The AI Operating Partner: Why an Old Form Reappears Whenever Technology Outpaces Absorption

A century of professional services traces a single pattern. One kind of firm — strategic judgment paired with operational accountability, working from inside the client — keeps disappearing from the consulting industry and returning under each new general-purpose technology. AI has summoned the same kind of firm again.

The Argument

Frederick Winslow Taylor stood on the shop floor at Bethlehem Steel in 1898 with a stopwatch in his hand and a contract in his pocket that paid him on tons of pig-iron moved, not on hours worked. He lived inside the plant. He hired the men, redrew the workflows, owned the outcome. He raised pig-iron loading from 12.5 tons per man per day to 47.5 tons across three years by redesigning the rest interval, the shovel weight, and the incentive scheme. He never submitted a recommendation. He did the work.^[1]

Charles Bedaux ran similar installations across roughly 700 American and European factories through the 1920s. General Electric, Du Pont, Standard Oil, and Imperial Chemical Industries each took a Bedaux engineer for months or years. The fees ran on a share of the measured productivity gains his teams produced, which gave Bedaux a direct financial stake in whether the work actually paid off.^[2] Edwin Booz, founding the practice that would become Booz Allen Hamilton in 1914, ran the same model: months on site, deliverables measured in operational change rather than in pages of analysis.^[3]

These firms invented the profession of management consulting. They were paid by outside companies. They ran operations from inside those companies. The word consultant, in its modern sense, had not yet been coined.

That model didn’t last. Marvin Bower took over McKinsey in 1950 and rebuilt the firm around the structure of a white-shoe law firm: partner track, hourly billing, departure when the engagement closed.^[4] From inside the consulting industry, the change looked like the profession growing up. From the perspective of the operations consultants used to run, it looked like abandonment. By 1990, the consultant who stood inside the factory had been replaced by the consultant who stood in front of a slide deck.

The buyout industry brought the embedded operator back. KKR and Hicks Muse, assembling themselves out of Drexel Burnham’s high-yield apparatus through the 1980s, discovered within five years that financial engineering alone didn’t produce the returns their model promised. Someone had to live inside the portfolio company and actually run it. The Operating Partner was the answer — at first an industry veteran on retainer, then a structured role with a board seat, equity, and carried interest tied to the buyout’s eventual sale price.^[5]

This kind of firm appears whenever the work can’t be done from the outside. It appeared in the 1980s when financial engineering alone fell short. It appears again now because AI touches every function inside the enterprise and arrives faster than internal teams can hire, train, and deploy against it. The field calls the current instance the AI Operating Partner.

Taylor didn’t have a word for the work he was doing. Neither did Bedaux. The work came first. The names caught up later.

Coase’s Variable and the Embedded Form

Embedded operator firms exist for a specific economic reason that Ronald Coase identified in 1937.

Coase asked a small question with large consequences. If markets are so efficient at allocating resources, why do companies exist at all? Why don’t businesses just contract every job out on the open market? His answer became foundational economics. Markets carry hidden costs: finding the right counterparty, negotiating the terms, monitoring whether the work gets done, recovering damages when contracts fail. Companies exist because organizing work under a boss who pays salaries is often cheaper than negotiating every task on the open market.^[6]

Oliver Williamson refined the argument across the four decades that followed. He named three forces that determine where a company’s boundaries should fall: bounded rationality (you can’t foresee every contingency in advance), opportunism (counterparties may cheat), and asset specificity (some investments have no value outside the specific relationship they were made for).^[7]

Embedded operator firms exist because three conditions hold at the same time. First, the work is too complex to write into a contract clearly enough. The job changes as it gets done, and no contract can anticipate everything that will come up. Second, the buyer’s company can’t hire the experts internally fast enough — there aren’t enough of those experts on the market to recruit a team. Third, the result takes years to develop, and consultants paid by the hour have no economic reason to stick around long enough to see it through.

These firms don’t fit either the vendor model or the employee model. The relationship runs too deep for a vendor — vendors don’t sit on the executive committee. The relationship is too time-bounded for employment — the firm leaves when the work is done. Instead, the firm signs a contract that ties its pay to the result. It works inside the client’s decision-making structure but isn’t on the client’s payroll. Williamson called this arrangement hybrid governance. Embedded operator firms are the clearest real-world example of it.

