The Second WorkforceA blueprint for the org chart when agents report for work

Opening

The experiment that lasted three days

On July 9, 2024, an HR technology company called Lattice announced that it had made history: AI agents would receive official employee records on its platform, alongside humans. Digital workers would be onboarded, trained, assigned goals and performance metrics, given appropriate systems access, and, remarkably, an accountable manager, a slot on the org chart. The backlash was immediate and withering. Commenters called it an insult to the humanity of real employees; HR professionals asked why a company would formalize software as staff while the industry’s human problems went unsolved. On July 12, three days after making history, Lattice unmade it. “This innovation sparked a lot of conversation and questions that have no clear answers yet,” the CEO conceded, and the feature was withdrawn.

It is tempting to file the episode as a cautionary tale about anthropomorphizing software, and it is partly that: giving a language model a personnel file mistakes the metaphor for the mechanism, and the market punished the category error within seventy-two hours. But dismissing Lattice entirely would be the second mistake, because look again at what the announcement actually contained: onboarding, goals, performance metrics, systems access, an accountable manager. Strip away the word “employee” and every single item on that list is something an enterprise running thousands of agents genuinely needs and mostly does not have. The instinct was early and the packaging was wrong. The questions were exactly right, and two years later they have stopped being philosophical. They are operational, they are unassigned, and in most enterprises they are quietly compounding.

So run the thought experiment this edition is built on. Picture your enterprise a short distance into the future: 30,000 employees and 10,000 agents. Not chatbots; agents, software that takes actions, touches systems, spends money within limits, and hands work to other agents. Now hold up your org chart, the document that has answered “who does what and who reports to whom” since the railroads invented it, and ask it the questions the next decade will ask: Where are the agents on this chart? Who deployed them, and who can retire them? What is each one allowed to do alone? Can one agent direct another, and under whose oversight? What happened to the managers whose teams are now one-third machine? Who reviews an agent’s performance? And when an agent makes a call at 2 a.m. that costs money or reputation, whose name is on the outcome? The chart has no row, no column and no convention for any of it. That is the gap this edition closes.

Exhibit 1The Org Chart, Then and From Now On

The Break

Why the org chart breaks, precisely

To fix the chart you first have to be precise about why it fails, because the failure is structural, not cosmetic. The organization chart is a technology, one of the most successful management technologies ever shipped, and like all technologies it encodes assumptions. It assumes each box is a person, so capacity is measured in headcount. It assumes work arrives through reporting lines, so authority and information flow together. It assumes a box costs roughly one salary, exists in one place, works one shift, and learns at human speed. Every planning ritual built on top, budgeting by headcount, spans of control, succession, performance cycles, inherits those assumptions.

Agents violate every one of them at once. An agent is not one unit of capacity; it can be cloned to a thousand instances by Friday and reduced to zero on Monday, which breaks headcount as a planning unit. It costs pennies per task rather than a salary, which breaks budget-by-box. It works every shift in every region simultaneously, which breaks the geography of the chart. It can be rolled back, versioned, and A/B tested, which no human role permits. And, most disruptive of all, it carries delegated authority: real permission to act on systems of record, granted the moment someone wired it to production, whether or not anyone with accountability actually signed that grant. Microsoft’s Work Trend Index, surveying 31,000 workers across 31 countries, found 82 percent of leaders calling this a pivotal year to rethink core operations and 81 percent expecting agents moderately or extensively integrated within 12 to 18 months; its researchers now describe traditional org charts giving way to fluid “work charts” assembled around outcomes. The direction is not in dispute. The management model underneath it is simply missing.

And the missing model has a cost that compounds quietly. Edition 04 documented where unmanaged intelligence leads: roughly a fifth of AI-related breaches already involve unsanctioned, shadow deployments nobody centrally tracked. The same 2026 research that celebrates the frontier finds the readiness gap runs top-down, 67 percent of leaders familiar with agents against 40 percent of employees, and concludes that organizational factors, culture, management, structure, drive roughly twice the AI impact of individual behavior. Translated out of survey language: the constraint is no longer the technology or even the people. It is the absence of an organizational design for the machine layer. What follows is that design.

