Rewiring Work vs. Automating It

Automating scales the process you have; redesigning rebuilds it around the technology

An organization facing a capable new technology has two moves available. It can automate, which means taking the existing sequence of tasks and having the machine do them faster or cheaper while the structure of the work stays put. Or it can redesign, which means rethinking what the work is for, recombining the tasks, and rebuilding the process around what the technology now makes possible. Both moves count as “adopting AI” in a survey. They sit at opposite ends of the value distribution, and the claim this paper assembles evidence for — from the management literature, from macroeconomics, and from recent generative-AI field studies — is that the first move captures a small, bounded gain while the second captures the large, compounding one.

The failure mode in the first move is older than AI. Bill Gates compressed it into two rules: “automation applied to an efficient operation will magnify the efficiency,” and “automation applied to an inefficient operation will magnify the inefficiency”¹⁶. A process you automate without first improving is a process whose flaws you have just hard-coded and scaled. Hammer & Champy put the same point more bluntly in their 1993 sequel: “Automating a mess yields an automated mess”¹. The idea often circulates as anonymous folklore. Its actual provenance is named authors with page numbers, and the verification note below records the one open question, the exact book page of the Gates formulation.

Hammer’s 1990 prescription: obliterate the process, and accept an all-or-nothing bet

The foundational text is Michael Hammer’s 1990 Harvard Business Review article, “Reengineering Work: Don’t Automate, Obliterate.” Hammer’s diagnosis was that companies had spent a decade using information technology to mechanize old ways of doing business, leaving inefficient processes intact and using computers simply to speed them up¹. His prescription was the inverse: “use the power of modern information technology to radically redesign our business processes in order to achieve dramatic improvements in their performance”¹. The metaphor that carried the argument is the one still quoted today:

Hammer was equally explicit about what the cure costs. “Reengineering cannot be planned meticulously and accomplished in small and cautious steps. It’s an all-or-nothing proposition with an uncertain result”¹. That concession matters as much as the metaphor, because it names up front what the failure statistics later in this paper confirm. The redesign path is a high-variance bet, and its founding author said so. The canonical illustration from the article, Ford’s accounts-payable function, made the upside of the bet concrete: by redesigning the procurement process around a shared database and “invoiceless” payment rather than automating the existing invoice-matching task, headcount in that function was cut dramatically.

Davenport & Short published the complementary founding text the same year. “The New Industrial Engineering” argued that “business process design and information technology are natural partners,” a relationship industrial engineers “have never fully exploited”⁹. Drawing on field studies of nineteen companies, they observed that organizations which “used IT to redesign boundary-crossing, customer-driven processes have benefited enormously,” while IT implementation often failed precisely because it automated flawed processes rather than reimagining them⁹. Davenport’s later work, Process Innovation (1993), extended the discipline. Thirty-two years on, he returned to it for the AI era, and the closing section picks that up.

The dynamo precedent: electrification paid off only after factories were rebuilt around it

The regularity behind Hammer’s prescription shows up across general-purpose technologies, and Paul David’s 1990 study “The Dynamo and the Computer” is the canonical case. Factory electrification produced almost no measured productivity gain for roughly two to three decades after Edison’s first generating stations opened in 1881; by 1900, motors powered less than about 5% of factory mechanical drive¹⁰. The surge arrived in the 1920s, once factories abandoned the centralized steam-era layout of shafts and belts for the “unit drive,” an individual electric motor on each machine. Unit drive let plants be laid out around the workflow rather than around a central power source. David’s explanation for the lag is the cost of replacing capital and work organization “adapted to the old regime”¹⁰. Swapping a steam engine for a big electric one bought little. Rebuilding the factory around what electricity now permitted bought the revolution.

Brynjolfsson, Rock & Syverson formalized the lag for the AI era. Their 2017 paper on the “modern productivity paradox” — AI capabilities soaring while measured productivity growth had halved — weighed four explanations (false hopes, mismeasurement, redistribution, implementation lags) and concluded that lags are “likely the biggest contributor”³. The mechanism is the complementary investment. Like other general-purpose technologies, AI’s “full effects won’t be realized until waves of complementary innovations are developed and implemented,” and “the required adjustment costs, organizational changes, and new skills can be modeled as a kind of intangible capital”³. Their 2018 follow-up gave the pattern its name and shape, the Productivity J-Curve: because that intangible investment is poorly captured in the accounts, measured productivity is understated early, while firms are investing, and overstated later, when those investments are harvested⁷. The dip is the cost of the reorganization, and the upswing is its harvest.