Williamson’s third condition — asset specificity — explains why this structure makes economic sense. The expertise an embedded firm builds inside one client engagement doesn’t transfer cleanly to another client. The firm learns how this client’s data flows, how this client’s people work, how this client’s customers behave. That knowledge has very little value outside the relationship. Williamson argued that when one party has to make investments that only matter for one specific relationship, both sides need a structure that protects the relationship itself. Pure contracts leave both sides exposed. Pure employment is too rigid. A hybrid arrangement, with shared incentives and outcome-linked pay, protects both sides. The buyout industry’s Operating Partner, the cloud-era hyperscale firm, and now the AI Operating Partner all settled on roughly the economic structure Williamson described.

Michael Polanyi made a related point even more simply in 1966: we know more than we can tell.^[18] Experts can’t fully explain everything they know. Some of their expertise lives in their judgment and their hands, not in any document. What Taylor learned about how a steel plant actually runs couldn’t be written down in a manual that another firm could pick up and execute. What an Operating Partner learns about a portfolio company over a five-year hold lives in operational instinct. Documentation captures some of it. It never captures all of it. That’s why a contract alone doesn’t work — you can’t write a contract for work that neither side can yet describe. And it’s why hiring alone doesn’t work — you can’t hire an employee to do work that your existing team doesn’t yet know how to do. The embedded firm operates in the gap between those two modes.

One more refinement matters. The three conditions explain why these firms exist. They don’t explain which of them are any good. The arrangement only works when the buyer can trust the firm to put the buyer’s interests above the firm’s own short-term profits. That trust takes years to build and seconds to lose. Williamson’s framework can’t tell you which firms will hold the trust and which will exploit it. The structural definition fits both kinds equally well. The buyer has to read the firm’s track record across prior engagements to know which one they are actually hiring. Price reflects scarcity. Quality reflects the specific firm.

The same explanation fits every prior instance. Taylor sat inside Bethlehem Steel because nobody had written down how to run a steel plant before he did the work, and the workforce had been shoveling for decades without ever being measured. Bedaux’s seven hundred installations had the same shape: the productivity science had to be performed on site, calibrated against the specific machinery and labor patterns of each plant. The Operating Partner KKR places inside a portfolio company is the modern version — too deeply involved in the operation to be a vendor, too clearly temporary to be management, paid based on the buyout’s eventual sale price rather than on a salary.

The pattern returns whenever three conditions hold together: the work is too complex to specify in advance, the buyer can’t build the team fast enough, and the outcome matters too much to leave to an advisor whose engagement ends before the result is clear. When all three conditions hold inside a market large enough to support a firm, an embedded operator firm appears. When any one of the conditions stops being true — the work becomes specifiable, the buyer builds its own team, or a consultant’s recommendation turns out to be enough — these firms recede.

The Withdrawal

Marvin Bower took over McKinsey in 1950 with an ethical mission. He believed management consulting needed a code of conduct strong enough to let the firm refuse work that compromised the consultant’s judgment. The white-shoe law firm gave him a template: a partner track, hourly billing, an obligation to walk away from clients who refused to take the consultant’s advice, a refusal of contingent fees that could bias the analysis.^[8] Those rules satisfied his ethical concerns. They also pulled consultants out of the operations they had previously run.

The lawyer charges by the hour because legal advice doesn’t produce a measurable physical artifact. The work product is judgment. The billing unit is the time spent producing it. Bower applied this lawyer’s architecture to a profession that had originally produced the kind of operational change Taylor produced at Bethlehem. Over thirty years, the architecture reshaped what the work actually was.

BCG launched the experience curve in 1968 and the growth-share matrix in 1970. These were tools for boardroom conversations, not for shop-floor execution. Bain built its practice around the partner-led case team — deeply analytical, deliberately at arm’s length from operations.^[9] By the time Tom Peters and Robert Waterman published In Search of Excellence in 1982, the management consultant had become an interpreter of organizations. The interpretation had real value. It did not change what was happening on the shop floor.

The embedded operator survived elsewhere. EDS, founded by Ross Perot in 1962, ran corporate data centers under five-year fixed-price contracts.^[10] Andersen Consulting ran multi-year ERP installations inside Fortune 500 manufacturers through the 1990s. The technology-services sector kept the embedded operator alive across the decades when the management consulting industry treated the role as beneath itself.

