Where AI Is Heading: The Absorption Gap Is the New Bottleneck

The Two Curves

There are two curves that matter for AI strategy. Most discussion blurs them.

The capability curve measures what AI can do. METR's research — the most rigorous public measurement of frontier AI agent capability — finds the length of tasks AI agents can complete autonomously has doubled approximately every 7 months for six years[1]. AISI's internal data, updated in February 2026, shrunk that doubling time to 4.7 months[8]. Claude Opus 4.5 now has a 50% completion horizon of nearly 5 hours on hard software tasks[2]. Inside Anthropic, internal merged pull-requests-per-engineer rose 67% on the back of agent adoption[9]. Sam Altman publicly stated his model is now smarter than him[14]; Dario Amodei has narrowed his "smarter than a Nobel laureate" prediction to 2027[3].

The capability curve is fast, broadly continuous, and roughly predictable. You can disagree with the timeline by a year or two. You cannot reasonably disagree with the direction.

The absorption curve measures what enterprises can actually deploy and sustain. The numbers there look nothing like the capability curve.

• 80%+ of AI projects fail to reach production[19]
• 42% of enterprises abandoned most AI initiatives in 2025, up from 17% the year before[12]
• 73% of AI customer service bots are eventually rolled back[5]
• 88% of organizations have AI in some business function, but only ~40% report any enterprise EBIT impact[11]
• 79% of enterprises face challenges adopting AI in 2026, with 54% of C-suite executives saying adoption is "tearing their company apart"[10]
• Of the $2.5 trillion forecast in AI spending in 2026, only an estimated 5% of pilots produce real value[7]

This isn't a capability problem. The same models that fail to deliver enterprise EBIT are the ones letting Cursor close 30%+ of their internal pull requests autonomously. The constraint isn't what AI can do. It's what organizations can absorb.

Why Absorption Is the Bottleneck

I've watched this pattern at scale. The capability gets demoed in a sandbox, the deployment runs into the operational reality, and the gap between the two is where the project dies. The recurring failure modes are organizational, not technical:

• Workflows weren't built for AI to do them. The enterprise process around an AI-generated artifact is the same process built for a human-generated artifact, complete with the same review steps, approvals, sign-offs. The result is AI-speed generation followed by human-speed bottlenecks. Net throughput barely moves.
• Data isn't ready. Gartner predicts 60% of AI projects will be abandoned by 2026 because the underlying data infrastructure can't support them[13]. AI doesn't work on PowerPoint architecture; it works on operational data.
• Validation is harder than generation. A METR randomized controlled trial found experienced developers were 19% slower with AI tools while feeling 24% faster[22]. If validators can't accurately judge AI output, "human-in-the-loop" stops being a safety net and starts being a rubber stamp.
• Org structure isn't designed for AI agents. Conway's Law applies. Agentic systems mirror the team structures that build them; an org with broken handoffs in human work will have broken handoffs in AI work.

Klarna is the canary case. They replaced 700 customer service roles with AI in 2024, then re-hired humans by May 2025[15]. The CEO's quote — "Cost was too predominant a factor; what you end up with is lower quality" — is the absorption story in one sentence. The capability worked. The deployment didn't.

What I Predict for 2026–2031

These are the calls I'm willing to defend. Each is specific enough to be wrong about.

1. Capability outruns business absorption by mid-2027

By the end of 2027, frontier agents will reliably complete day-long autonomous software tasks. By 2030, multi-day. Enterprise deployment will be lagging by 18–36 months for the median company. The companies that close that gap will compound disproportionately.

2. The CAIO role bifurcates

The CAIO role exploded — 76% of large orgs had one by 2026, up from 26% in 2025[6]. By 2028, the role will split. Most "CAIO" titles today are doing capability strategy. The harder, scarcer, better-paid version will be operations: redesigning workflows, killing PoCs, shipping fewer projects but at deeper integration. Watch for a new title — Chief AI Operations Officer, Chief Transformation Officer, or something close — focused entirely on absorption.

3. The second-generation walkbacks arrive — but not where you expect

The first wave of walkbacks (Klarna pattern) hit customer-facing functions. The next wave, 2027–2029, hits engineering, legal, and finance — functions that auto-generate output where validation is harder than humans realize[16]. Expect public stories of code generated by AI causing significant production incidents at named companies, followed by sober internal reports nobody reads.

