The Organizational-Change Frameworks

Generative AI has reproduced the gap the frameworks were built to explain

There is a pattern in business old enough to predate the computer: leading companies repeatedly lose position when the technology or the market underneath them changes.¹⁶ The intuitive explanation, that the incumbents were complacent or badly led, fits the cases poorly. The explanation that has held up, developed across organizational theory from the 1960s onward, is that the organization is a system, and that a system built to do one thing well is structurally hostile to doing a different thing at the same time.

The pattern matters now because generative AI has reproduced it at speed. By 2025, up to 88% of surveyed organizations reported using AI, 70% reported generative AI in at least one business function, and the population-level adoption curve ran faster than the PC or the internet.⁶ Over the same period, a widely-cited MIT study of enterprise deployments found roughly 95% of organizations getting no measurable return, with only about 5% of integrated pilots extracting real value.⁵ Adoption, in short, has run far ahead of value capture, and that gap is the territory these frameworks map.

Leavitt’s Diamond: change one variable and the system compensates

The oldest of the four statements belongs to Harold Leavitt, whose 1965 chapter “Applied Organizational Change in Industry” appeared in March’s Handbook of Organizations.¹ Leavitt modeled the organization as an interdependent socio-technical system of four interacting variable classes: task, structure, people (the human actors), and technology. His central claim is the one this whole corpus rests on. A change in any single variable produces compensatory or retaliatory change in the others, so the system tends to re-equilibrate around its prior state.¹ Effective change therefore requires acting on several variables at once, and a technology-only intervention is structurally absorbed by the unchanged task, structure, and people around it.

Sixty years on, the generative-AI deployment record supplies the field test Leavitt never ran. In the MIT sample, 60% of organizations evaluated enterprise-grade GenAI tools, only 20% reached pilot stage, and just 5% reached production; most failed, in the report’s words, “due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.”⁵ That list of failure causes is the unchanged task and structure rejecting the new technology. The same report finds seven of nine sectors showing no real structural change.⁵ This is the diamond’s prediction at population scale: the tool arrives alone, the other three variables hold their shape, and the system re-equilibrates with the tool expelled or idling.

Where the technology has taken hold, it has taken hold around the organization. The MIT report documents what it calls a “shadow AI economy” — employees using personal ChatGPT accounts, Claude subscriptions, and other consumer tools “to automate significant portions of their jobs, often without IT knowledge or approval.”⁵ Only 40% of the surveyed companies had purchased an official LLM subscription, yet workers at over 90% of them reported regular use of personal AI tools for work, in many cases several times a day, while their employers’ official initiatives “remained stalled in pilot phase.”⁵ For Leavitt’s model this is the vivid case. A technology change that demands nothing of task, structure, or people succeeds instantly at the level of the individual, the organizational deployment that demands all three stalls, and usage simply reroutes around the structure that failed to change.

March 1991: exploitation wins the short run and can destroy the long run

If Leavitt explains why a technology-only change gets absorbed, James March explains why the absorption is rational at every step. His 1991 Organization Science paper, “Exploration and Exploitation in Organizational Learning,” is the canonical statement.³ March defines exploration as “search, variation, risk taking, experimentation, play, flexibility, discovery, innovation,” and exploitation as “refinement, choice, production, efficiency, selection, implementation, execution.”³ Both are essential, both compete for the same scarce resources, and the organization must continuously choose between them.

Returns from exploitation are “positive, proximate, and predictable”; returns from exploration are “uncertain, distant, and often negative.”³ Because feedback ties exploitation to its consequences more quickly and precisely, organizations improve at what they already do faster than they improve at anything new, and those advantages compound: each gain in competence at an activity raises the reward for repeating it, which raises competence again.³ That asymmetry carries the whole argument. March’s stark conclusion is that “it is quite possible for competence in an inferior activity to become great enough to exclude superior activities with which an organization has little experience” — the mechanism later writers call the competency trap — and a firm exploiting to the exclusion of exploration becomes “trapped in suboptimal stable equilibria.”³