Generative AI reproduces the pattern: adoption near-universal, absorbed value rare

The generative-AI evidence is reproducing the same finding on a compressed timescale. Adoption moved fast by the survey numbers. Stanford’s 2025 AI Index reports organizational AI use jumped to 78% in 2024 from 55% in 2023, and the share of organizations using generative AI in at least one business function more than doubled over the same year, from 33% to 71%¹⁴. Bottom-line impact has lagged well behind. McKinsey finds just 39% of organizations reporting EBIT impact at the enterprise level⁵, and only about 11% of companies worldwide using generative AI at scale¹³.

Nearly everyone has adopted; few have absorbed the technology into how the work is done. On the J-curve, that is the early stretch of the dip.

The industry evidence puts the effort in people and process, and the gains in reshaped workflows

Two consultancies, working from different survey bases, locate the missing value in the same place. BCG’s 10-20-70 rule distributes the effort required to capture value from AI: roughly 10% to algorithms, 20% to the underlying technology and data, and 70% to people and processes⁸. The model, the thing most executives instinctively treat as the project, is explicitly the smallest slice. BCG’s complementary framing sorts AI plays into three tiers. Deploy covers off-the-shelf tools that streamline everyday tasks and “boost workforce productivity by 10–15%”; Reshape is the “re-imagination of functions through workflow re-engineering,” driving “30–50% improvements in efficiency, effectiveness across affected functions”; Invent builds AI-native offerings and new business models, a “new revenue play”⁸. Between Deploy and Reshape, between inserting the tool and redesigning the work, sits roughly a three-to-five-fold difference in function-level impact. BCG’s own summary is that “this is a people transformation not a tech transformation”⁸.

McKinsey’s survey evidence triangulates the same conclusion from the demand side. In its March 2025 “State of AI: How organizations are rewiring to capture value,” McKinsey reports that “out of 25 attributes tested … the redesign of workflows has the biggest effect on an organization’s ability to see EBIT impact from its use of gen AI”⁵. The roughly 6% of respondents who qualify as AI high performers, attributing at least 5% of EBIT to AI and reporting significant value, are “nearly 3x more likely to have fundamentally redesigned workflows as part of their AI efforts”⁵. One caveat belongs on both findings. They are cross-sectional survey correlations, and organizations already capable enough to redesign their workflows may be the ones capable enough to capture value from AI, so selection can carry part of the association. Read with that limit, the direction still matches the rest of the section. McKinsey Global Institute’s $2.6–4.4 trillion annual estimate carries the same condition in its own framing, since it is explicitly contingent on leaders “re-examining and redesigning core business processes”¹².

Brynjolfsson’s augmentation case, and what a tool-insertion field study actually showed

If Hammer supplies the “redesign, don’t automate” half of the argument, Erik Brynjolfsson supplies the “augment, don’t imitate” half. His 2022 essay “The Turing Trap” draws the distinction precisely: automation systems “substitute for human labor,” while augmentation systems are “focused on augmenting humans rather than mimicking them”². The trap is that the field has, since Turing, implicitly aimed at human-imitating AI, and that this target carries two costs. The first cost is distributional. “As machines become better substitutes for human labor, workers lose economic and political bargaining power and become increasingly dependent on those who control the technology,” whereas augmentation lets “humans retain the power to insist on a share of the value created”². The second cost is the one that decides the strategy question, and it is about value:

The nearest field test of augmentation is also, read carefully, a test of plain tool insertion. Brynjolfsson, Li & Raymond studied 5,179 customer-support agents at a Fortune 500 firm during a staggered rollout of a generative-AI assistant into the existing support process; the treatment the study measures is access to the tool, and the study describes no accompanying workflow redesign. Productivity, measured as issues resolved per hour, rose 14% on average, with sharp heterogeneity underneath. Novice and low-skilled workers gained 34% while the effect on “experienced and highly skilled workers” was minimal⁴. The study’s own account of the mechanism is that the AI “disseminates the best practices of more able workers and helps newer workers move down the experience curve”⁴, and the same rollout improved customer sentiment and employee retention⁴.

Two readings follow, and they should be kept apart. As direct evidence, the study shows access to the tool paying on its own. A Deploy-tier move, in BCG’s terms, delivered a double-digit average gain and 34% for the workers who needed it most. The redesign reading — that codifying the tacit knowledge of high performers into a workflow available to everyone amounts to reorganizing how expertise flows through the team — is this paper’s synthesis, and it carries that label here.