Exhibit 2Seven Questions Your Org Chart Cannot Answer

The Distinction

Using AI is not employing it

Before the model, one distinction, because it separates the enterprises this edition can help from the ones that will merely feel helped. Most large companies today can truthfully say “we use AI everywhere”: copilots in the documents, chat in the browser, autocomplete in the code. That is the AI-assisted organization, and it is genuinely valuable, and it changes almost nothing structural. The chart is identical; every box is still a person; each person is somewhat faster at the job they already had. The gains are real, personal, and capped, the well-documented 10 to 30 percent of individual productivity, and the metric that gets reported to the board, adoption percentage, measures enthusiasm rather than capacity. Knowledge stays trapped in private chat histories. Accountability belongs to whoever happened to click. When the license renewal comes up, the entire capability can be switched off, which is the surest sign it was never organizational.

The agentic organization is a different species that happens to run on the same models. Intelligence is not a feature inside each person’s tools; it is a workforce on the chart: registered agents with named owners, tiered authority, scorecards and a lifecycle, doing work rather than assisting it. What improves is not each person’s speed but capacity itself, decoupled from headcount. Knowledge compounds in the registry instead of evaporating in chats. The reported metrics are cost per outcome and the human-agent ratio, numbers a board can steer by. The tell that distinguishes the two takes one question: who owns agent number forty-seven? The AI-assisted organization does not understand the question. The agentic organization answers with a name. The same vendors happily power both, which is precisely why the difference cannot be bought, only built, and why the rest of this edition is the build.

Exhibit 3The AI-Assisted Firm vs the Agentic Firm

The Model

The Machine Layer model

Here is the central design decision, and it is the one Lattice got wrong under pressure and most enterprises are still getting wrong by default: agents do not belong inside the human hierarchy as pseudo-employees, and they do not belong outside it as unmanaged tools. They belong in a distinct machine layer: a formally chartered stratum of the organization that sits inside the org chart’s authority structure without pretending to personhood. The model has five planes, and the whole of it fits on one page.

At the top, nothing changes and everything sharpens. The board owns the risk appetite for delegated machine authority, exactly as it owns credit risk appetite or safety tolerances; this is not a metaphor, it is a policy document the board should be asked to approve. The CEO and executive team own the human-agent operating model itself: the decision of which work belongs to people, which to machines, and which to both, the question Microsoft’s researchers reduce to the human-agent ratio. Human leadership below them sets goals and owns outcomes, including the outcomes of every agent in their domain, a sentence that should appear, verbatim, in leadership role descriptions.

Beneath leadership sits the layer where the real redesign happens: the hybrid operating layer, teams composed of people and agents working the same workflows. The people judge, design, and handle exceptions; the agents execute, draft, monitor, and retrieve. The composition is deliberate and tuned, a ratio someone owns, not an accident of whoever installed what. And at the bottom, the layer that did not exist before: autonomous workflows, where tier-three agents (defined below) act alone inside hard bounds, budgets, and rollback plans. Spanning all five planes, drawn deliberately as a plane and not a box, is governance: the agent registry, the authority tiers, the audit trail, the scorecards, and a kill switch that is tested rather than assumed. Everything Edition 04 argued about securing the intelligence layer applies here with full force; the machine layer is that argument given an organizational address.

Exhibit 4The Machine Layer Model

The Lifecycle

Agent HR: the lifecycle nobody built

Human resources is, at its core, a lifecycle discipline: define the role, hire, onboard, supervise, review, promote, exit. The reason the Lattice instinct resonated even as its framing failed is that agents genuinely need the same lifecycle discipline, run by operations rather than HR, with the anthropomorphism stripped out. Seven stages, each with a named owner, none optional.

Define the role before anything is built or bought: the task, the scope, the systems and data it will touch, the human alternative it augments or replaces. Deploy is hiring, and the rule that prevents most downstream damage is brutally simple: no agent goes live before a human owner is named in the registry. Onboard means what it means for people, translated: context (the knowledge and data it needs), access (least-privilege credentials, exactly as Edition 04 prescribed), and guardrails (its authority tier, its budgets, its escalation paths). Supervise covers the probation period: early output reviewed by humans at high sampling rates before trust is dialed up. Review is the scorecard, on a cadence, covered below. Expand is promotion: authority tiers are earned by track record and signed by a named human, never assumed by default. And retire is the exit interview nobody performs on software: decommissioning, access revocation, and a log entry, because an orphaned agent with live credentials is precisely how a fifth of breaches now begin.