This is the precise reason a capable organization underinvests in a genuinely new technology. Management following the local feedback loops is doing its job. The loops reward the safe, near-term refinement of the existing business over the uncertain, distant payoff of the new one, and every quarter of refinement makes the next quarter’s choice easier to make the same way. The 2025 deployment record shows the trade-off in tool form. MIT’s gloss on its own divide is that ChatGPT “dominates for ad-hoc tasks but fails at critical workflows.”⁵ In March’s terms, the ad-hoc use carries exploitation’s return profile and the workflow integration carries exploration’s. The perception data sit on the same boundary. In the Stanford survey, inaccuracy rose to the top-cited AI risk, named by 74% of respondents, up from 60% a year earlier; an error rate that costs little in ad-hoc drafting costs a great deal inside a core process, which is where integration keeps stalling.⁷

March’s models add a second-order point that bears directly on AI deployment. In the mutual learning between an organization and its members, fast convergence is dangerous. If individuals adapt to the organizational “code” before the code can learn from them, variety collapses and the organization’s long-run knowledge degrades. Slow socialization and moderate turnover, counterintuitively, preserve the exploratory variety that keeps the system intelligent.³

Ambidexterity: exploration and exploitation need opposite alignments, and the prescription comes with boundary conditions

If exploitation reliably crowds out exploration, the design question becomes: how can one organization do both? Tushman and O’Reilly’s answer is organizational ambidexterity — “the ability to simultaneously pursue both incremental and discontinuous innovation… from hosting multiple contradictory structures, processes, and cultures within the same firm.”⁸ They frame March’s trade-off as the adaptive challenge it implies: a firm must “engage in sufficient exploitation to ensure its current viability and, at the same time, devote enough energy to exploration to ensure its future viability,” against a standing “bias in favor of exploitation with its greater certainty of short-term success.”⁸

Because exploitation and exploration require opposite alignments (efficiency, control and incremental improvement on one side; flexibility, autonomy and experimentation on the other), their structural prescription is to establish “autonomous explore and exploit subunits that were structurally separated, each with its own alignment of people, structure, processes and cultures,” joined by “targeted integration” at the senior level.⁸ The exploratory unit is protected from the parent’s resource-allocation and management logic precisely so that the competency trap cannot reach it. Their 2013 review is careful to add that this is “at heart, a leadership issue more than a structural one”; the structural separation is necessary but not sufficient.⁸

The 2013 review reports ambidexterity positively associated with sales growth, innovation, market valuation (Tobin’s Q), and firm survival, across studies at the firm, business-unit, project, and individual levels.⁸ One study can stand for the genre. Geerts and colleagues followed more than 500 firms over four years and found a positive effect of ambidexterity on firm growth, and the other large longitudinal samples in the review point the same way.⁸ The review also bounds the claim. A study of 605 technology companies found an inverted-U relationship between ambidexterity and financial performance, corroborated in a second large sample; an inverted U means the under-use of ambidexterity costs firms and so does the over-use, with the payoff peaking between.⁸ The figure that does the most work for this paper is the field measurement of March’s trap. In one large sample, an estimated 80% of firms under-emphasized exploration and over-emphasized exploitation.⁸

The prescription itself comes with conditions the review states plainly. It distinguishes three ways of achieving the balance — sequential (shifting structures over time), structural (simultaneous, separated units), and contextual (designing the unit so individuals divide their own time between the two demands) — and concludes that “all three are potentially viable.”⁸ Firms in practice combine and sequence them. In the review’s account, Raisch and Tushman found that “incumbent firms created new business by initially employing structural ambidexterity and switched to integrated designs when the exploratory unit achieved political and economic legitimacy,” so the separate unit reads as an opening move with an exit condition.⁸ The benefits are contingent as well. Ambidexterity is more valuable under environmental uncertainty and when sufficient resources are available, which is more often the case for larger firms, and in stable environments the review suggests firms “may be able to afford a sequential approach.”⁸

Christensen: the same autonomous-unit remedy, reached through resource allocation

Christensen’s disruption thesis is the same mechanism viewed through resource allocation. “Disruptive Technologies: Catching the Wave” (Bower & Christensen, HBR 1995) opens on the observation that “one of the most consistent patterns in business is the failure of leading companies to stay at the top of their industries when technologies or markets change.”¹⁶ The Innovator’s Dilemma (1997) sharpens the paradox: “even the most outstanding companies can do everything right — yet still lose market leadership.”¹⁷ The cause Christensen identifies is discipline. Resources flow to the sustaining innovations demanded by existing high-value customers, which systematically starves disruptive bets that initially serve small, low-margin, or non-existent markets.¹⁶ His Resources–Processes–Values framework makes the point structural. The resources, processes, and values that make a firm excellent at its current business actively disqualify it from pursuing a disruptive one.¹⁷ The integrated steel mills that earned only 7% margins on rebar were behaving rationally when they ceded that segment to minimills, and that rationality is exactly what let the disruptor climb upmarket.¹⁷ The remedy Bower and Christensen observed is the design already on the table: pursue the disruptive business through an autonomous unit freed from the mainstream organization’s resource-allocation logic.¹⁶¹⁷