By the late 1980s, corporate buyers had begun to notice the gap the withdrawal had produced. The slide decks were excellent. The outcomes were uneven. Recommendations on paper rarely produced operational results that compounded. The buyout industry, assembling itself in the same decade, saw the gap clearly and built around it.

The Rediscovery

KKR’s first major leveraged transaction — Houdaille Industries in 1979 — closed at $355 million and proved that the buyout structure could be executed at scale.^[11] The decade that followed produced the canonical names: Beatrice, RJR Nabisco, Safeway. Within five years of those transactions, the buyout firms saw a pattern. Financial engineering produced part of the return. Operational improvement during the holding period produced the rest. And the buyout firms didn’t have the people to execute the operational improvement.

The Operating Partner filled that gap. Lou Gerstner spent eleven years as president of American Express and four years as chairman of RJR Nabisco — a KKR portfolio company — before IBM recruited him in 1993. Larry Bossidy served as Jack Welch’s vice chairman at GE through the 1980s before becoming chairman of Allied Signal in 1991.^[12] Both men had run real businesses before they advised anyone else’s. The buyout firms began hiring such operators on structured terms: equity, a board seat, a defined portfolio engagement, and carried interest tied to the eventual sale price.

By 2010, every major buyout firm had institutionalized this role. KKR Capstone grew to several hundred professionals. Blackstone built a comparable practice. Bain & Company’s research has traced the structural consequence: multiple expansion peaked as a source of buyout returns in the industry’s first decades and has compressed steadily since the global financial crisis, while operational improvement now accounts for the largest single share of post-2010 buyout returns.^[13]

KKR Capstone’s mechanics became the template the industry studied. The group maintained a bench of operational specialists organized by function (revenue, supply chain, working capital, technology) and by industry vertical. When KKR closed a deal, Capstone deployed a senior operator inside the portfolio company within thirty days. The Capstone partner sat on the portfolio company’s executive committee, attended the board, and built the value-creation plan around the operational thesis the deal had been underwritten against. Blackstone’s portfolio operations group followed. Smaller buyout firms either built their own benches or retained external operators on structured terms. By 2015, the Operating Partner had become a standard line item in every PE fund’s organizational chart. The model spread because it captured what good operators do without freezing the work into a fixed methodology. The role was clear enough to replicate across deals and specific enough to add measurable value during each hold.

The compensation carried the substance of the model. The Operating Partner didn’t bill hours. He took a fraction of the fund’s carry, sat on the board, lived inside the portfolio company’s operations for the length of the hold, and earned his money when the exit closed. The role inverted every structural feature of the McKinsey partner. The engagement ran through deployment, not through advice. The relationship deepened the more involved the operator got. Pay tracked the buyout’s eventual outcome rather than the hours spent.

The buyout industry rediscovered this model because Coase’s three conditions required it. Taking a portfolio company through a five-year turnaround was too complex to specify in a contract. It exceeded the portfolio company’s existing management (which had often been part of the problem in the first place). It couldn’t be delegated to advisory consulting whose engagement ended before the turnaround did. Only the embedded operator structure produced the returns the firms needed.

The consulting industry watched from a distance and tried to adapt. McKinsey launched private equity practice groups and operations-focused arms. BCG opened Platinion. The Big Four built systems integration practices. None of those moves restructured the firm at its core. Each of them added outcome-adjacent capacity to a model that still ran on hourly billing. The partner pyramid required the predictability of billable hours, and an outcome-aligned engagement model would have collapsed the compensation structure that funds the partnership.

The McKinseys could see the problem clearly. They couldn’t restructure around it without dismantling the economics of partnership itself. The firms best positioned to observe the gap were the same firms whose internal economics would have collapsed if they had tried to fill it.

Two professional services models had split on a question nobody had been asking openly: who carries the risk that the outcome will fail? The Operating Partner carries it. The hourly consultant — by design — doesn’t. The split mattered modestly through 2020. It matters very differently now.

