The Absorption Gap

For the first few years of the modern AI cycle, it made sense for the conversation to orbit capability. The practical question was whether these systems could write, code, reason, use tools, and stay coherent long enough to matter inside real work. Every twelve months or so, the answer changed enough to expand the conversation. Each turn pushed the frontier a little further of what the technology could do.

That was the right conversation for the first phase. It is no longer the most important one. The scarce capability is no longer access to AI; it is the ability to absorb AI into work faster than the frontier moves. The capability is already enough to alter real work. Powerful models are broadly accessible, interfaces are easier to use, and the cost of running intelligence over ordinary work has fallen so far that tasks uneconomic to automate two years ago now run for pennies. Every board has asked about AI, every CEO has an answer, and knowledge workers are already using the tools daily. Yet the value is still uneven, lumpy, and weirdly hard to find inside the P&L. Firms can buy the same tools, give employees access to the same models, and still end up in very different places.

That distinction matters because adoption and absorption are different problems. Adoption begins when a firm gives people AI tools. It is a diffusion problem. Absorption begins when the firm changes the work around the tools. It is an operating-model problem: workflows made legible, judgment codified, evaluations built, escalation paths defined, and roles redesigned around what the technology can now own. The gap between those two states is why capability is everywhere and value is still uneven.

The operating-model rewrite is hard because three forces compound. Pace is external: the frontier is not one model crossing from impossible to obvious but many curves moving at once, faster than the firm can finish scoping the last deployment. Drag is internal: the operating model that made the firm successful now slows the absorption loop precisely when it needs to tighten. The Void is developmental: the disciplines and roles the new operating model runs on never existed in the prior era, so the firm has no muscle for them yet.

The firms that integrate AI well are the ones willing to act on today's capability and build the operating agility to ride the next wave. They do not wait for the stack to settle, because it will not settle soon. Reacting to industry buzz about 2025 being the “year of agents,” Andrej Karpathy, who co-founded OpenAI, reset the clock: “In my mind, this is more accurately described as the decade of agents.” The frontier will keep compounding across that decade, and the surface area addressable by agents will expand. So the durable work must happen alongside it. Make work legible, instrument it, deploy against it, study the failures, rebuild the process, and repeat. Each cycle leaves the firm more legible to the next model, tool, or agent than it was to the last.

Pace moves the target. Drag slows the organization. The Void leaves the firm without the muscles required to adapt. The bottleneck occurs because absorption moves at the pace of organizational change while the technology moves at the speed of the frontier. This is not a software rollout. It is an operating-model rewrite.

That is why speed of absorption is critical. The winners will not merely use AI more often. They will become more absorbent companies whose systems are built to convert capability into operating leverage faster than their competitors. They absorb more capability first, then more margin, more share, and eventually the companies that failed to keep up. Carried across a market over years, the gap decides not only how the leaders rank but which companies are still on the list.

Software Starts to Own the Work

Software is moving from applications people operate to systems that operate processes. For most of the SaaS era, software gave the user a place to enter data, manage workflow, view dashboards, and coordinate human execution. The work still belonged to the person. The system recorded it, structured it, and made it more manageable.

That boundary is moving. The agent is becoming the active layer with the interface, database, integrations, workflow engine, and audit trail receding into infrastructure underneath it. The product is less the screen a person uses and more the process the system completes.

Three changes made this possible.

The first is model capability. Models became good enough for knowledgeable users to delegate meaningful pieces of real work, inspect the output, correct the failure, and try again. The leap was not that the model knew more. It was that the model became better at staying oriented inside a task: holding an objective in context, decomposing the work, choosing a tool, inspecting what came back, and continuing. METR's time-horizon work puts a number on the shift: frontier agents' reliable task horizon has recently been doubling about every four months, reaching tasks that take human experts more than 14 hours. The exact benchmark will move. The operating meaning is what matters: bounded persistence against real work is becoming practical.