Exhibit 5Agent HR: The Seven-Stage Lifecycle

The Authority Question

The authority ladder: what an agent may do alone

Of the seven questions in Exhibit 2, the one with the sharpest edge is authority, because it is the one that converts silently into incidents. Every agent in your enterprise already has an authority level; the only question is whether it was designed or defaulted. The blueprint answer is a ladder with four rungs and a ceiling, coarse enough to govern, fine enough to matter.

Tier 0, observe: read-only agents that summarize, monitor and report; a team lead can grant this. Tier 1, draft: agents whose output always lands in front of a human before it goes anywhere; a function head signs. Tier 2, act with review: agents that execute, with a human approving before effect; function head plus risk sign, because from here the agent touches systems of record. Tier 3, act alone within bounds: autonomous execution inside hard limits, spend caps, rate limits, rollback plans; this is an executive and CISO signature, and it should feel like one. Above the ladder sits the never list, the enterprise’s constitutional clause, set as board policy: legal commitments, safety-critical calls, and decisions about people, hiring, firing, discipline, compensation, are never delegated to the machine layer, at any tier, regardless of capability. Not because a model could not draft them, but because accountability for them is the one thing an organization cannot delegate and remain an organization.

Two design notes from the field. First, promotion between tiers must be earned and signed, exactly like the lifecycle’s expand stage; an agent’s track record is auditable in a way no employee’s is, so use it. Second, this ladder answers the question that unsettles every executive who first hears it, can an agent manage another agent? Yes, orchestrator agents already delegate to sub-agents, and the ladder makes the arrangement governable: an orchestrator holds a tier like any other agent, its sub-agents cannot exceed the authority of the orchestrator that directs them, and the human owner of the orchestrator answers for the whole tree. Machine-to-machine delegation is fine. Machine-to-machine delegation that terminates in no human name is not.

Exhibit 6The Orchestrator Tree

Exhibit 7The Authority Ladder

The Accountability Question

Who answers for what

Underneath authority sits the question this entire edition, and frankly this entire series, keeps arriving at: when the agent acts, who answers? The machine layer’s rule is single, simple, and absolutely non-negotiable: every agent has exactly one named human owner, and the owner answers for the agent’s actions the way a manager answers for a team’s outcomes. Not a committee, not a cost center, not “IT”: a name. The chain around that rule has four links. The agent does the work. The owner answers for it. The orchestrator, the redefined middle manager we meet next, watches the fleet, samples output, and handles what escalates. And governance audits the system itself: the registry, the logs, the tiers, the reviews.

Notice what this chain does to the org chart’s oldest deliverable. The classic chart is a map of task allocation: who does what. In a hybrid enterprise that map still matters, but it is no longer the load-bearing one, because “who does what” increasingly answers “an agent does.” The map that now carries the weight is who answers for what: the accountability overlay that assigns every agent, every autonomous workflow, and every machine decision to a human name. Regulators are converging on the same demand from the outside in, the EU AI Act’s human-oversight provisions being only the most explicit; enterprises that build the accountability map now will find compliance a byproduct rather than a scramble.

Exhibit 8The Accountability Chain

The Middle

What actually happens to middle management

Now the question everyone asks with a career in mind, and it deserves a more honest answer than either the doom take or the reassurance take. Middle management, as a layer, exists to do two different jobs that a century of org charts happened to bundle: relaying, passing information, assignments and status up and down the hierarchy, and judging, designing how work is delegated, deciding exceptions, and developing people. The machine layer unbundles them ruthlessly. The relay job is precisely what agents do supremely well, at zero marginal cost, around the clock; that portion of the middle is exposed, and no amount of sentiment will protect it. The judging job becomes more valuable, because a hybrid team multiplies the number of delegation decisions, exception calls, and design choices that need a human with context and accountability.