The General-Purpose Technology Law

Timothy Bresnahan and Manuel Trajtenberg published the definitive paper on general-purpose technologies in 1995 in the Journal of Econometrics.^[14] A general-purpose technology — electricity, the semiconductor, the computer, AI — is one that touches every sector of the economy. Its productivity gains depend on co-invention by the adopting industries. The technology by itself produces nothing. The complementary investments — process redesign, organizational restructuring, skill development — produce the gains, and those investments lag the technology by years or decades.

Paul David made the lag concrete in his 1990 paper on the dynamo.^[15] Electrification took roughly forty years from the first commercial deployment of central station power in the 1880s to the productivity surge in American manufacturing of the 1920s. The lag wasn’t technological. It was organizational. Factories that had been built around steam-driven central drive shafts had to be physically reconstituted around distributed electric motors before the productivity case for electrification showed up in the data. Companies that reorganized first captured the gains. Companies that just bolted electric motors into the old steam-shaft layout captured only a fraction.

Brynjolfsson and McAfee read AI through the same lens in The Second Machine Age and in Brynjolfsson’s subsequent paper on the productivity J-curve.^[16] When a general-purpose technology arrives, productivity initially falls or stays flat as companies spend on the complementary changes that are needed before the gains arrive. Measured productivity rises only after enough of the reorganization is done.

When a general-purpose technology arrives, it produces exactly the conditions Coase’s framework identifies. Rebuilding a company around the new technology can’t be fully specified in advance — the technology keeps evolving while the rebuilding happens. The buyer’s company can’t hire the people it needs fast enough — the senior practitioners cluster inside a small set of firms during the early adoption window. And the outcome takes years to develop, longer than a billable-hours consultant has economic reason to stay. An embedded operator firm emerges to do the work.

The mainframe wave produced EDS. Ross Perot founded the firm in 1962 with a $1,000 stake and a single observation: corporations had bought mainframes they couldn’t operate. EDS didn’t sell machines. It absorbed the client’s entire computing operation under five-year fixed-price contracts that wove the firm into the client’s daily operations. By 1984, EDS sold to General Motors for $2.5 billion. The embedded operator captured the value the hardware vendor couldn’t.^[10]

The client-server wave reorganized IBM. Lou Gerstner, arriving in 1993, refused the breakup that IBM’s board had drafted and rebuilt the company around end-to-end delivery of integrated operating capability. He sold the result the client needed, against contracts that ran for years. By 2007, IBM Global Services generated roughly $55 billion in revenue and accounted for the majority of IBM’s profit.^[17]

The cloud wave produced Accenture’s hyperscale practice, AWS Professional Services, Slalom Build, and a generation of cloud-native integrators. AWS Professional Services lived inside Amazon Web Services as the cloud era took shape, working at depths the product organization couldn’t reach. Slalom, founded in 2001 and grown into a multi-billion-dollar firm by the mid-2020s, organized its entire delivery model around what it called “local market” embedment — consultants living in the same city as the clients they were working with, so the relationship could run at the depth the work required.^[19] None of those firms invented something new. They recovered an old model and calibrated it to the cloud era’s specific contours. Issue 005 documents the investor returns this pattern has produced across sixty years.

Every wave summoned the same kind of firm. Every wave gave it a different name. The underlying structure did not change.

The Form Under AI: What an AI Operating Partner Is

The definition does three things at once that no adjacent firm structure does. First, the thesis. The firm arrives with a first-principles analysis of where AI creates measurable value inside the specific business — which functions, on what time horizon, against which competitive pressures, under which governance constraints. Second, the build. The firm doesn’t just recommend systems — it builds them. Agent architectures. Evaluation harnesses. Orchestration layers that decide what AI agents should do when the goals aren’t fully clear. Mechanisms that hand decisions to humans when uncertainty crosses a threshold. Observability systems that catch drift before it becomes failure. All of it installed inside the client’s production environment. Third, the duration. The engagement runs through deployment, through the first production failure, through the iterations that follow, and through the point at which the client’s team can run the system without the firm’s daily help.

Adjacent forms

The MBB or Big Four consultant produces a thesis and delivers it as a slide deck. The pricing model — hours billed against a statement of work that ends at the recommendation — can’t accommodate operational accountability across deployment. The consulting firm’s billable hour ends where the recommendation lands. The AI Operating Partner’s engagement begins there.