The convergence deserves one honest weighing. Three frameworks arriving at the same autonomous-unit prescription looks like independent triangulation, and part of it is inheritance. The ambidexterity literature takes March’s trade-off as its explicit starting point — the review quoted above states the core problem in March’s own words — so those two witnesses share one line of thought.³⁸ Christensen reached the remedy along a different path, from field observation of resource allocation, which makes his agreement worth more as corroboration.¹⁶¹⁷ What survives the discount is still substantial. The frameworks are elaborations of one tradition, their prescriptions fit together because of it, and the empirical support rests on the studies reviewed above rather than on the agreement itself.

Raffaelli: incumbents can re-engage a legacy technology when leaders reframe its value

Every framework so far explains why the incumbent organization defaults to absorption. Ryan Raffaelli’s work supplies the boundary of that account, the conditions under which incumbents recover a position the default had cost them. He studies technology re-emergence — how a field given up for dead can come back — and his primary case is Swiss mechanical watchmaking.¹³ After Japanese quartz watches arrived in the 1970s, the Swiss share of global watch export value fell from 55% to roughly 30% within a decade.¹¹ Quartz won on cost as well as precision; Swatch’s production costs ran about 80% below the incumbent approach.¹¹ The rest of the record confirms the scale of the collapse. Export volume fell further still, employment dropped from about 90,000 to 33,000 by one industry account, and roughly two-thirds of Swiss watch companies were lost.¹¹ On a pure disruption reading, mechanical watchmaking should have died. By 2008 Switzerland was again the world’s leading watch exporter, having reclaimed 55% of total export value.¹¹

Raffaelli’s explanation runs through identity. The surviving firms stopped competing on precision, the dimension quartz had won, and redefined the value and identity of the mechanical watch as craftsmanship, heritage, and emotional meaning.¹³ In his words, “successful companies may be able to reposition a ‘dying’ technology by redefining its identity and value for the customer.”¹³ The mechanism is sociocognitive: re-emergence required institutional guardians (collectors and loyal employees who preserved the legacy technology and its meaning) in productive tension with institutional entrepreneurs (leaders who pursued new markets).¹¹ Raffaelli calls the firm-level version identity ambidexterity, preserving legacy capabilities while adapting to new markets, and notes that “a lot of companies fail because they cannot do both things simultaneously.”¹³

His companion paper, “Frame Flexibility,” locates incumbent resistance one level deeper, in the top management team’s cognition. It asks why incumbent firms reject non-incremental innovations and answers with frame flexibility: the capability “to perceptually expand an innovation’s categorical boundaries and to cast the innovation as emotionally-resonant with the organization’s identity, competencies, and competitive boundaries.”⁹ Forces of inertia generally constrict how leaders perceive an innovation; frame flexibility relaxes the usual assumption that cognitive frames are static and shows how reframing raises the likelihood of adoption.⁹

Mapped onto AI, this leg does two jobs the other frameworks leave open. Frame flexibility sits upstream of everything Leavitt and March describe. Whether the exploratory work is funded at all depends on whether the leadership team can cast the technology as belonging to the organization’s identity and competence, so the cognitive step precedes the structural one.⁹ And re-emergence bounds the pessimism the other frameworks invite. Incumbency constrains, on this record, and the Swiss case shows the constraint being undone where leaders reframed what the technology was for — a live option both for firms deciding what AI is for, and for firms whose legacy offering AI now threatens.¹¹¹³

The productivity J-curve: the value lag is expected, and it explains only part of the divide

There is a macroeconomic reason to expect the value lag and to hold off on panic. Brynjolfsson, Rock, and Syverson’s “Productivity J-Curve” shows that general-purpose technologies like AI require large intangible complementary investments — new processes, skills, and business models — that standard accounts capture poorly.¹⁰¹² Because those investments are made before their benefits are harvested, measured productivity is underestimated early in a GPT’s diffusion and overestimated later as the intangibles pay off, tracing a J.¹² Their intangibles-adjusted total factor productivity ran 11.3% above official measures at the end of 2004 and 15.9% above by the end of 2017; both points measure the prior computing wave, and they sit on the schematic curve below as a precedent for AI.¹⁰ The organizational rewiring this paper has been arguing for is, in macroeconomic terms, exactly the intangible investment the J-curve says must precede the payoff.