The jagged frontier: the same tool helps inside its capability and hurts outside it

Augmentation has an operational edge condition. AI capability is “jagged,” strong on some tasks and abruptly weak on adjacent ones, so the gains depend on matching AI to the right work. The largest field experiment on this, Dell’Acqua, Mollick and colleagues’ study with 758 BCG consultants, quantifies both edges of the frontier, and here too the design was tool insertion, with consultants handed the tool and no organizational restructuring around it. On tasks inside the frontier (work AI is good at), consultants using AI completed 12.2% more tasks, roughly 25% faster, at roughly 40% higher quality than the control group⁶. On tasks outside the frontier (beyond current AI capability), consultants using AI were 19 percentage points more likely to produce incorrect answers⁶. The same tool and the same people produced opposite effects, depending entirely on whether the work was matched to the capability.

Among the treated consultants, the study observed two usage patterns in the effective users: the Centaur, a human-led division of labor that switches between AI and one’s own work by task, and the Cyborg, fine-grained integration of human and AI effort throughout the task⁶. Both patterns emerged from individuals adapting; neither was designed by the organization. The step from there to “redesign the division of labor deliberately” is this paper’s inference, drawn because the patterns that worked look like small-scale versions of the workflow re-engineering the industry data reward. The study itself, though, tested a handed-out tool, and its inside-frontier gains arrived without any organizational redesign at all. Read against the paper’s thesis, the jagged-frontier result supports the heterogeneity and task-matching points directly, and supports the redesign point only by extension.

The OECD’s vacancy-level evidence shows what reorganization looks like inside a single job. Its 2024 study of online vacancies across ten OECD countries records a concrete case: an insurance company adopted an AI tool that flags which customers are likely to escalate a service issue, and “the job of a sales agent changed to emphasise greater customer interaction with less time needed to analyse customer files”¹⁹. The work moved toward the human-strong activity.

On the broader question of where skill demand is heading, the same source cuts both ways, and its own caveats should travel with it. Across AI-exposed occupations, demand for business and management skills rose about 8% and for emotional, digital and social skills about 15% over the period studied. The OECD is explicit that this rise is not isolable to AI, since the same increases appear in less-exposed occupations (“factors other than just AI may be driving these changing skills demands, such as the general trend towards increased digitalisation”)¹⁹. When the study instead isolates the AI effect at the workplace level, the direction reverses for those skill groups. Demand for management, business and digital skills fell by over three percentage points in workplaces becoming more exposed to AI, a decline the authors call “relatively small” but worth monitoring as adoption grows¹⁹. The insurance example is the cleaner support for the augmentation reading; the aggregate skill-demand direction is contested within the source itself.

Source: Dell’Acqua, McFowland, Mollick et al., “Navigating the Jagged Technological Frontier,” HBS/BCG WP 24-013, 2023.

The price of redesign: over two-thirds of transformations fall short of expectations

Everything so far prices the upside. The same BCG deck that supplies the 30–50% Reshape range also prices the risk. BCG reports that “over 2/3 of transformations fall short of expectations (in terms of time, budget, meeting ambition),” against a backdrop it sizes at $1 trillion of wasted IT spend across the S&P 1200⁸. The people-and-process challenges it names are the ones a redesign program runs straight into, “resistance, opposition, and fear about AI impacting jobs” and “challenges with implementing new processes and reimagining workflows”⁸. Hammer had already conceded the shape of this risk in 1990 when he called reengineering an all-or-nothing proposition with an uncertain result¹. The failure base rate is the empirical version of his caveat.

Set the two paths side by side as an expected-value problem. Deploy is cheap and fast, and its 10–15% arrives without an organizational rebuild⁸; the customer-support study shows what that looks like in the field, a double-digit gain from tool access alone, with no redesign described⁴. Reshape offers 30–50%, conditional on landing in the minority of transformations that deliver on their ambition⁸. The comparison can still favor Reshape, since the conditional payoff is three to five times larger⁸. It remains a comparison between a small, likely gain and a large, uncertain one, and a leader who picks Reshape is accepting variance along with the bigger range.

There is a second warning in the same dataset. BCG finds 68% of GenAI-using companies already have Reshape plays in motion⁸, while McKinsey finds only 39% of organizations reporting enterprise-level EBIT impact and about 6% qualifying as high performers⁵. The two surveys have different samples and different questions, so the two numbers are not strictly comparable. Even so, self-described reshaping is near-universal while measured results are rare, and the plainest reading of that gap is that much of what companies call workflow redesign runs shallower than the Reshape tier BCG describes.