Concretely, the manager of eight people becomes the orchestrator of six people and forty agents, and the job description rewrites itself line by line. Task assigner becomes workflow designer: deciding what gets delegated, to a person or to a machine, the single highest-leverage decision in the hybrid enterprise. Approver of everything becomes exception handler: touching only what the system escalates, which is why escalation quality appears on the agent scorecard. Reviewer of all output becomes sampler and auditor: statistical inspection, not exhaustive reading. Progress chaser becomes ratio owner: tuning the human-agent mix as the work evolves, the operating decision Microsoft’s researchers put at the center of the frontier firm. And coach of people remains coach of people, the one line of the old job description that transfers untouched, because the six humans on the team need development, judgment and career sense more than ever, and no agent supplies it. Microsoft calls the emerging role the “agent boss” and reports leaders already expecting agent management inside their scope within five years. The title will not survive; the job will. Enterprises should start writing it into role descriptions, grading structures and promotion criteria now, because the alternative is discovering in 2028 that an entire management layer was rebuilt informally, without design, by whoever happened to be standing there.

Exhibit 9Middle Management, Redefined

The Talent

The roles this organization hires

Redefining the middle is only half the talent story; the machine layer also creates roles that do not exist on today’s chart, and enterprises that wait for job boards to standardize the titles will staff them two years late. Six missions recur in every serious deployment. The orchestrator, already met, runs the hybrid team and owns its ratio. The agent product manager treats a domain’s agents as a product line: what gets built, what gets promoted up the authority ladder, what gets retired. The evaluation engineer builds the tests agents must pass before deployment and promotion, the machine layer’s equivalent of a hiring bar. The context engineer owns what agents know: the knowledge, data and tool pipelines that determine whether a fleet is grounded or guessing. The fleet reliability engineer owns uptime, rollback and the kill-switch drills. And the AI governance lead runs the registry, the tiers and the audit cadence, the role regulators will effectively require whether or not it appears on the chart.

Notice what unites them: none is a coding role, and none is replaceable by the agents they manage, because each sits exactly where the ladder keeps humans, at judgment, design and accountability. The underlying skills, judgment under ambiguity, delegation design, statistical sampling, agent literacy, are trainable, which is what makes the reskilling path in the change plan real rather than rhetorical. The talent architecture of the hybrid enterprise, which roles to hire versus grow, how career ladders and compensation adapt, and what this means for the graduate entering the workforce, deserves its own edition, and it will get one.

Exhibit 10The Roles the Machine Layer Creates

The Measurement Question

The agent scorecard: performance review without personhood

The Lattice backlash centered on giving agents performance metrics like employees, and the criticism was half right: agents should not get employee reviews. They should get something stricter, because unlike employees, their every action is loggable, samplable and comparable. The scorecard has six metrics, each answering a management question the org chart used to answer with intuition. Task success rate: does it do the job to standard? Escalation rate: does it know what it does not know, the single most underrated agent virtue? Error and drift trend: is quality moving, and in which direction, against its own baseline? Cost per outcome: fully loaded, against the human alternative and against last quarter, the number that connects this edition to the agent-productivity metric Edition 05 placed on the enterprise’s new scorecard. Incident count: boundary violations, near misses, rollbacks, feeding straight into the governance plane. And utilization: because a fleet of shelfware agents is a cost line pretending to be a capability.

Cadence matters as much as content: automated metrics weekly, owner-sampled quality monthly, finance and utilization quarterly, incidents continuously. Reviews roll up from agent to fleet to function, and the fleet-level view lands on the executive dashboard beside revenue and headcount, because in the hybrid enterprise the machine layer is a workforce line item. The board that today asks “what is our headcount” will, within a few budget cycles, also ask “what is our agent count, what does it do, and who answers for it,” and the enterprises able to answer from a registry rather than a scramble will be recognizable by their composure.

Exhibit 11The Agent Scorecard

The Proof

One function, rebuilt

Models persuade architects; operators want to see one function actually rebuilt. So take the least glamorous corner of any enterprise, accounts payable, as an illustrative composite drawn from the shape of published automation deployments, and run it through everything above. Today the function is forty people; every invoice is touched by hands; cost per invoice sits around twelve dollars; cycle time runs about nine days; capacity is fixed at whatever forty people can process; and the deepest process knowledge lives in three veterans who could resign next quarter.

