From Pilots to P&L

Most companies no longer have an AI adoption problem. They have an AI accountability problem.

• Adoption is close to universal. Roughly 88% of organizations now use AI in at least one function (McKinsey, The State of AI, Nov 2025).
• Confidence is not. Only about 39% of leaders are even confident their AI investments will improve financial performance (Gartner, 2026), and far fewer can prove it.

That gap between usage and the P&L is where every serious AI conversation now sits.

The winners of the next 24 months will not be the companies with the most pilots or the biggest model budgets. They will be the ones that:

• redesign the work itself,
• manage agents like labor,
• reprice what they sell, and
• run AI through the same financial discipline as any other capital program.

Below are five topics where a firm, defensible point of view wins the room, with the evidence under each.

The funnel, in one view. The distance from "we use AI" to "it moved the P&L" is not a rounding error. It is the whole strategic gap:

Every step down is a place where conviction ran ahead of evidence. The rest of this piece is how to climb back up.

1. The ROI reckoning: AI needs a value ledger, not an innovation dashboard

The take: the AI ROI problem is not a model problem. It is a management-accounting problem, which makes it a CFO problem.

Spending is speeding up into the accountability gap, not away from it:

• Budgets are doubling. BCG's AI Radar 2026 found companies plan to roughly double AI spend this year, to about 1.7% of revenue.
• Investment continues without returns. 94% intend to keep investing even without near-term returns, and half of CEOs believe their job depends on getting AI right.
• Most pilots show nothing. MIT's widely cited GenAI Divide study put the share of initiatives with no measurable P&L impact near 95%. Treat that figure as directional, since its method is contested.
• Confidence agrees. Gartner's 39%, the share of technology leaders even confident AI will improve financial performance, tells the same story more politely.

The fix is not better models. It is three finance habits:

• Size the value first, by where the money is, instead of keeping a list of use cases.
• Measure unit economics per workflow, instead of making enterprise-wide claims.
• Keep a benefits ledger with the same sign-off discipline finance applies to synergies in a deal.

If a benefit cannot survive an FP&A review, it is not a benefit. The full argument, covering value cards and how to fund and stop agents by the numbers, is in The Harness Era: Fund Agents Like Value Streams.

2. Agents are a new class of labor, not a new layer of software

The take: stop managing agents like applications. Manage them like a workforce, with role definitions, access rights, supervision, performance measures and rules for shutting them down.

The numbers show how far deployment runs ahead of redesign:

• Scaling is narrow. McKinsey finds 23% of organizations scaling agentic AI somewhere, but no more than 10% scaling agents in any single function.
• Use is not production. 50–75% of organizations say they are using agents (KPMG, Forrester), yet only about 17–19% run them in production (Gartner, Snowflake, Databricks).
• Autonomy and governance are rare. 7% run fully autonomous agents (Bain), and just 21% have a mature governance model for the ones they have (Deloitte).
• The work has not changed. PwC's Agent Survey shows 79% of executives say agents are being adopted, yet only 45% are rethinking operating models, 42% are redesigning processes and 34% use agents in finance at all. Deloitte adds that 84% of companies have not redesigned a single job around AI.

Agents dropped into unchanged workflows produce demos, not results. The control document that makes agent accountability assignable (owner, tools, risk tier, tests, link to value) is in The Harness Era: The Agent Card.

3. The billable hour is melting: sell assets, not effort

The take: AI's most under-covered disruption is not inside the enterprise. It is the pricing model of everyone who advises the enterprise. When 60 hours of analysis becomes six, time stops measuring value, and every hours-based business faces the same choice: reprice or shrink.

This is no longer theoretical:

• McKinsey reportedly earns more than 30% of global fees from outcome-tied pricing.
• BCG's leadership has said AI-related work would grow from roughly 20% of 2024 revenue toward 40%.
• EY leaders openly describe a "service-as-software" future.
• Microsoft launched Frontier Company on July 2: $2.5B and 6,000 embedded experts, explicitly an "outcome-driven engineering organization."
• AWS stood up a $1B embedded-engineer unit two days earlier.

The firms that turn expertise into reusable assets they own and can price (ontologies, diagnostics, agent suites, benchmarks) will compound. The firms that keep selling hours will hand their productivity gain to the client and watch platform vendors sell outcomes over their heads.

4. Human-in-the-loop is becoming the industry's favorite checkbox

The take: "human in the loop" is on its way to being the most over-claimed control in enterprise AI. A person who lacks the authority, context, time and tools to spot, challenge and stop an agent is not a control. They are an audit finding waiting to happen.

The data supports the skepticism, and now so do the regulators:

• The Bank of England. On June 30, Deputy Governor Sarah Breeden told the ECB's Sintra forum that existing frameworks "were not built to contemplate autonomous agents" and that relying on a human in the loop for every agent action "is unlikely to be realistic." She also floated market-wide circuit breakers for AI-driven trading.
• The FSB. Its June consultation recommends treating agents as "synthetic employees" with boundaries, approval gates and audit trails.
• McKinsey. High performers are nearly three times more likely to have defined human review of AI output (65% vs 23%). Its AI Trust work found roughly two-thirds of organizations cite security and risk, not technology, as the top barrier to scaling agents.

The board question is shifting from "is a human in the loop?" to "can that human actually step in, and can you prove it?" That is ICFR thinking applied to agents, and it is coming for every autonomous workflow.

5. The EU moved the deadline. It did not move the obligation.

The take: the loudest AI-regulation headline of 2026, the EU's Digital Omnibus delaying high-risk AI rules, is being read backwards. A delay is not relief; it is time to scope the work. Companies that treat it as a pause will reach December 2027 with no inventory, unclassified systems and a queue at the notified bodies.

What actually moved, and what did not:

• Moved: Annex III high-risk obligations go from August 2, 2026 to December 2, 2027. Annex I embedded systems go to 2028.
• Already in force: general-purpose-AI obligations have applied since August 2025.
• Still on schedule: most transparency obligations still land August 2, 2026.
• Watermarking: generative-content marking must be in place by December 2, 2026 for systems already on the market. New systems mark from August 2.

Governance built into delivery scales. Governance parked in a policy PDF does not. Use the extra time to build the controls, not to postpone them.

The bottom line

One funnel frames the whole agenda: 88% adoption, 39% confidence, 7% with ROI they can prove. Everything above is a way to close that gap:

• Measure AI like capital.
• Manage agents like labor.
• Price outcomes, not hours.
• Make controls real, not ceremonial.
• Use the regulatory delay to build, not to wait.

The model was never the moat, and the pilot was never the point. The pilot era is over. The reckoning has started.