The second is reliability. Bounded agents became production-usable in workflows where the task is observable, the cost of failure is tolerable, and escalation paths can be defined. Reliability is not a model property in the abstract. It is a workflow property. A 95% reliable agent may be production-grade for Tier 1 customer support but unacceptable for clinical decisions or credit approvals. The relevant question is where the current reliability level is good enough for the cost of failure, and where the firm can catch failures before they become expensive. Lemonade's claims agent shows the shape in a regulated industry: in its 2025 annual report the insurer disclosed that AI handled 96 percent of first notices of loss with no human involved and settled 55 percent of claims end to end.

The third is inference cost. The cost of applying intelligence to work has collapsed, and the practical unit is no longer cost per token but cost per resolved task. Epoch AI has tracked the slope: the cost to reach a fixed capability level on benchmarks has fallen nearly 10x a year. More importantly, firms can now route work across frontier, mid-tier, and workhorse models instead of sending every step to the most expensive system. When inference is expensive, AI gets reserved for high-value expert work. When inference is cheap and routable, it becomes rational to apply intelligence to the ordinary connective tissue of operations: intake, routing, checking, summarizing, reconciling, documenting, and following up. The deployment map widens by an order of magnitude when the unit economics flip, and by another when the routing layer matures. Both have happened.

Together, those changes alter what software can be. It no longer has to stop at helping a person do the work. It can increasingly run the work, ask for help when needed, and leave behind a record of what happened.

That does not make human judgment disappear. It changes where judgment sits in the operating model. The agent can click, draft, reconcile, route, summarize, and follow up, but the firm still has to define what good looks like, which exceptions matter, when to escalate, and what risk the business will tolerate.

Capability has crossed. Reliability has crossed inside bounded surfaces. Cost has collapsed. Software has changed shape. The value is bankable at today's capability. As the stack around the model matured (cheap routable inference, reliability engineering, packaged scaffolding), even mid-tier and year-old models became good enough to run bounded work in production and clear their cost many times over. Capability is no longer the scarce input; it is already here, broadly and cheaply. As Bret Taylor, who chairs OpenAI's board and runs the AI agent company Sierra, put it, even “if we paused model development, we'd still have trillions of dollars of economic value ... that have yet to be realized.” That flips the constraint: the question is not whether the model is good enough, but whether the firm can absorb what is in front of it. And the unit that has to change is the process and the firm, not the task.

Adoption Has Outrun Absorption

Once capability, reliability, cost, and software shape crossed together, the question shifted from what AI could do to how fast diffusion would happen. That part has largely been answered. The technology is already in market at every layer of the economy: foundation models are scaling into massive consumer and enterprise platforms, application-layer AI companies are growing at speeds that would have looked impossible a few years ago, and large enterprises and small businesses alike are putting generative and agentic tools into daily use. AI has cleared procurement, entered the workflow, and reached the employee. McKinsey's late-2025 survey put 88 percent of organizations using AI in at least one function, up from 78 percent a year earlier. The diffusion itself is the fastest on record. A St. Louis Fed study found generative AI reached about 40 percent of US adults within 2 years of ChatGPT's launch, faster than either the personal computer or the internet, both of which sat near 20 percent at the same age. And the spend is already enormous. Enterprise generative-AI spending went from $1.7 billion in 2023 to about $37 billion in 2025 with some 2026 estimates as high as $100B making it the fastest-scaling category in software history. Spending at that rate buys tools, not the capacity to use them.

PwC found that 56 percent of firms see neither lower cost nor higher revenue from AI deployments. And the failure rate is climbing: 42 percent of organizations abandoned most of their AI initiatives in 2025, up from 17 percent the year before. Tools have reached someone's screen; the work itself has not necessarily changed. Most firms running an AI scoreboard are counting seats, logins, or queries which are vanity metrics borrowed from the prior software cycle because they are easy to produce. Or they are measuring tokens which shows use but not utility. The absorbent firm counts the work: cycle time on a workflow, cost per resolved task, capacity reallocated, exceptions caught upstream, outcomes achieved. Usage is the metric of a diffused tool. Value is the metric of an absorbed one.

Aggravated Absorption

Three forces are causing friction in absorption. One moves the target. One slows the firm. Another is the missing muscle underneath both.

