Adoption sprints, EBIT stalls: a scaling-ROI gap

Enterprise AI’s 2026 data tells a counterintuitive story: deployment is sprinting, returns are standing still. In McKinsey’s August annual survey, companies with over $1 billion in revenue pushed “scaling AI agents in at least one business function” from 27% to 40% in a year; in the same data, the share reporting any EBIT impact from AI has been stuck at 37% for two years, and only 6% qualify as high performers (at least 5% of EBIT attributed to AI). The gap between adoption and return is widening. Our judgment: this is not a technology problem, it is a conversion problem — individual productivity does not automatically flow into the income statement.

The numbers: deployment sprints, profit stalls

McKinsey’s “The State of AI in 2026: On the road to ROI” (August 25, 2026, 1,719 respondents) delivers the core figures (source: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ; analysis: https://bigaiagent.tech/ai-agent-scaling-gap-2026-mckinsey-survey/ ):

Scaling jumped from 27% to 40% — but only among large enterprises; smaller companies did not move, holding flat at 22%. The share reporting any EBIT impact is 37%, unchanged for two years. High performers are 6%, also unchanged. About one in five companies says operating costs — token spend being the visible line — are already constraining AI usage. Eighty percent report AI improved their personal productivity and half say it improves their decisions — yet individual gains are not becoming enterprise numbers. Thirty-two percent skipped buying a software product because agentic coding tools let them build it in-house; among high performers that share nears 47%. Expectation of AI-driven headcount decline rose from 32% to 39%, while actual reported reductions over the past year were just 14% — conviction is running ahead of evidence, on both cost savings and headcount.

What high performers did: workflow redesign, not model choice

The sharpest dividing line in the McKinsey data has almost nothing to do with models or platforms: nearly three-quarters of high performers report fundamentally redesigning their workflows around AI, versus about a quarter of everyone else. High performers are more than three times as likely to be scaling agents across most business functions, and more than twice as likely to spend over 15% of their IT budget on AI (source: https://bigaiagent.tech/ai-agent-scaling-gap-2026-mckinsey-survey/ ).

Salesforce’s survey of 2,025 decision-makers (double-blind, fielded May 14-28, 2026, across 20 countries) cuts the same conclusion differently: the three most predictive success factors are “clean, accessible data” (36%), “narrowly scoped use case” (36%), and “human escalation paths defined before launch” (35%) — all ahead of model quality, platform choice, and a unified orchestration layer (about 30%). Deployers reached positive ROI in about eight months on average, with 53% employee adoption and a 29% lift in customer satisfaction. Professional and business services were among the slowest to deploy yet the fastest to meaningful ROI (6.5 months); high tech is one of the biggest deployers yet one of the slowest to returns (10.1 months) — being first to deploy does not mean first to returns (source: https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/ ).

Our judgment: these numbers point to one thing — returns come not from “a stronger model” or “more agents” but from “which workflow got rebuilt.” Bolting agents onto old processes caps the return. High tech is the footnote: it is best at deploying and most habituated to bolt-on adoption, and it is the slowest to convert.

The gap is a missing conversion layer

80% individual productivity gains, 37% enterprise EBIT impact — the conversion mechanism between them does not exist. We call this missing mechanism the conversion layer, and it has three components:

First, workflow redesign. Draw workflows around what agents can actually do, instead of replaying every step of the old process with an agent. Second, data readiness scoped to the use case. Salesforce: only 31% of deployers fully unified data before launching (7.3 months to ROI); 69% reached ROI with iterative integration (8.8 months). Assemble the data one high-value use case needs and ship — do not wait for enterprise-wide unification. “Waiting for perfect data” is the wrong gate. Third, operating costs in the budget. Forrester’s Microsoft-commissioned TEI study offers a cost structure worth copying into your budget model: roughly 25% of labor savings go to “paying” digital workers — subscriptions and consumption are explicit liabilities, not after-the-fact expenses. The composite organization saw $44.5M in benefits versus $20.2M in costs over three years: $24.2M NPV, 120% ROI, 15-month payback (source: https://tei.forrester.com/go/microsoft/agenticaisolutions/docs/Forrester_TEI_The_Total_Economic_Impact%E2%84%A2_Of_Microsoft%E2%80%99s_Agentic_AI_Solutions.pdf ; commissioned study — figures are a vendor-supplied composite model).