The systems integrator implements a specification the client (or the client’s strategy firm) has already produced. The AI Operating Partner writes the specification, because the specification follows from the operating thesis the firm carries into the engagement. The integrator brings execution capacity. The embedded operator brings the strategic direction that capacity serves.

Internal AI teams and Chief AI Officers run the systems a company already has. An embedded operator builds the systems the company doesn’t yet have, then transfers them. The internal team is the institutional heir to the engagement. The strongest engagements end with the client’s internal team running a capability the embedded firm installed.

The private equity Operating Partner brings cross-portfolio operational pattern matching — CEO succession, EBITDA improvement, preparation for exit. The AI-specific firm brings technical fluency in the architecture, governance, and deployment of agentic systems at a depth a generalist operating partner doesn’t have. The two roles share the structural logic. The substantive expertise is different.

The shape of the engagement

A typical engagement begins with six to twelve weeks of thesis validation. The firm works alongside the executive team to produce the operating roadmap that the rest of the engagement will build against. The architecture phase converts the roadmap into the specific agent systems, evaluation methods, and governance instruments the client will install. The execution phase runs for twelve to twenty-four months while the firm’s engineers work inside the client’s production stack and ship capability against the milestones in the roadmap. The transfer phase ends with the client’s operators running the capability and the firm shifting into a quarterly governance and roadmap role. Each phase produces something the next phase requires. Skipping any phase breaks the model.

Why now

The unit cost of autonomous reasoning has dropped by approximately 99.7% across the past three years on a per-token basis, and the capability frontier has advanced through three model generations in the same period. A single embedded firm can now take outcome-based accountability, profitably, for work that would previously have required permanent in-house headcount. Issue 006 develops this argument in detail.

The senior practitioners of agentic systems cluster inside roughly a dozen firms globally — the frontier labs, the hyperscaler practices, a handful of specialized boutiques. The Fortune 500 buyer cannot hire from this population at the scale the technology demands. Embedding is the only practical way for the buyer to reach this scarce expertise.

Every Fortune 500 board has approved an AI strategy without appointing an executive whose specific charter is to deliver it across the function set. The embedded operator firm closes the coordination gap that board approval has opened.

All three conditions are converging at the same time. Embedded operator firms are emerging to fill the gap.

The Discipline the Form Imposes

Holding this kind of position responsibly requires the firm to make choices against its own short-term commercial interest. Five specific disciplines come up during a typical engagement. A buyer can recognize them by looking at what excellent embedded firms have done in prior waves.

The work is the transfer

The knowledge an embedded firm acquires inside an engagement does not belong to the firm. It belongs to the engagement, and through the engagement to the buyer. The firm’s job is to leave that knowledge behind in usable form: a runbook the client’s team can actually run from, documentation that survives the firm’s departure, training that turns the client’s internal staff into operators of the system the firm installed. EDS’s strongest engagements in the 1970s left behind real operating discipline that the EDS team had spent years installing. When clients later tried to bring those data centers back in-house, they discovered which engagements had transferred the knowledge well and which hadn’t. The transfer of knowledge isn’t a closing artifact. It is the work itself, and it gets measured at the end of the engagement, not at the start.

The exit is architected from the first week

A management consulting engagement ends when the deck is delivered. An embedded operator engagement ends when the capability runs without the firm. The engagement therefore has to be designed backward from that ending. Success criteria specified up front. Timeline scheduled. Milestones at which the firm steps back and the client steps forward. KKR’s strongest Operating Partners worked to the fund’s exit timeline. The way the firm structures the engagement reveals what kind of firm the buyer is dealing with.

The board needs an instrument it can read

Hybrid governance is harder for boards to oversee than either pure vendor relationships or pure employment. The board can’t read the embedded firm’s work the way it reads an internal team’s work, and it can’t run an RFP on the firm’s work the way it can on a vendor’s. The firm has to give the board something it can read instead: operating reviews the firm proposes, capability scorecards published quarterly, governance escalations the firm itself triggers when things start going wrong. Gerstner published IBM Global Services’ operating disciplines openly because he understood that the legitimacy of an embedded firm depends on the board’s ability to read what the firm is actually doing.