Pace is the moving target, and it is external. The frontier does not stand still while a firm works through its absorption window. Models improve, release cycles compress, reliability tiers expand, and new domains become production-ready. A deployment scoped around what models could do in January may be underwritten against a different capability stack by May. The release cadence itself has compressed: the gap between one lab's frontier models fell from roughly 6 months in 2023 to about 2 months by 2025 and leading labs shipped 7 frontier models between February and April 2026 alone, roughly one every 11 days. The vendor landscape adds another layer of motion: new products, model providers, workflow agents, vertical platforms, orchestration tools, and integration layers arrive constantly, each one creating another route to evaluate. The relevant change is not one launch, but the compression of the planning cycle. The skill set is moving too. Prompt engineering was the visible skill of the first wave. Context engineering, eval design, agent orchestration, workflow instrumentation, and process codification are becoming more important as firms move from experimentation to production. A firm running hard can still fall behind, because the surface itself is moving. Microsoft's CTO, Kevin Scott, named the same gap from inside the frontier in mid-2026: AI “breakthroughs are arriving faster than institutions, workflows, and human systems can absorb them.”

Drag is the weight of the operating model you're still carrying, and it is internal. Layered governance designed for stable decisions. Legacy systems built around older workflows. Data fragmented across functions and vendors. Processes optimized for human coordination rather than machine execution. Compliance routines that assume software is deterministic. Incentives that reward visible experimentation more than the harder work of process redesign. Tacit judgment trapped in operators' heads. None of this was irrational when it was created. Much of it was the residue of good management in the prior era. But the accumulation slows the absorption loop precisely when the loop needs to tighten.

The Void is what you have to build as you shed the drag and try to catch pace. It is the absence of capability the prior era never asked for: workflow articulation maintained as a discipline, evaluation infrastructure that catches regressions, trust calibration built through repeated use, context maintenance with a named owner, agent governance, and tooling that compounds individual reasoning across a team instead of trapping it in single-player context. None of this existed in the prior operating model because none of it was needed.

Leaders feel that absence. CEOs can call AI a top priority and still be unprepared to absorb it. The return shows up only when firms build into the Void: PwC's 2026 AI performance study found the firms pulling ahead were twice as likely as other firms to redesign their workflows around AI rather than bolt tools onto the ones they had.

Pace, Drag, and the Void multiply. The frontier keeps moving while the firm is still trying to make its own work readable and still hasn't built the muscle the new work demands. The heavier the inherited operating model and the further behind the new disciplines, the more each external shift compounds internal delay. Firms that treat AI as a tool wave respond by waiting for the stack to settle, choosing pilots carefully, and asking for proof before they rewire anything important. That response feels prudent because it worked in slower software cycles. Here it risks leaving the durable work undone.

That is the absorption gap. AI diffuses through access. Value diffuses through absorption. Access is now broadly available. Absorption is still scarce, uneven, organizationally expensive, and operator-led.

Ability to Absorb

Firms do not absorb AI all at once because absorption changes who owns the work in layers. At the surface, a person uses a tool. Deeper in, the person still owns the task, but AI assists. Deeper still, an agent owns a bounded task and escalates exceptions. Then an agentic system owns a process that used to move across several roles. Eventually the firm rebuilds roles, accountability, and management layers around agentic capacity.

The first is Adopt. A person uses a general-purpose AI tool. The firm is building literacy and reps, not changing the operating model yet. Firms are putting general-purpose agentic tools (Claude Code, Cowork, Replit, Cursor, Codex) into the hands of the workforce. Walleye Capital mandated Claude Code across its 400 employees, Rokt ran a 700-person 24-hour hackathon that shipped 135 internal apps, and Deloitte rolled Claude out to 470,000 staff. The bar is participation and fluency: learning how to work with models, use skills, apply simple harnesses, and move beyond prompt-engineering a chat interface. The person's skills have changed but the work hasn't.

The second is Augment. The person still owns the task, but AI makes the task faster, better, or cheaper. The human invokes the tool, reviews the output, and decides what happens next. The general-purpose tools from the first layer and vertical vendor platforms get pointed at specific work. The user builds skills, schedules recurring jobs, and runs simple agentic workflows against a single task. One person now directs a handful of agentic components instead of doing each step by hand, so the work gets faster and usually better. But the person still owns the task, and the agent works inside a loop the human still drives.