Rebuilt on the machine layer, the same function is twelve people and roughly sixty registered agents. Extraction, coding, matching and vendor-chasing agents run at tiers one and two, drafting and executing with review; a payment-run workflow operates at tier three inside hard caps, one of the few workflows in the enterprise that has earned it. The twelve humans do what the ladder reserves for them: exceptions, vendor judgment, dispute strategy, and the orchestrator role that owns the ratio and samples the output. Cost per invoice falls toward three dollars; cycle time toward a day; seventy percent of invoices go through untouched; and when an acquisition doubles invoice volume for a quarter, the fleet scales in days without a single requisition. The number that belongs in the CFO conversation is none of these, though: it is what happened to the other twenty-eight people, who moved to collections strategy, vendor negotiation and spend analysis, judgment work the old function never had the capacity to staff. The machine layer did not just make the function cheaper. It changed what the function is for.

Exhibit 12One Function, Rebuilt

The Strategy

The payoff: capacity becomes a dial

Everything to this point could still be mistaken for very good operational hygiene, so be explicit about the prize, because it is strategic and it compounds. An organization at level three holds options that a fixed-headcount competitor cannot exercise at any price. When demand doubles, it scales the fleet in days and hires only for judgment, while the competitor opens requisitions and waits three quarters. It can enter thin-margin segments where the headcount business case never closes, because its marginal cost of execution is machine-priced. It can run a hundred process experiments a year, A/B testing agents the way software teams test features, where the competitor cannot staff a single pilot. Its nights, weekends and regions are covered by default. When a key expert resigns, the playbook stays, because it lives in the registry and the fleet rather than in a head. And its unit costs fall with volume while the competitor’s rise roughly linearly with headcount.

None of these are efficiency gains, and pricing them as cost savings is how finance committees undervalue the entire program. They are asymmetries: moves one organization can make that the other structurally cannot. This is also where this edition joins the argument of the series. Edition 02 named the learning loop the last durable moat; Edition 05 showed the market beginning to price it. The machine layer is that loop’s organizational home: every exception a human resolves teaches the fleet, every scorecard cycle compounds, every workflow promoted up the authority ladder is captured, auditable, institutional learning that no departing employee and no competitor’s checkbook can take away. An enterprise that builds the machine layer is not automating its org chart. It is converting its organization into the compounding asset the whole series has been pointing at.

Exhibit 13The Strategic Asymmetry

The New Math

The new math, previewed

The asymmetry eventually shows up in the numbers, and one number in particular is about to become the most watched in business: revenue per employee. A mature software company runs somewhere between two hundred and four hundred thousand dollars of revenue per employee, and that band has been stable for a generation because capacity has always been made of people. The earliest AI-native firms are reporting figures past one million dollars per employee, with celebrated outliers past five million, tiny judgment cores running enormous machine leverage. Those are outliers, not medians, and this series will not pretend otherwise; survivorship and stage effects flatter every one of them. But the direction is structural, because it follows directly from the model in this edition: when execution capacity moves to the fleet, headcount stops being the denominator of output, and the ratio between the two becomes a designed number rather than an accident.

That is why the comparison that matters is not company against company but shape against shape. The traditional software firm is a pyramid whose capacity, cost and knowledge all live in headcount: it grows by hiring, its unit costs are roughly linear, and its playbooks resign along with its people. The machine-layer firm is a diamond: a small human core of judgment roles, the six above plus leadership, sitting over a large, registered, governed fleet; it grows by scaling agents, its unit costs fall as the fleet learns, and its knowledge compounds in the registry. Boards will learn to read human-agent ratio and revenue per employee together the way they read growth and margin together today. The full economics, real benchmarks, where the analogy to past leverage revolutions holds and breaks, and what it all implies for how these companies get valued, is a dedicated edition in this series, and it connects directly back to the repricing thesis of Edition 05: the market will price what compounds, and this is the org chart of compounding.

