Adoption is widespread, but scaled impact is rare — and the difference lives in how work actually gets done.
Blake Aber · Predicate Ventures · 2026
The adoption number hides the real problem
Most companies now use AI somewhere. A large share of organizations report regularly using AI in at least one business function, according to McKinsey's 2025 State of AI survey.
That statistic gets quoted often. It is also misleading if you stop there.
Using AI in one function is not the same as running AI workflows across an organization. The first is a pilot. The second changes how output gets produced. Most companies are still in the first stage.
McKinsey found that only a minority of respondents report having begun to scale AI across the enterprise, with nearly two-thirds not yet scaling. An even smaller group reports that AI has been fully scaled across their organization.
So the picture is not one of slow uptake. It is one of wide experimentation paired with shallow integration.
What an AI workflow actually is
An AI workflow is a sequence of steps where AI handles defined tasks inside a process that produces a business result. The key word is process. A prompt is not a workflow. A chatbot tab open in a browser is not a workflow.
A workflow has inputs, handoffs, checks, and an owner. It connects to systems the business already runs — the CRM, the ticketing queue, the document store, the billing engine.
When AI lives inside that structure, it produces measurable output. When it lives beside it, in a separate tool people open occasionally, it produces anecdotes.
This distinction explains the McKinsey gap. Experimentation is cheap and individual. Workflow integration is expensive and organizational. Companies cross the first threshold easily and stall at the second.
The EBIT signal
The survey offers a useful tell on where value concentrates. Only a share of respondents report enterprise-level EBIT impact attributable to AI.
That group is small. It also overlaps heavily with the companies that have moved past pilots into scaled deployment.
The inference is straightforward. Financial impact follows workflow integration, not tool access. Buying model credits does not move EBIT. Rebuilding how a function operates does.
This is the part most buying decisions skip. Procurement evaluates the model or the vendor. The return depends on the process redesign that surrounds it, which procurement rarely scopes.
Agents change the unit of work
The 2025 survey, which polled a sizable respondent base over its field period, found that a large share of organizations are at least experimenting with AI agents.
Agents matter for workflows specifically because they change what a single step can contain. A traditional automation rule executes one instruction. An agent can take a goal, decide intermediate steps, call tools, and check its own output.
That shifts the design question. Instead of mapping every branch of a process, you define the outcome and the constraints, then let the agent fill the middle.
The risk shifts too. An agent that can act needs guardrails an automation rule does not. Most organizations are still experimenting precisely because they are working out where the boundaries go. Experimentation here is sensible, not laggard behavior.
Why most workflow projects stall
The stall point is rarely the model. It is everything around it.
Data access. A workflow needs clean inputs from the systems of record. If the data is fragmented or permissioned poorly, the AI step starves.
Handoffs. Real processes pass work between people and systems. A workflow that automates one step but breaks the handoff into the next step creates more work, not less.
Ownership. Pilots have champions. Scaled workflows need owners who are accountable for output quality after the launch enthusiasm fades.
Measurement. If no one defined what the workflow was supposed to improve, no one can tell whether it did. This is how companies end up with widespread use and no EBIT impact at the same time.
None of these are AI problems. They are operations problems that AI exposes.
How to evaluate AI workflow vendors
If you are buying, the commercial question is not whose model is best. It is who helps you cross from experimentation to scale.
Ask vendors to describe the full workflow, not the AI step. Where does data come from. What happens before and after the model runs. Who owns the output. Vendors who only talk about the model are selling a pilot.
Ask for the integration surface. A workflow tool that connects to your existing systems beats a better model that lives in isolation. The McKinsey scaling gap is largely an integration gap.
Ask how the vendor measures impact. The credible ones tie their work to a process metric — cycle time, error rate, cost per case — not to usage statistics.
Ask what guardrails exist for agentic steps. As workflows give AI more authority to act, the vendor's controls become part of your risk posture.
The practical sequence
Start with one workflow that has a clear owner and a measurable outcome. Pick a process where the data is already accessible. Resist the urge to start with the most complex case.
Instrument it before you automate it. If you cannot measure the current process, you will not be able to prove the new one is better.
Design the handoffs explicitly. The AI step is the easy part. The connections to people and systems are where projects fail.
Then expand to adjacent workflows that share data and owners. Scaling is mostly a matter of repeating a pattern that already works, not inventing a new one each time.
The bottom line
The market has settled the adoption question. Most organizations use AI somewhere.
The open question is integration. A minority have scaled, an even smaller group reports EBIT impact, and the two groups overlap. That overlap is the whole story.
Workflows are the bridge between using AI and getting paid for it. The companies that build that bridge deliberately — with data access, clear handoffs, named owners, and real measurement — are the ones showing up in the EBIT numbers. Everyone else has a pilot.
Top comments (0)