The third is Automate. A narrow agent owns a bounded job end to end inside one role, queue, or seat. The human no longer drives each instance of the task; they define the boundaries, review exceptions, and improve the agent. The work may involve several steps, but it does not yet redesign the broader process or cross multiple handoffs. Avoca is a good example: its agent answers and books inbound calls for home-services contractors. The unit of work is still one seat, but ownership has crossed from person to agent. This is the layer where firms stall on the climb.

The fourth is Transform. An agentic system owns a workflow that used to move across several roles, systems, or handoffs. The firm is no longer automating one job inside the old process; it is redesigning the process around what agents can own, where humans enter, and how exceptions, quality, and accountability work. Drillbit, an Austin-based startup, runs this inside an operating business: an integrated agent system handles a home-services contractor's office work end to end, answering the inbound call, qualifying the lead, building the quote, scheduling the job, and chasing payment, the chain that used to move through a receptionist, an estimator, a dispatcher, and a back-office clerk. The atomic unit of work becomes the process itself, not the task.

The fifth is Rebuild. The firm reorganizes around agents as labor. Roles, management layers, accountability, and coordination routines change because the production model has changed. Layers come out. Roles get redrawn. Managers become player-coaches. Block shows the rebuild in a public company: in March 2026 Jack Dorsey cut roughly 40 percent of its 10,000-plus staff and, in an essay titled “From Hierarchy to Intelligence,” redrew the org around three roles, individual contributors, directly responsible individuals, and player-coaches. The org chart starts to look less like an industrial hierarchy and more like a group of process owners directing a much larger pool of agentic capacity. The organization changes around the new production model.

The ladder is not about how advanced the tool looks. It is about how far ownership moves: from person with tool, to person-plus-AI, to task agent, to process system, to firm design.

To make the layers concrete, picture the same role across all five.

A sales rep using Claude chat for meeting prep has adopted AI. A rep using AI to draft follow-ups, refresh pipeline notes, and prepare account research while reviewing every output is augmented. When an inbound lead is enriched, answered, booked, and logged by an agent without the rep driving each step, a bounded task has been automated. When prospecting, qualification, outreach, reply handling, scheduling, CRM logging, and handoff all run through an agentic workflow, the SDR motion has transformed. When the firm no longer organizes revenue work around handoffs between SDRs, AEs, RevOps, and managers, but around accountable owners directing agentic capacity against outcomes, the firm has begun to rebuild.

These are not neat stages every company moves through in order. They are depths of absorption. And not all work should move deeper. The first workflows to absorb are high-volume, observable, language-mediated, economically meaningful, and bounded by clear escalation paths. Trust-heavy, irreversible, ambiguous, or relationship-sensitive work stays human-owned longer.

The firms that identify the right work at the right layers, and keep pushing deeper, are not merely adopting AI faster. They are building the muscle to become more absorbent. The firms that do not run it may still have widespread AI usage, but the usage will sit on top of the old operating model. They will look modern at the tool layer while remaining slow at the work layer.

The Competitive Surface

Most firms hear “AI absorption” and reach for the IT budget. The data substrate, security architecture, integration layer, model routing, and agent infrastructure are real work. They are expensive, unavoidable, and easy to underestimate. But they are not where most of the durable advantage forms. They are the cost of entry.

The competitive surface sits above the technical foundation. It is the operating muscle the firm builds around the technology: workflows made legible, judgment captured from the people who know the work, evaluations that define what good looks like, escalation paths that make failure survivable, context maintained by named owners, and agents governed by people whose job is to make them better over time.

That advantage is already concentrating. PwC's April 2026 study found that 74 percent of AI's measured economic value sits with 20 percent of organizations, and the firms pulling ahead index on growth rather than cost savings. The management literature points to the mechanism: durable edge comes from codifying tacit judgment without flattening it into generic process documentation.

The counterargument is that waiting gets easier every year. Models improve, agent tooling matures, vendors package the scaffolding, and the late firm can buy what the early firm had to build. That is partly true. The generic layer will commoditize. Evals, orchestration, monitoring, context tools, and agent templates will get cheaper and better. A firm can buy the same model, hire the same systems integrator, and deploy the same vendor platform as a competitor.