Exhibit 14The New Math, Previewed

The Other Side

Objections, taken seriously

Four objections deserve the floor, because each contains something true. The first: agents are not reliable enough for autonomy, so this is premature. For most workflows today, correct, and the blueprint agrees; that is why the ladder exists. Tiers zero through two assume imperfect agents and extract value from them safely, and tier three is earned by track record, not granted by enthusiasm. The design point is sequencing: governance must be standing before capability arrives, because the alternative, retrofitting control onto a thousand already-deployed agents, is precisely the level-zero swamp most enterprises are in. The second objection inverts the first: this bureaucratizes the one advantage agents have, which is speed. But unmanaged speed is borrowed speed, repaid at incident rates, and the registry is not friction; it is the instrument that makes delegation safe to increase. Brakes are not what makes a car slow. Brakes are what make it possible to drive fast. The enterprises that will run tier-three workflows confidently in two years are the ones building the visibility to justify that confidence now.

The third objection comes from the regulated world this series knows best: autonomous workflows are legally impossible in my industry. Mostly true today, at tier three, for qualified processes, and the model is built for that truth: bounded autonomy with named human ownership is not a workaround of oversight obligations, it is their implementation. The accountability chain is, almost clause for clause, what the EU AI Act’s human-oversight provisions and two decades of validated-automation precedent in GxP and financial controls already demand. Regulated enterprises will reach level four later and more narrowly than others; they need levels one and two sooner, because an unregistered agent touching regulated data is not a future risk but a present audit finding. And the fourth objection: the survey numbers are hype. Partly conceded, and this edition has flagged its statistics as directional throughout. But notice that the case for the blueprint never rested on any multiplier. It rests on a single, checkable fact about your own enterprise: somewhere in it, right now, an agent with production credentials is running that no central registry lists and no named human owns. If that sentence might be true, the ninety days below are not a bet on a trend. They are the closing of an open exposure.

Since the honest treatment of objections deserves honest evidence, here is what the record now shows about doing this wrong, and it strengthens the blueprint rather than the skeptics. MIT’s 2025 GenAI Divide research found roughly 95 percent of enterprise AI pilots delivering no measurable P&L return, and Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, on escalating costs, unclear value and inadequate risk controls: in both cases the diagnosis is the absence of exactly what this edition specifies, the registry, the owners, the tiers, the scorecards. The security record says the same thing in dollars: IBM’s 2025 breach research found shadow AI, unsanctioned tools quietly shipping enterprise data to external models, involved in one in five breaches and adding $670,000 to the average incident, with 97 percent of AI-breached organizations lacking basic AI access controls; which is why what information may be sent to which model is a constitutional decision in the machine layer, not a browser habit. Observability follows the same logic: an agent that is not instrumented, traced and measured cannot be governed, billed, or switched off, and the scorecard is what makes all three possible. And the policy clock has a date on it: the EU AI Act’s high-risk and transparency obligations are scheduled to apply from August 2026, with fines reaching seven percent of global turnover, so the never list, the registry and the named owners are not merely good design. They are compliance, banked early. For the architecture-level treatment, the AI-SAFE framework at akhawat.com/aisafe catalogs the decisions underneath this blueprint, thirty-six of them across six layers, and pairs each with the trust and value tests this edition’s governance plane applies.

Exhibit 15The Objections, Answered

The People

Leading the humans through it

Finally, the constraint that decides whether any of this ships, and it is not technical. Microsoft’s own data locates the readiness gap top-down, 67 percent of leaders familiar with agents against 40 percent of employees, and finds organizational factors driving roughly twice the impact of individual attitude. Translated: the workforce is not the obstacle; unled change is. A machine layer announced lands as a threat. A machine layer led moves through four stages, and each has a specific leadership action attached.

Fear comes first and should be answered first, which is why the never list is not only a control but the single most credible trust signal a leadership team can send: publish it early, and say plainly that decisions about people, hiring, firing, discipline, pay, are constitutionally off-limits to the machine layer, at any tier, forever. Clarity follows: show the map of which work transfers and which appreciates, and put the orchestrator career path, its grading and its expectations in writing before the first team is rebuilt, not after. Capability is a training investment: agent literacy for everyone, fleet skills for managers, and deliberate reskilling of relay-heavy roles toward exception and design work, the direction the middle-management section already drew. Ownership is the end state and it has a tell: people talk about their fleet and their numbers, escalations get celebrated as judgment rather than logged as failure, and agent leverage shows up in growth conversations, not just cost reviews. In jurisdictions with works councils and consultation obligations, start that dialogue in the census phase, not the cadence phase; the blueprint’s transparency is an asset in that room. Companies willing to change do not have a technology problem. They have a trust sequence, and it is entirely designable.