Our judgment: the missing conversion layer is the gap. Individual productivity is an input, not an outcome; without a conversion mechanism, “efficiency gain” reports will keep sitting between “80% of employees say AI helps” and “37% of companies see returns,” never converging.

The speed-versus-risk trade-off is real

The Salesforce survey carries an easy-to-miss pair of numbers: deployers averaged two governance structures (real-time monitoring, escalation frameworks, audit logs) before deploying and three after; lighter-governance organizations reached positive ROI in 7.2 months versus 9.3 months for heavier ones — but 32% of the lighter group (versus 18%) only discovered an agent operating outside parameters after a consequential error (source: https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/ ).

An industry counterexample footnotes the same lesson: Tata Steel deployed 300+ specialized agents in nine months (per Google Cloud’s account) with no outcome metrics — agent count does not prove value, and deployment speed does not prove risk is controlled (source: https://www.agenticism.co/post/luxury-retailer-scales-agentic-ai-across-workforce-with-high-volume-interactions-and-measurable-effi ).

Our judgment: ROI speed and risk visibility are not simultaneously available — that is a design trade-off, not an execution defect. The board’s “faster returns” and “more stability” are two columns of the same table, not a sequence. Write down what you are choosing before budget review: the faster the returns you want, the lighter the governance you accept, and the price of light governance is discovering roughly a third of out-of-bounds behavior after it happens.

Our judgment

First, replace “agent deployments” with “workflow rebuilds” as the return metric. Deployment counts are vanity metrics — Tata Steel shipped 300+ agents with no outcome numbers; EY’s fifty thousand agents matter because it built tooling that let business units self-serve agent creation, not because of the count. Second, the missing conversion layer is the root of the gap: personal productivity does not flow to the income statement by itself; it needs workflow redesign, use-case-scoped data readiness, and explicit operating-cost budgets. Third, data does not have to be perfect first: the 7.3-versus-8.8-month difference is smaller than the opportunity cost of waiting another year for unified data. Fourth, governance intensity and risk tolerance belong on the same table: the 7.2-versus-9.3-month ROI difference corresponds to 32% versus 18% discovery-after-incident — either choice can be right; pretending you can have both is wrong. Fifth, token cost is 2026’s silent ceiling: one in five companies is already usage-constrained by operating costs, and an agent plan without an operating-cost budget will hit that wall at scale.

A buyer’s action list

First, list every agent launched or scaled in the past six months and answer one question per row: “was its workflow rebuilt?” If more than half say no, that is your conversion-layer gap signal. Second, report individual-productivity metrics (the 80% category) separately from financial metrics. Do not let usage masquerade as return; the board reads the second column. Third, scope data readiness to the use case. Pick one high-value use case, assemble the data it needs, ship — 69% of deployers proved this path works. Fourth, put token and operating costs in the budget model explicitly. Use the Forrester 25% rule as a reference: about a quarter of labor savings goes to “paying” digital workers; then cap cost per agent. Fifth, put governance intensity and ROI speed on the same decision table. High-autonomy scenarios get audit logs and out-of-bounds detection; low-risk scenarios can iterate fast with light governance — but write the cost of each cell down.

The decision question left for the C-suite: in last month’s agent plan, which line was “rebuild a workflow,” and which was just “insert an agent”?

References: McKinsey “The State of AI in 2026: On the road to ROI” (2026-08-25) — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ; BigAIAgent analysis “AI Agent Scaling Gap 2026” (2026-09-05) — https://bigaiagent.tech/ai-agent-scaling-gap-2026-mckinsey-survey/ ; Salesforce “Agentic AI Study: Preparation Beats Speed for ROI” (2026-06) — https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/ ; Forrester “The Total Economic Impact of Microsoft’s Agentic AI Solutions” (commissioned by Microsoft) — https://tei.forrester.com/go/microsoft/agenticaisolutions/docs/Forrester_TEI_The_Total_Economic_Impact%E2%84%A2_Of_Microsoft%E2%80%99s_Agentic_AI_Solutions.pdf ; Enterprise Agenticism “Luxury retailer scales agentic AI across workforce” (2026-09-07) — https://www.agenticism.co/post/luxury-retailer-scales-agentic-ai-across-workforce-with-high-volume-interactions-and-measurable-effi