4. The single-person billion-dollar company arrives — but as an outlier, not a category

Amodei has given the first single-person billion-dollar company a 70–80% probability "soon"[18]. I'd take the over on it happening, the under on it becoming a category. It will be an extreme outlier, not a template. Most billion-dollar companies will still need 50–500 humans, but with a fundamentally different shape than today's org charts.

5. The frontier of business-and-tech moves from "AI as feature" to "AI as architecture"

Most current enterprise AI is bolted onto existing products. The companies that rebuild their stack as AI-native — agent-to-agent orchestration, real-time context graphs, eval as a core engineering discipline — will compound 10x faster than retrofitters[20]. By 2030, the gap between AI-native and AI-bolted will be unbridgeable for most incumbents.

6. A category I'm watching: agent-to-agent commerce

Two AI agents negotiating a contract on behalf of two companies, settling within seconds, instrumented for audit. The technical pieces exist. What's missing is the trust framework and the legal infrastructure. By 2030, this category exists, materially.

Counter-Perspective: "You're Underweighting Capability"

The strongest opposing view comes from the optimist side, not the skeptic side. It runs like this: "You're treating capability as 'on a curve.' But the curve has a discontinuity. When AI becomes genuinely smarter than humans across most domains — what Amodei calls 'powerful AI'[4] — absorption becomes trivial because the system reorganizes itself. You don't need humans to redesign the workflow; the AI does."

It's a serious argument. Two reasons I'm not persuaded yet.

First, the absorption gap exists today even when AI is provably capable. Cursor's cloud agents close 30% of merged PRs internally. Most enterprises with the same tools see no movement on engineering throughput[17]. The bottleneck isn't model intelligence — it's everything around the model. Smarter models don't fix legacy data architectures, regulatory regimes, or three-decade-old change-management practices.

Second, the historical precedent isn't kind to discontinuity arguments. Cloud computing was a capability discontinuity. PC, mobile, internet — each one promised to reorganize the enterprise. Each took 10–15 years to actually do it. The lag between capability and absorption is the most reliable pattern in technology history. Maybe AI breaks the pattern. That's a possibility, not a base case.

The honest version of the optimist view: by 2030, some enterprises will be running on AI-native architectures so different from today that comparing them to current orgs makes no sense. I agree. The number of those enterprises will be small. The gap between them and everyone else will be the most important business story of the decade[21].

The capability curve is the easy curve. Operators who only watch it will be surprised by 2028. The absorption curve is the curve that decides which companies are still here in 2030.

What Changes My View

A genuinely useful Perspective should commit to its own falsifiability. Here's what would update me toward "capability is the binding constraint after all":

• METR-style benchmarks plateau — if autonomous task duration stops doubling, the capability curve breaks. Possible but no evidence yet[24].
• A reproducible AI-native operating model emerges — if a small number of companies show they can build AI-native and scale it, the absorption argument weakens because the playbook becomes copyable.
• The walkback rate falls dramatically by 2028 — if enterprises stop abandoning AI projects at 40%+ rates, my "absorption is the bottleneck" call fades.
• A regulatory shift forces architectural change — if EU AI Act enforcement (or a US equivalent) makes "bolt-on AI" non-compliant, every enterprise has to rebuild whether they were ready or not.

I'm watching all four.

What This Means for Operators

If you're running an enterprise AI program: stop measuring capability, start measuring absorption. Track the rate at which AI-generated work makes it through your full operational pipeline (review, deployment, monitoring, iteration). Track the failure mode of every walked-back project. Hire for absorption talent — change managers, ops architects, eval engineers — not just AI talent. Resist the temptation to chase the latest model. The model that wins isn't the most capable; it's the one you can actually integrate.

If you're a board director or investor: ask the absorption question first. "How many of your AI initiatives have you killed in the last 12 months?" A healthy answer is a number greater than zero. If a company hasn't killed any, they haven't been honest about what's working.

The next five years aren't going to be won by the smartest AI. They're going to be won by the operators who turn capability into compound business outcomes. The frontier of business and tech is the frontier of absorption.