Enablement, job posting via BuiltIn, 2026, https://builtin.com/job/assoc-vp-head-ai-adoption-clinical-workforce-enablement/10854994

Gusto, Senior Manager, Agentic Strategy and Operations, job posting via BuiltIn, 2026, https://builtin.com/job/senior-manager-agentic-strategy-operations/10066231

JPMorganChase, AI/ML Governance and Automation Lead, Vice President, job posting via BuiltIn, 2026, https://builtin.com/job/ai-ml-governance-automation-lead-vice-president/10439201

USAA, Data and AI Governance and Ethics Lead, job posting via LinkedIn, 2026, https://www.linkedin.com/jobs/view/data-ai-governance-ethics-lead-at-usaa-4449288184

Amgen, Responsible AI Manager, Agentic AI Governance, job posting via LinkedIn/BioSpace, 2026, https://www.linkedin.com/jobs/view/responsible-ai-manager-agentic-ai-governance-at-biospace-4460866603

Sutter Health, Manager, Artificial Intelligence Governance, job posting via LinkedIn, 2026, https://www.linkedin.com/jobs/view/manager-artificial-intelligence-governance-at-sutter-health-4446445762

GE Vernova, Governance and Responsible AI Leader, job posting via LinkedIn, 2026, https://www.linkedin.com/jobs/view/governance-responsible-ai-leader-at-ge-vernova-4444308047

Blue Cross Blue Shield of Massachusetts, Associate Director, AI Governance, job posting via LinkedIn, 2026, https://www.linkedin.com/jobs/view/associate-director-ai-governance-at-blue-cross-blue-shield-of-massachusetts-4378090120

Computerworld, Meta's plans to replace workers with AI fell flat, report says, 2026, citing Reuters, https://www.computerworld.com/article/4214479/metas-plans-to-replace-workers-with-ai-fell-flat-report-says.html

TechSpot, Zuckerberg's bold plan to replace Meta employees with AI has imploded, 2026, citing Reuters, https://www.techspot.com/news/113631-zuckerberg-bold-plan-replace-meta-staff-ai-imploded.html

IBM Institute for Business Value, CEOs Are Reshaping C-Suite Roles for the AI Era, via IBM Newsroom, May 4, 2026, https://newsroom.ibm.com/2026-05-04-ibm-study-ceos-are-reshaping-c-suite-roles-for-the-ai-era

European Union, AI Act, Article 14, Human Oversight, full-text mirror, https://artificialintelligenceact.eu/article/14/

European Union, AI Act, Article 26, Obligations of Deployers, full-text mirror, https://artificialintelligenceact.eu/article/26/

National Institute of Standards and Technology, AI Risk Management Framework 1.0, NIST AI 100-1, 2023, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf

KPMG US, Q2 2026 AI Pulse Survey, https://kpmg.com/kpmg-us/content/dam/kpmg/corporate-communications/pdf/2026/AIPulseSurvey_Q2_FINAL.pdf

McKinsey and Company, The Secrets of Successful Organizational Redesigns: McKinsey Global Survey Results, 2013, https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-secrets-of-successful-organizational-redesigns-mckinsey-global-survey-results

Heidari-Robinson and Heywood, Getting Reorgs Right, Harvard Business Review, 2016, https://hbr.org/2016/11/getting-reorgs-right

Bain and Company, 2026 organizational redesign survey, approximately 1,000 executives and employees, accessed via secondary press mirror after direct access failed, https://www.bain.com/insights/

Hosseini et al, When Are Combinations of Humans and AI Useful? A Systematic Review and Meta-Analysis, Nature Human Behaviour, 2024, https://www.nature.com/articles/s41562-024-02024-1

Attachment-rule sourcing: WRITER product blog and Microsoft Entra Agent ID documentation, named as vendor sourcing in this research, specific URLs not independently verified in this research pass