Emily Lewis-Pinnell, Driving AI Adoption at Evaila.
Stanford’s 2026 AI Index reports that 88% of organizations now use AI in at least one business function, and generative AI shows up in 70% of them. AI agent deployment, meanwhile, remains in single digits across nearly every function. Adoption is nearly universal. Transformation is rare. And the gap is now showing up in earnings conversations.
I shared an article earlier this year, arguing that winning with GenAI requires two tracks running together: broad enablement to build organizational AI fluency, and deep focus in one domain to produce transformation. That framework still holds. What has changed is which track leads. Agents change the sequence.
The Broad Track Worked, Then Hit Its Ceiling
Give leaders credit. The enablement playbook was executed at remarkable speed. Co-pilots rolled out, training programs launched and employees across functions learned to draft, summarize and analyze with AI. Organizations built the muscle.
Most never put it to work. McKinsey calls this the GenAI paradox: Nearly eight in 10 companies have deployed generative AI, and roughly the same share reports no material impact on earnings. The diagnosis matches what we see inside organizations every week. Horizontal tools such as co-pilots scaled quickly because they were easy to distribute, but their benefits spread thinly across thousands of individual tasks. The vertical use cases that reshape how a function operates rarely make it out of the pilot phase.
PwC’s 29th Global CEO Survey put a number on the consequence: 56% of 4,454 CEOs surveyed said their companies have realized no significant financial benefit from AI to date.
Read those numbers carefully before concluding that AI underdelivered. Co-pilots made individuals faster at existing tasks. They were never designed to change how work flows through a function. The broad track did what it was built to do: It created fluency. Fluency was the prerequisite. It was never the payoff.
Agents Change The Unit Of Value
With chat-based tools, the unit of value is the individual task, and the cost of a bad output is low. A weak draft gets edited. A wrong summary gets caught. Broad distribution made sense because every employee could capture small gains safely.
Agents invert both conditions. The unit of value becomes the workflow, because an agent’s usefulness comes from executing multistep processes across systems. And autonomy raises the cost of error, because outputs turn into actions. An agent that misfires books the wrong shipment or approves the wrong invoice.
That inversion is why the enablement playbook cannot simply be rerun with agents. Fluency can be democratized. Autonomy cannot. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Those are the failure modes of organizations distributing agents the way they distributed co-pilots: broadly, optimistically and without redesigning the work underneath.
The organizations seeing returns pick one bounded workflow, redesign it end to end around what the agent can reliably do, define what remains human and measure the before and after with precision. PwC’s data shows the same pattern: The 12% of CEOs reporting both revenue and cost gains are two to three times more likely to have embedded AI extensively into specific products, services and decisions.
Leaders Are Measuring The Wrong Track
A second problem hides inside the ROI numbers, and it belongs to leadership. Deloitte’s 2026 pulse research, drawing on nearly 3,700 professionals, found that 48% of organizations have introduced AI without redesigning the workflows or roles it sits within. Only 12% report redesign at scale. Deloitte’s conclusion is one every executive team should sit with: Adoption metrics and transformation metrics are not the same.
Seats provisioned, logins and training completions measured the broad track well. Leaders then carried those dashboards into the agent era and expected them to show business results. When the dashboards showed activity and the P&L showed nothing, many concluded AI wasn’t working. The measurement system failed before the technology did.
This is the reallocation risk I flagged in January, now visible at survey scale. Time saved that is never redirected becomes organizational slack, and slack is invisible in financial statements. If you are not capturing a baseline before an agent goes live, you have already decided not to know whether it worked.
What To Do Now
To help maximize agentic potential:
Keep the broad track running, and change its job.
Fluency is now table stakes. The new purpose of broad enablement is surfacing agent candidates: The people closest to the work know which workflows are bounded, repetitive and measurable. Build the channel that moves those observations to the teams deciding where agents go.
Pick one workflow and redesign it end to end.
Choose a process with structured inputs, established rules, measurable outputs and short feedback loops. Assign a named owner and a single success metric. Capture baseline data before deployment, and resist expanding until the metric moves.
Set decision rights before the agent acts.
Define what the agent may do alone, what requires human approval and who owns the outcome. Governance designed this way makes speed sustainable. Retrofitting it after an incident costs far more.
Measure transformation, and retire the adoption dashboards.
Report cycle time, quality and outcomes against your baseline. Stop presenting usage data to the board as evidence of value.
Sequence matters in both directions
None of this is an argument for shutting down broad access. Organizations that skipped the fluency phase and jumped straight to agents are struggling too, because a workforce that has never worked with AI cannot supervise it or spot its failure modes. Broad enablement built the capability. Deep focus is where it compounds.
In January, the dual strategy was a question of balance. In the agent era, it is a question of sequence. The question for leaders heading into 2027 is no longer whether your people are using AI. It is whether you have chosen the workflow where AI stops assisting and starts operating, and whether you would know how to measure the difference.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
