Agents are vertical. Organisations are horizontal
Human work flows sideways: across people, review loops, teams, and steercos. Each handoff in those horizontal workflows was designed for human speed and manual oversight. What used to require multiple steps, and handoffs to other apps, colleagues or teams, agents have collapsed into integrated, vertical tasks.
The mismatch creates an AI absorption and coordination "tax" now that agents transform AI into a capability that's producing more coherently, and faster than organisations can process, and integrate. People in existing horizontal workflows carry a new burden: reviewing, validating, and integrating agent output.
Can organisations adapt? Coordination, context, judgment, and organisational memory all need to land where horizontal work meets vertical agent capability.
A quick, practical diagnostic: People Per Process (PPP). Count the number of people a piece of work touches before completion. Low PPP signals vertical-ready work. High PPP reveals the horizontal organisation. Decompose who's there and why. Some people are load-bearing (core function, real expertise, real approval rights). Other people aren't (awareness, CYA, legacy inclusion). Increasingly, agents' vertical task integration will expose these PPP tensions.
Is the absorption gap compounding? Here's what the research agents are finding.
-
Artificial Analysis·evidence
The leading frontier model fully satisfies every requirement on only 3% of multi-week knowledge-work tasks.
The first independent, non-vendor evaluation of frontier models on realistic multi-week knowledge work scores them against a rubric built to catch output that looks polished but is wrong, and the current leader satisfies every requirement on just 3% of tasks. Vertical agent output still lands on a person to finish, check, and integrate before it is usable. That puts a measured ceiling on the absorption tax Layer 1 names: capability produced faster and more coherently than the horizontal organisation can review and integrate.
-
Satya Nadella, Microsoft·illustration
Human capital appreciates as token capital grows: a firm's durable IP is a learning loop on top of models, not the models themselves.
Nadella separates human capital (knowledge, judgment, relationships, pattern recognition) from token capital (the AI capability a firm builds and owns) and argues the first appreciates as the second grows, direct CEO-level pushback on experience-erosion pessimism. The operational core is a learning loop on top of models (private evals against business outcomes, RL on internal traces, a queryable knowledge base) that becomes the firm's new IP and survives a model swap without losing the "company veteran" expertise. The most senior external voice yet to state the owned-intelligence and memory-loop thesis the diagnosis rests on, though Microsoft has a commercial interest in "the durable layer is the ecosystem, not the model" being true.
-
Anthropic Institute·evidence
More than 80% of the code merged into Anthropic's production codebase is now written by Claude.
The shift happened without the firm stalling under review burden: engineers now ship roughly eight times more code per quarter than they did across 2021 to 2025. That makes it the clearest data point this window that the bottleneck is absorbing agent output rather than generating it. The horizontal work of validating and integrating what agents produce is the binding constraint the diagnosis names.
-
Demirer, Musolff & Yang (SSRN)·evidence
Autonomous coding agents lift commits 180%, but the gain attenuates sharply at the project and release stages downstream.
The lift does not survive the production chain: the cumulative commit gain falls to 50% at the project level and 30% at actual releases.
-
Azeem Azhar, Exponential View·evidence
Individual AI speed does not reach the firm's bottom line until the organisation rewires around decision speed between workflows.
Azhar's three-stage ladder is organised by logic rather than capability: Stage 1 speeds individuals, Stage 2 speeds workflows (today's agent-bolted-to-process pattern), and Stage 3 speeds the firm through closed-loop autonomy.
-
Anthropic·illustration
Maintainers asked Anthropic to slow vulnerability disclosures because human triage and patching cannot keep pace with the model's find-rate.
-
Krishna Rao, Anthropic CFO·evidence
Anthropic's enterprise customer spend multiplied five-fold in twelve months.
-
Yu et al., arXiv cs.MA·evidence
EntCollabBench: role-specialised agents struggle with delegation, context transfer, and workflow closure under realistic enterprise constraints.
-
Aaron Levie, Box·evidence
Cloud took 15 years to diffuse 1,000×. The same shape calibrates the AI absorption timeline.
-
Azeem Azhar, Exponential View·evidence
Citadel data: jobs in highly AI-exposed occupations are rising, not falling.
Software engineering postings up 18% YoY; customer service and accountancy also rising. Azhar reads the data through three mechanisms: complementarity, supervision overhead, demand expansion. Meaningful pushback on the simple substitution narrative; needs a Layer 1 reading that allows for compounding gain without proportional headcount compression.
Contested · evidence · 19 jun 2026Challenger, Gray & Christmas attribute roughly 40% of US May job cuts to AI, echoed in the 19 June Azhar roundup. The fresh layoff data cuts against the "rising, not falling" reading on the same AI-exposed occupations, so this card is held contested pending occupation-level corroboration.