I have spent eight years building remote engineering teams, and the sharpest question I hear this year is not about which model to adopt. It is whether my team is actually using what it already has. There is a wide gap hiding inside that question. McKinsey State of AI 2026, published in August 2026 from a survey of more than 1,700 executives across 97 markets, finds that 80% of people say AI improved their individual productivity. The same report says only 6% of organizations qualify as AI high performers. The distance between those two numbers is not talent. It is leverage.
Leverage is the distance between one person doing something faster and a whole team doing something different. McKinsey says 44% of organizations now scale AI enterprise-wide, up from 38% the prior year. Agentic AI scaling at large organizations hit 40%, up from 27%, while small organizations stayed flat at 22%. Coding agents reached about 31% among large enterprises. Yet the share of organizations reporting an EBIT impact held flat at 37% for a second consecutive year.
So the tools arrived, the usage arrived, and the margin line did not move. My hypothesis after eight years of watching this: usage without structure is just faster noise. A team needs a level model, a governance layer, and a review cadence. This post gives you all three.
Why the gap between usage and impact exists
Usage is individual; impact is structural
An engineer who drafts faster still has to review, ship, and get paid for the same sprint. If every teammate gets a small personal gain, the team total gains proportionally and then plateaus at the sprint planning boundary. Structural impact only appears when a workflow itself is re-shapeable, which means someone has to redesign it on purpose.
80% say AI improves individual productivity. 37% say AI moved EBIT. That 43-point spread is where every dollar of AI budget is quietly disappearing.
Tool sprawl is real overhead
Forty percent of IT executives in the McKinsey survey said AI tool proliferation is harming collaboration and efficiency. I see it in my own team. Three code assistants, two meeting summarizers, one doc pipeline. Each one works. Together they add review overhead, because every artifact now arrives half-finished and needs a human second pass.
Cost quietly caps the ambitious teams
About 20% of organizations said they were constrained in AI use because of operating or token cost, and 28% now spend over 10% of their IT budget on AI. Meanwhile 39% expect AI-related workforce declines, up from 32%. The pressure is coming from both ends. You cannot just run harder. You have to run deeper on fewer workflows.
The five levels of AI leverage
I defined these from behavior, not from licenses. A team at level four can look exactly like a team at level two in a tools inventory. You can only tell them apart by watching what happens when the owner is offline.
L0 No AI
The team has no defined AI use. Prompts happen privately, inconsistently, and are never reused. There is no shared library and no budget line.
L1 Individual assistant
Teammates prompt their way through docs, code, and tickets. Time saved is real but personal. Observable signals: a shared prompt library exists but is not versioned, and quality depends on who is at the keyboard.
L2 Team workflow
One repeatable workflow runs end to end with AI. Ticket triage, release notes, or support draft response, for example. The output is stable enough that a new teammate can run the same workflow on day one without asking questions.
L3 Automated agents
An agent acts on a queue with a human approval gate. It files the ticket, updates the issue, and routes the review. Observable signals: there is a written rollback trigger, and there is a spend ceiling on the automation.
L4 Governed scale
Multiple workflows run under one registry, each with a named owner, a token budget, and one business metric tied to it. Reviews happen monthly and changes happen on purpose.
Observable behaviors, failure modes, and the governance each level needs
This is the table I bring to a team review. It is deliberately blunt about failure modes.
| Level | Observable behavior | Typical failure mode | Required governance |
|---|---|---|---|
| L0 No AI | AI use is ad hoc or absent | Opportunity cost, vendor seats keep growing | None in place |
| L1 Individual assistant | Teammates prompt in docs, code, and tickets | Inconsistent quality, no shared prompts | Data handling rule, one shared prompt library |
| L2 Team workflow | One repeatable workflow runs end to end with AI | Works for one person, breaks on handoff | Named owner, per-workflow token budget |
| L3 Automated agents | Agents work a queue behind a human approval gate | Silent regression, unbounded token bill | Approval gate, rollback trigger, spend ceiling |
| L4 Governed scale | Multiple workflows reviewed monthly against a business metric | Tool sprawl, orphaned automations | Central registry, quarterly owner review |
Move exactly one level at a time. Teams that try to jump from L1 to L4 in a quarter end up with automations nobody owns.
The governance layer that makes scale safe
One accountable owner per workflow
If no single person is on the hook for a workflow’s quality, the workflow decays. I require a named owner in the registry row, and the owner is the person who writes the rollback trigger.
A token budget per workflow, not per team
Team-level budgets get gamed. Per-workflow budgets are honest. Set a monthly ceiling, choose the cheapest model that clears your quality bar, and route heavier runs to a larger model only when a specific gate fires.
A human approval point for anything that ships externally
Agents can draft. They should not release. The approval gate is where quality, tone, and compliance get checked, and it is also where your audit trail starts.
An audit trail you can actually read
Every agent run needs a durable record: input, model, output, approver, and timestamp. When something regresses silently, this is the only thing that tells you when it started.
The monthly leverage review
Two pages, thirty minutes, every month. I keep it to three lines of content.
- Maturity level per workflow, and whether anything moved.
- Governance compliance: owner present, budget respected, approval gate used, audit trail intact.
- The one business metric tied to the workflow, with the delta from last month.
When something breaks governance, it goes on the same page as the win. Teams that hide governance failures tend to break the same workflow twice.
How to move up without buying anything new
The single most important finding in the McKinsey data for this framework is that concentration beats breadth. Small organizations scaled agentic AI at a flat 22% while large organizations moved to 40%. Size did not help. Focus did.
Pick one workflow with real volume and a measurable outcome. Drive it to L2 first, because L2 is the level where the workflow becomes independent of who runs it. Only then move to L3, where an agent takes the queue and a human reviews. Only then widen.
I have watched teams go from L1 to L4 in six months using only the tools they already had. What changed was not the tooling. It was the owner, the budget, the review cadence, and the written rollback trigger.
The 6% of AI high performers are not running better prompts than the rest of the market. They are running fewer workflows with more structure on each one. Five levels, one workflow at a time, a budget that caps token spend, and a monthly review that ties the whole thing back to a business metric. That is the framework. The gap between 80% individual productivity and 37% enterprise impact closes when a team stops measuring prompts and starts measuring leverage.