But the generic layer is not the absorption layer. A firm cannot drop an agent on top of its data and expect it to run the coordination and judgment work of the business. The agent still needs to know which workflows matter, which exceptions break the business, what good output looks like, where escalation belongs, what risk the firm will tolerate, which roles change, and which humans still own judgment. None of that arrives in the box, because most of it was never written down. It lives in the operators' hands and heads, and a model cannot package work it never gets to observe.

That is why speed compounds. The early absorber is not merely using today's comparatively weaker tools. It is building the operating infrastructure that lets every stronger tool land faster. The late firm gets the better agent. The early firm gets the better agent plus the map.

The advantage may narrow where generic scaffolding replaces work firms once had to do themselves. But the frontier also expands the surface area of what can be absorbed. As agents become capable of running more of the company, the absorption burden moves deeper into roles, accountability, trust, incentives, and operating design. The firm that has been learning all along activates the frontier faster. The firm that waited still must become absorbent.

Once built, it compounds. The workflow documented at the Augment layer becomes the input for automation. The evaluation built for a bounded task becomes the control system for a broader process. The escalation path defined for one agent becomes the governance pattern for the next. Each layer crossed leaves behind operating infrastructure the next layer can use.

Closing

The absorption gap has a name in the literature. Brynjolfsson, Rock, and Syverson called it the Productivity J-Curve: the period when a general-purpose technology has arrived, but measured productivity stays flat or even falls while firms pay the complement costs required to use it well. The technology is visible before the value is visible. The costs show up first: new processes, new roles, new skills, new systems, new ways of organizing work.

That is the shape of the current AI cycle. The disappointing part of enterprise AI is not that the technology has failed to land. It is that the complements are expensive. Adoption can happen through procurement. Absorption requires the firm to document how work actually happens, codify judgment, build evaluations, define escalation paths, rebuild processes, and eventually restructure roles around agentic capacity. Those are the complement costs of this transition.

The standard lesson from prior technology waves is patience. Electrification took decades to show up fully in productivity because factories first installed electric motors into operating models designed for steam. The gains came later, after managers rebuilt the factory around the new power source. The historical pattern is lag before lift.

But this wave changes the catch-up math. In prior waves, the technology stabilized while firms climbed the J-Curve. The laggard could copy the leader's tools, hire similar talent, run the same implementation playbook, and close much of the gap a few years later. The complement work was hard, but the target was relatively stable.

AI is not giving firms that kind of stationary target. The frontier moves while the firm absorbs it. Models improve, costs fall, reliability tiers expand, interfaces change, and agentic capability reaches new workflows before most firms have finished rewiring the last one. The complement costs are no longer a one-time payment required to reach a new steady state. They are the operating muscle required to keep absorbing as the frontier moves.

That is why speed compounds. A firm that documented its workflows, codified its judgment, built its evals, and rebuilt its first processes in 2024 does not merely have a head start. Each turn through the loop converts more operating knowledge into something the system can run on. The next deployment is faster because the work is more legible, the next process easier because the evaluation muscle exists, the next agent safer because the escalation paths are already understood.

The firm that waits does not stand still. It falls behind a moving curve. Each cycle of waiting leaves more work undocumented, more judgment trapped in people's heads, more processes optimized for human coordination, and more organizational debt sitting between the firm and the capability now available to it. The gap widens not because the leader bought a better tool, but because the leader became better at absorbing whatever capability comes next.

Across the decade, the gap shows up as response time, service quality, cost structure, management layers, pricing power, and market share. Eventually it shows up as ownership. Firms that absorb AI faster will not merely operate better than firms that do not. They will take share from them, compress their margins, and in many cases acquire what remains.

The scarce capability in this phase is not access to AI. Access has diffused. The scarce capability is absorption: the ability to turn new technical capability into operating leverage faster than the frontier moves. The firms that build that muscle now become absorbent companies, compounding operating knowledge the next wave of capability can run on.

That is the compounding logic of the absorption gap. The firms that absorb faster do not merely get more from today's tools; the advantage builds on itself. Over time, they absorb the tools, then the workflows, then the margin, then the market.

Far enough out, the list of leaders in a market turns over, and the names that stay are the ones the frontier finds easiest to improve.