Exhibit 16From Fear to Ownership

The Mirror

The maturity ladder: where you actually are

Blueprints flatter; maturity models tell the truth. Five levels describe the road from where most enterprises are to where the model in Exhibit 4 operates, and the honest placement for the majority of large organizations today is the bottom rung. Level 0, shadow: agents exist, procured by teams, wired by enthusiasts, and nobody centrally knows where; this is risk without record, and Edition 04’s finding that roughly a fifth of breaches already involve unsanctioned AI is its price tag. Level 1, registered: one registry, every agent listed, every agent owned; unglamorous, transformative. Level 2, governed: authority tiers assigned, scorecards running, reviews on cadence. Level 3, orchestrated: fleets managed by design, orchestrator roles formalized, ratios tuned as an operating decision. Level 4, autonomous in bounds: tier-three workflows running alone inside hard limits, continuously audited, with kill switches that have actually been pulled in a drill. The distance from L0 to L1 is one quarter of disciplined work, requires no new platform, and removes the largest silent risk in the building. That is where the ninety days begin.

Exhibit 17The Maturity Ladder

The Blueprint

The first ninety days

Everything above compresses into a plan a leadership team can start Monday, in three phases of thirty days, each with a named executive owner, none requiring new headcount. Days 1 to 30, the census, owned by the CIO: inventory every agent, bot, and autonomous automation in the enterprise, including the ones marketing bought and the ones engineering built on a weekend; stand up the registry; name an owner for each entry or switch it off. Expect the count to be a multiple of what anyone guesses, and treat that gap itself as the first finding. Days 31 to 60, the constitution, owned jointly by the COO and CISO: assign every registered agent an authority tier; write the never list and take it to the board for signature; define what promotion between tiers requires and who signs it. Days 61 to 90, the cadence, owned by the CHRO and CDO together, and the pairing is deliberate, one owns the org design, the other the measurement: stand up the six-metric scorecard, set the review rhythm, rewrite one pilot team’s manager role as an orchestrator role, formally, with the grading and expectations in writing, and run one workflow through the full lifecycle end to end.

At day ninety you will not have finished building the machine layer; nobody has. You will have made it visible, owned, and reviewable, which puts you at Level 1 to 2 on the ladder and, on the evidence of this edition, ahead of most of your industry. More importantly, you will have converted the seven unanswerable questions of Exhibit 2 into seven assigned ones, and assignment, as every operator knows, is where answers start.

Exhibit 18The 90-Day Blueprint

Exhibit 19The Board’s Five Questions

A note on the inevitable procurement question, because the first thing many leadership teams will ask after “yes” is “which platform do we buy.” Platforms genuinely help: registries, orchestration, logging and evaluation tooling are maturing fast, and building all of it from scratch is rarely wise. But the machine layer is an operating discipline before it is a product, and the three things that make it work, ownership, authority and accountability, cannot be outsourced to a vendor at any price. Buy the plumbing where it is good. The constitution, the never list, the signatures and the names must be yours. The author writes this from the practitioner’s side of the fence: a two-year, phased platform modernization in the author’s own organization, independently documented by AWS with measured outcomes on scale, cost and reliability, followed exactly this discipline of piloting first, measuring everything, and owning the operating model while buying the plumbing, and it is that program’s second phase, the agentic conversion, that this blueprint describes.

The Horizon

The chart is the tell

Step back from the mechanics and notice what the org chart actually is: the most honest document in the enterprise. Strategy decks describe intentions; the chart describes commitments, who exists, who decides, who answers. That is why Lattice’s three-day experiment struck such a nerve, and why the chart is the right place to watch this transformation land. Edition 05 argued that as intelligence commoditizes, value migrates into the enterprise’s own compounding layer; this edition has drawn what that layer looks like when it reports for work. The repricing determines who captures the value. The machine layer determines whether your organization can actually hold it.

