The Next Value Wave — and the Paradox

Start with the uncomfortable headline every board should sit with: near-universal adoption, negligible bottom-line impact.

McKinsey frames it as the "gen AI paradox": roughly 80% of companies use generative AI, yet roughly 80% report no material impact on the P&L (McKinsey, Seizing the Agentic AI Advantage, 2025). The broader State of AI survey puts adoption near-universal — ~88% of organizations use AI in at least one function — but only ~39% report enterprise-level EBIT impact, and most of those attribute under 5% of EBIT to it (McKinsey State of AI, Nov 2025).

This is not a one-firm finding. Independent research converges on the same gap:

The diagnosis: the problem is almost never the model. MIT NANDA names the root cause the "learning gap" — flawed workflow integration and process fit, not technology quality. If a pilot isn't moving a number, look at the workflow, not the model.

(All figures are time-stamped survey findings that refresh roughly annually — treat them as a 2025–2026 snapshot and re-verify against the live source before quoting.)

If adoption is everywhere and value is nowhere, where did the value leak out? McKinsey's answer is the heart of this lesson: most spend went into shallow, horizontal copilots — broad assistants that make many people a little faster at many tasks — when the prize sits in deep, vertical agentic workflows that reengineer a specific core process end-to-end.

The distinction is the whole story:

Copilots make existing tasks faster — incremental productivity that's genuinely real but spreads too thin to surface on the income statement. Agents can do a whole workflow end-to-end — a step change that requires more redesign and more governance, but lands where value concentrates.

This is why the same research that documents the paradox also documents the cure: McKinsey finds workflow redesign has the single biggest effect on whether AI moves EBIT — yet only about a fifth of firms have fundamentally redesigned a workflow; most simply bolt AI onto a legacy process and capture little.

The corrective McKinsey proposes is not "add more agents." It is to stop treating AI as a feature bolted onto today's org chart and start reinventing core processes around human + agent coworkers — what McKinsey calls "the agentic organization."

The logic chains together:

1. Chat and copilots gave broad but shallow gains — useful, but they don't reengineer the process, so they don't move the P&L.
2. Agents execute whole workflows, which is what makes them an operating-model change rather than a productivity feature.
3. The value only materializes when the workflow is redesigned around the agent — not when an agent is dropped into an unchanged legacy process.

McKinsey's prescription for getting there (its "four pillars") is deliberately the opposite of "agents everywhere":

The recurring governance principle underneath all of it: automation for execution, humans for judgment. Agents run the steps; people own the decisions, the exceptions, and the accountability.

Before championing agents internally, calibrate to reality — agent adoption today is early, concentrated, and unevenly proven. The hype is running ahead of the reliability, and a clear-eyed leader is more credible than an evangelist.

What the time-stamped evidence shows:

Where agents actually work today is narrow and structured: the most common enterprise use cases are IT / service-desk triage and resolution and knowledge "deep research" agents that pull from internal sources (McKinsey, 2025). Production agents "tend to be simple and human-supervised" (Anthropic) — reliable for bounded, well-structured, lower-stakes tasks, not autonomous "set-and-forget" digital employees.

Two cautions for the boardroom. First, "agent washing" is rampant — Gartner estimates only ~130 of thousands of self-styled "agentic" vendors are genuine; the anti-hype test is autonomous reasoning + tool orchestration + persistent context. Second, data and integration readiness gates everything — about 8 in 10 firms cite data limitations as the roadblock to scaling agents.

There is a bigger reason agents are positioned as the next wave and not just a better copilot — and it's an economic one, not a technical one. The investor framing (a16z) is "services as software" / digital workers: the unit of sale shifts from software seats to work delivered.

The distinction reshapes the size of the prize:

Instead of licensing a tool that a person uses to do the work, an agent (internal or vendor) delivers the outcome the work used to produce — a reconciled invoice batch, a reviewed contract, a researched supplier comparison. This is why a16z frames agents as a far larger market than SaaS: they go after the services and labor budget, not just the software budget.

For an executive, this reframes two things at once:

• The opportunity — AI stops being an IT cost line and becomes a lever on the largest cost in most businesses: people and outsourced services.
• The risk and the responsibility — "digital workers" sit under human oversight, not instead of it. The augmentation narrative (work redesigned around human + agent, not wholesale replacement) is both the realistic near-term path and the defensible one.

You don't need to evaluate the technology to govern it well — you need to ask the questions that separate a real, value-bearing agentic bet from a shallow copilot rollout or an "agent-washed" demo. Bring these to the table:

On the value thesis

• Is this a tool rollout or a workflow redesign? (Only the second reliably moves EBIT.)
• Which one core process does it reengineer end-to-end — and what P&L number do we expect to move?
• Are we concentrating on 2–3 vertical use cases, or spreading thin across scores of pilots?

On readiness

• Is our data clean, governed, and accessible enough to feed this? (8 in 10 firms stall here.)
• Have we redesigned the workflow, or are we bolting an agent onto a legacy process?

On "is it even a real agent"

• Does it show autonomous reasoning + tool orchestration + persistent context — or is it a rebranded chatbot? (The anti–agent-washing test.)

On governance and measurement

• What is our human-in-the-loop policy for a system that acts, and can we halt or roll back?
• Are we tracking quality (CSAT/NPS, error rates), not just volume? (The Klarna lesson.)
• What is our plan if this pilot is one of the >40% Gartner expects to be canceled — kill it, or reshape it?

The throughline: champion agents where the workflow is redesigned, the data is ready, the value is concentrated, and a human owns the judgment. Everywhere else, you are funding activity, not impact.