JD Rico What changes when AI becomes part of the workflow instead of another tool sitting beside it?
A company can deploy AI without making many changes at all.
That came through clearly in a series of conversations we held with operators working across enterprise software, retail, telecom, logistics, and growth.
The most useful example was also a failed one.
Sayali Patil described an internal automation project in which the team initially added a chatbot to an existing process. The technology worked. Adoption did not.
“We gave people a slower way to do the same slow thing.”
The team eventually removed the chat layer and rebuilt the workflow underneath it.
That distinction kept appearing in different forms across our conversations.
The more useful question is whether companies are changing the work itself once AI enters the process.
Whether the work itself changes once AI enters the process.
The market is moving past the interface
Much of the first wave of enterprise AI has been easy to recognize: copilots, chatbots, generated content, summaries and assistants.
Those applications can be useful.
But they can also leave the operating model almost untouched.
The stronger examples we heard started somewhere else: a recurring task already mattered to the business, AI took responsibility for a specific part of it, and something changed downstream.
At Mr. Checkout, for example, AI now handles the first pass on incoming brand opportunities: classification, disqualifiers and a routing recommendation. The person still makes the referral.
Joel Goldstein described the division simply:
“It prepares, a human commits.”
Speed helped, but his more interesting observation was about consistency.
“The tenth inquiry of the day got the same quality of attention as the first.”
That is a different kind of productivity gain.
The work is not simply faster. The decision arrives in a more consistent state.
In telecom and retail, Priyank Jain described a similar progression from manual forecasting to a system that supports decision-making across roughly 3,000 stores.
The initial application helped with staffing and marketing. The same underlying capability later became useful for questions such as site selection and out-of-home placement.
Jain’s professional work at Boost Mobile focuses on production machine-learning systems across forecasting, anomaly detection, and store recommendations.
What changed was larger than “Excel versus machine learning.”
A one-time forecasting task began to behave more like reusable decision infrastructure.
The operational effect becomes easier to see downstream
This is where logistics becomes particularly revealing.
At Finmile, the AI does not stop after suggesting a route. Its platform is designed around planning, dispatch, live conditions, exceptions, and customer communication. Finmile describes the product as an operating system for execution rather than another planning tool.
That matters because the value sits closer to the actual operation.
A recommendation can be ignored.
A route that is replanned, a delivery reassigned, or a customer proactively updated changes what happens next.
Rich Pleeth’s case shows what changes as AI moves closer to execution. The operational effect becomes easier to observe, while the consequence of getting an action wrong becomes more important.
The same principle showed up in a very different context.
Heath Squier of EVKII put it this way:
“The operational value is not more copy. It is a shorter distance between a customer sentence and a shipped change.”
That is a useful way to think about AI inside an operating workflow.
Companies already have enormous amounts of information.
Customer comments. Support tickets. Search behavior. Sales notes. Research. Analytics. Competitive signals.
The bottleneck is often what happens between receiving the information and changing something as a result.
AI becomes more interesting when it shortens that distance.
Three questions separate AI activity from workflow change
Across the cases, three questions were consistently more useful than asking whether AI had been deployed.
What happened before?
The old process has to be understood first. Otherwise the organization risks automating friction it never needed.
What does AI actually do now?
Classify? Forecast? Retrieve? Recommend? Route? Execute?
“AI-powered” tells us almost nothing about the operating change.
What happens differently afterward?
Does someone make a faster decision?
Does a customer receive a response?
Does capacity move?
Does a product change?
Does the next team receive better information?
If nothing important changes downstream, the organization may have created more AI activity without changing much work.
That also gives leadership a more useful way to evaluate proposed implementations:
| Question | What to examine |
|---|---|
| Before | Where does the work slow, fragment or lose quality today? |
| AI role | What specific part of the work should AI perform? |
| After | Which decision, handoff, or action changes because of it? |
| Outcome | What should improve for the business or customer? |
The table is simple enough to use before approving another pilot.
The next problem is what moves
There is another reason to examine the whole workflow.
Improving one step does not necessarily remove the constraint.
It may move it.
Goldstein saw that after AI made intake triage less expensive. Distributor evaluation became the next capacity problem.
His summary:
“The bottleneck moved, it didn’t disappear.”
That is a useful warning for companies measuring AI productivity one task at a time.
A process can become dramatically faster at the front and still produce no additional value if the next stage cannot absorb the work.
The same issue appears with quality.
A system may process more records, generate more recommendations, or increase aggregate accuracy while making the metric that actually matters to the decision worse.
That is why the next phase of AI implementation will probably require more attention to workflow economics, not just model performance.
Where did the time go?
Where did the work move?
What new constraint appeared?
What became easier to do?
And did any of it improve the outcome?
The implementation question is getting better
A common enterprise AI question has been:
Where can we use AI?
The cases in this research suggest a more useful one:
Where does important work currently slow, degrade, fragment or fail to turn information into action?
Sometimes AI will be the right intervention.
Sometimes the real problem will be data, process design, authority, integration or capacity somewhere else.
That is exactly why the distinction matters.
The strongest cases in this research started with a specific piece of consequential work and changed how that work moved through the organization.
If AI is already in your stack but the workflow still behaves largely the same, that is probably the place worth examining next.
Working through one of those decisions? Email Us.
Next in this series: what happens when AI makes one part of the workflow faster, and the constraint moves somewhere else.
Sources
Praxable operator research, August 2026 — interviews and operating examples from Sayali Patil, Joel Goldstein, Priyank Jain, Rich Pleeth, Heath Squier and other contributors.
McKinsey & Company — The State of AI: How Organizations Are Rewiring to Capture Value
Read the research
Brynjolfsson, Li & Raymond — Generative AI at Work, National Bureau of Economic Research
Read the research