The failure base rate turns the thesis into a decision rule. When change capacity is constrained, when the process is a commodity that would not repay rebuilding, or when the organization cannot absorb the two-thirds shortfall risk, the bounded gain is the correct choice. Take the 10–15%, bank it, and build the redesign muscle on a narrower front. Reshape earns its risk where the function is core, where the payoff window is long, and where leadership is prepared to fund the intangible investment the J-curve says comes first⁷.

Why organizations stall between adopting and absorbing, and where redesign should aim

The sequence explains the stall. Adopting a general-purpose technology is cheap and fast, while the productive reorganization around it is expensive, slow, and intangible; that asymmetry is the J-curve’s dip⁷, and most organizations are sitting exactly where the model predicts, tools adopted and reorganization pending.

What gets the few across, in McKinsey’s operations research, is process discipline. “A disciplined, stage-gated review process with clear go/no-go criteria separates the merely promising deployments from the ones most likely to be productive,” and two-thirds of respondents set a three-to-five-year timeline for realizing full value¹³. That three-to-five-year figure is the implementation lag from the economics literature, now showing up in survey answers.

Redesign also needs a target, and the targeting logic borrowed here — with the caveat in the callout below — is Eliyahu Goldratt’s Theory of Constraints: a system’s throughput is governed by its single binding constraint, improvement effort anywhere else yields little, and once a constraint is broken the bottleneck moves. Applied to AI, the logic says that accelerating a step that was never the bottleneck produces no system-level gain; it builds inventory in front of the next constraint. A generative-AI assistant can make one step twice as fast and leave end-to-end throughput unchanged, because the binding constraint sat elsewhere. Redesign works on this logic because it re-architects the whole flow and relocates effort to the constraint.

The counterargument: “so-so” automation displaces labor without productivity to show for it

The strongest economic pushback on enthusiasm for automation comes from Acemoglu & Restrepo. Their 2019 work decomposes automation’s effect on labor demand into a productivity effect and a displacement effect, and shows the displacement effect “always reduces the labor share”¹¹. Their counter-intuitive conclusion is that the real danger to employment and wages is not “brilliant” automation but “so-so technologies that generate small productivity improvements” — automation just good enough to be adopted, but not good enough for its productivity gain to offset the labor it displaces¹¹. Their named example is “automated customer service, which has displaced human service representatives but is generally deemed to be low quality”¹¹, and they note this category “may also include several of the applications of artificial intelligence technology to tasks that are currently challenging for machines”¹¹.

So-so automation is the formal economic statement of the Gates and Hammer warnings. Take the automation path without redesign and you displace the labor, hard-code the flawed process, and capture too little productivity to justify either. Acemoglu & Restrepo warn that “further automation, especially when it is induced by … excessive enthusiasm about automating everything, would take the form of such so-so technologies and would not bring much in productivity gains”¹¹. The escape route they leave open is the one Brynjolfsson prescribes, augmentation and the reinstatement of new, higher-value tasks, which in operational terms means redesigning the work. For the paper’s thesis, the counterargument raises the stakes. If automation without redesign can leave workers worse off while capturing little productivity, the cost of stopping at pure automation runs beyond the foregone gain.

A decision frame: Hammer’s axis crossed with Brynjolfsson’s

The literatures align on two axes. One is Hammer’s, whether you embed the existing process in software or rebuild it. The other is Brynjolfsson’s, whether the AI substitutes for people or extends what they can do. The high-value quadrant, redesign plus augmentation, is where BCG’s “Reshape,” McKinsey’s “rewiring,” Davenport’s process management, and the Centaur and Cyborg patterns all live. The diagonal opposite, an unchanged process run by substituting machines, is Acemoglu & Restrepo’s so-so automation.

Davenport, who helped found the discipline in 1990, returned to it in 2025 with the observation that process management “is experiencing a renaissance, thanks to AI,” because “AI helps firms significantly scale up improved processes, and well-managed processes make it easier to obtain the high-quality data needed to train AI”¹⁸. Wilson & Daugherty add that natural-language interfaces now let non-technical employees participate in redesign directly, an evolution of Toyota’s continuous-improvement model they call “kaizen 2.0”¹⁷. Thirty-five years after Hammer told managers to stop paving the cow paths, both halves of his article still hold. The large gains sit in the redesign, and the redesign is the move that most often falls short. Choosing between the bounded gain and the priced bet is the actual strategy decision.

References