Talent stewardship is the paradox

The firm’s senior practitioners are the firm’s main asset. They build compounding domain knowledge across engagements that no internal counterpart can match. And yet the work requires the firm to send that talent into the client’s organization as the engagement closes. The firm that hires its own practitioners into the client at engagement close, that allows the client to recruit against the engagement team, that sees itself as an institution producing operators for the client side — that firm holds the position. The firm that retains the talent it has developed for its own future engagements has quietly turned the model into vendor extraction. For a firm that wants to hold the position responsibly, the paradox has only one resolution: send the people in.

Indispensability is the test the firm must fail

The hardest discipline cuts against the firm’s own commercial logic. An embedded firm earns its position by becoming indispensable. The stated end-state of the engagement is to make the firm dispensable. A firm that genuinely transfers capability shortens its own engagement. A firm that hoards capability extends it. No contract can resolve this tension. The arrangement depends on the firm choosing — every time, on every engagement — to fail the indispensability test. The behaviors that show a firm passing that test are specific and visible: shipping the runbook the client’s team will operate without the firm, hiring the firm’s own talent into the client’s organization at engagement close, scheduling the exit on a defined timeline before the engagement even begins. Firms that hoard the work signal it differently: they stay perpetually one quarter away from a clean handoff. The early years of the engagement look identical in both cases. The buyer has to do the reading, and the reading gets clearer only as time passes.

The five disciplines compound. A firm that transfers knowledge but resists oversight will eventually break the buyer’s trust. A firm that builds the oversight but hoards the talent will struggle to compound across engagements. Firms that hold all five disciplines compound across engagements. Firms that hold only some of them erode their position over time.

Those of us who work inside this business model treat these disciplines as definitional. Without them, the model isn’t legitimate. The model’s structural conditions can change — a new technology, a new way of building internal capability, a new regulatory regime — and when they change, the embedded operator firm recedes again, the way it has before. In the 1930s, as standardized industrial engineering programs started producing internal practitioners at scale, Bedaux’s market for installations receded. Companies could now do the work internally because they finally had the people to do it.

AI has summoned this kind of firm back. The same kind of firm was summoned earlier by the mainframe, then by client-server, then by cloud. Whatever follows AI will summon it again.

What We See Now

The pattern is now 125 years old. Taylor walked into Bethlehem Steel in 1898 because the technology had arrived before the company had built the capability to operate it. Bedaux installed productivity science across seven hundred factories because the work depended on systematic measurement the workforce had never been trained to perform. Corporations had purchased mainframes they couldn’t operate, so Perot ran their data centers for them. Financial engineering alone produces only part of an LBO’s return, so the Operating Partner sits inside the portfolio company to produce the rest.

Each time the same arrangement appears. Embedded operator firms return because Coase’s three conditions converge inside markets large enough to support them. The names change. The arrangement stays the same.

How these firms get paid has shifted across the instances, and the shifts are worth examining. Taylor was paid in pig-iron tons moved. Bedaux’s installations earned the firm a share of the measured productivity gains. The Operating Partner takes carry on the buyout’s eventual sale price. Each pay structure is the era’s available answer to the same problem: how to align the embedded firm’s economics with the outcome the buyer needs, while preserving the boundary that keeps the firm distinct from the client’s own employees. The AI era will settle on its own answer. The early signal is that engagement economics will pay out against capability transfer milestones — pay tied to whether the client’s team can actually operate the system the firm built. That structure handles the model’s most visible failure mode, which is the firm that extends the engagement by hoarding the work.

The AI Operating Partner is the current label. The firms operating inside that label aren’t inventing anything. They are occupying the position the structural pattern has created. The work earns legitimacy from what the firm leaves behind in the client’s organization, not from what the firm retains. Those of us working inside this model treat the dependence we create as temporary — something the engagement is designed to dissolve.

The AI era will produce its embedded operator firms. The buyer’s task — anyone engaging this kind of firm in 2026, 2027, or 2029 — is to figure out which firms hold the position responsibly and which don’t. The structural pattern guarantees that these firms will exist. The work they do won’t be done by any other kind of firm. A buyer who needs the work and doesn’t engage someone to do it will eventually discover that the work didn’t get done.

Embedded AI firms will exist. Whether the current generation produces firms worth engaging is the question the era leaves for the buyer to answer.

Caerus Alpha’s mission is to be the leading AI Operating Partner in the new era.