The J-curve, though, cuts against the paper’s own use of the 95% figure, and the tension needs facing. If GPT payoffs lag by years while intangibles accumulate (the verified points span thirteen years of the prior wave), then a 95% no-measurable-return rate two to three years into GenAI diffusion is close to what the model predicts even for organizations doing everything right. On that reading the reorganization is underway and simply unmeasured, and the figure indicts nobody. The saved evidence offers discriminators, all from the same MIT source and carrying its directional caveat. The crossers of the divide were already showing measurable results inside the window: mid-market top performers moved from pilot to full implementation in about 90 days while enterprises took nine months or longer, and specialized vendor partnerships reached deployment roughly 67% of the time against about one-third for internal builds.⁵¹⁵ The winners’ approach was also distinct and describable — they “pick one pain point, execute well, and partner smartly” — where a pure measurement lag would leave winners and losers looking alike until the curve turned.¹⁵ And the shadow economy shows value arriving instantly wherever the tool meets a task without organizational friction, which a lag story has no way to explain.⁵ The split tracks what firms do, which is the frameworks’ account. The lag story survives for the middle of the distribution, where intangible investment may be quietly underway, and the two readings will only be fully separated as the returns data mature.

There is a second reading of the divide that deserves to stand in its own words, because it comes from the paper’s main modern source. The MIT report names learning as the core barrier to scaling, finding that most GenAI systems “do not retain feedback, adapt to context, or improve over time,” and its executive summary concludes that “learning-capable systems, when targeted at specific processes, can deliver real value, even without major organizational restructuring.”⁵ On that reading the binding constraint is tool capability, and the remedy is to buy learning-capable, process-specific systems, drive adoption from the front lines, and hold vendors accountable to business metrics.⁵ The frameworks’ reading and MIT’s overlap without collapsing into each other. Selecting, customizing, and embedding a learning system in one process is itself exploratory work, precisely the kind an exploitation-aligned organization starves, so the tool-capability story still needs an organizational explanation for why some 95% of buyers failed to do it. But MIT’s “even without major organizational restructuring” is a genuine caution against this paper’s thesis at full strength. If a well-chosen tool in one process can pay off inside an otherwise unchanged firm, the reorganization the frameworks prescribe may be the ceiling on value capture and something short of it may clear the floor. The claim this paper retains is the weaker, better-supported one — the gap is organizational in origin — while the size of reorganization required to close it stays open.

Who the separate-unit answer is for, and what the frameworks predict

The prescription that survives this record is conditional, and the conditions are usable. For a large firm facing genuine technological uncertainty (the position most enterprises occupy with generative AI), the evidence favors structural ambidexterity: an exploratory unit with its own alignment of people, structure, processes, and culture, integrated at the senior level and treated first as a leadership problem.⁸¹⁷ The sequencing evidence refines the design. The separate unit worked in the reviewed cases as an opening move, with firms switching to integrated structures once the exploratory unit had earned political and economic legitimacy, so the separation is a stage with an exit condition, and the exit condition is legitimacy.⁸ The benefits of the full structural design are strongest for larger, resource-rich firms, and in stable environments the review says firms “may be able to afford a sequential approach.”⁸ The inverted-U stands as the warning on the other side, that exploration can be over-built as well as starved.⁸

Around that structural core, the other frameworks set the operating constraints. Budgeting for the tool without budgeting for change to task, structure, and people buys absorption; measuring the exploratory work on the exploitation business’s cadence starves it through feedback loops doing exactly what they were built to do; and no structure protects work the leadership team cannot yet frame as the organization’s own.¹³⁹

Read through all of this, the 95%-zero-return finding is a report on organizational defaults, carrying its directional caveat. Most buyers acquired the technology and held task, structure, people, and incentives constant, which is the outcome every framework here predicts for a technology-only intervention. The small group extracting real value behaved differently in describable ways — one pain point, a learning-capable system, clear accountability — and whether closing the gap for everyone else demands full ambidextrous redesign or only that narrower discipline is the live question the next few years of returns data will answer.⁵ The frameworks’ bet, and this paper’s, is that the system decides what the tool becomes.

References