The deeper architecture beneath the chart, how the enterprise’s intelligence is structured, how knowledge compounds through it, and how the whole stack is designed rather than accreted, is where this series is heading. It arrives as the capstone of this arc, the Cognition Stack, six editions out, and it sits at the heart of my forthcoming book, publishing this September. The road there is deliberately practical, one operating question per edition, because blueprints are built layer by layer. The org chart took a century to perfect its answer to who reports to whom. The enterprises that win the next decade will be the ones that answer the new question first, and answer it on paper, with names: who answers for what.

Exhibit 20The Series Roadmap

The Author

About the author

Prashant Akhawat is a technology and AI leader with over 25 years building and scaling enterprise platforms, and Chief Technology and AI Officer at an AI-native platform company whose work serves more than 50 regulated enterprises across life sciences, pharmaceuticals, financial services, media and government. The two-year platform modernization he led is the subject of a published AWS case study: a 10x increase in scale, 60 percent lower total cost of ownership, and 99.7 percent SLA reliability, with the agentic second phase, the subject of this series, now underway. He is the creator of AI-SAFE, the AI Substrate Architecture Framework for Enterprises, and of the Cognition Stack; a keynote speaker at AWS Summit Bengaluru 2026 and AWS Summit Mumbai 2025; and has been featured in The Economic Times CIO and Mint. An alumnus of BITS Pilani and IMI Delhi, he is the author of a forthcoming book on how AI rewires the modern firm, publishing this September.

Read every edition at akhawat.com/writing  ·  Subscribe to the CXO Intelligence Series  ·  Connect on LinkedIn

Selected Sources & Further Reading

1. AWS-published case study independently documenting the two-year platform modernization in the author’s organization, with measured outcomes on scale, total cost of ownership and service levels: aws.amazon.com/solutions/case-studies/ninestar-case-study.
2. MIT NANDA, The GenAI Divide: State of AI in Business 2025 (enterprise pilot outcomes and causes); Gartner, June 2025 press release (agentic AI project cancellation forecast to end-2027 and “agent washing”); IBM and Ponemon Institute, Cost of a Data Breach Report 2025 (shadow AI incidence, the $670,000 cost premium, and AI access-control findings).
3. EU Artificial Intelligence Act (Regulation 2024/1689): phased application, including high-risk and transparency obligations scheduled from August 2, 2026, with penalties up to 7 percent of global turnover; the proposed Digital Omnibus adjustments remained under negotiation at the time of writing.
4. Prashant Akhawat, AI-SAFE: a reference framework of thirty-six architectural decisions across six layers for the AI-native enterprise, each paired with Trust Ring and Value Ring tests: akhawat.com/aisafe.
5. Reporting on the Lattice digital-workers episode, July 9 to 12, 2024: Inc., SHRM, Forbes, and industry coverage of the announcement, backlash, and reversal, including CEO Sarah Franklin’s statements.
6. Microsoft, 2025 Work Trend Index Annual Report: The Year the Frontier Firm Is Born (human-agent teams, the agent boss role, the human-agent ratio; survey of 31,000 workers across 31 countries).
7. Microsoft, 2026 Work Trend Index: Agents, Human Agency, and Opportunity (the transformation paradox; organizational factors driving roughly twice the impact of individual factors; adoption patterns by industry). Survey figures cited as reported.
8. Moderna company disclosures and reporting on its 3,000+ internal AI assistants and the merger of HR and technology leadership (see Edition 03 for full treatment).
9. EU Artificial Intelligence Act, human-oversight provisions (Article 14) and related enterprise-governance analyses.
10. CXO Intelligence Series: Edition 03, The Redesign (decision-centric organization, work planning over workforce planning); Edition 04, The Fragile Moat (shadow AI, least-privilege access, the governance case); Edition 05, The Great Repricing (agent productivity and the enterprise’s new scorecard).

Survey-based statistics (Work Trend Index and similar) are directional signals rather than precise measurements and are cited as reported by their publishers. Organizational-design guidance in this edition is a framework for adaptation to each enterprise’s regulatory and operating context, not one-size-fits-all prescription.