CIOs Improve AI Cost Efficiency By Matching Technology To Each Workflow’s Requirements
Prafulla Patil, Founder and CEO of OneSolve, and JP Mehlhaff, Partner at OneSolve, outlined an integration-first approach to choosing where automation and AI belong.

Enterprise AI costs accumulate through decisions made one workflow step at a time: which tasks follow fixed rules, which need judgment, and which simply need systems connected. CIOs can make those choices explicit through an architecture that combines integration, automation, and agents around a defined business outcome. That approach also gives teams a basis for evaluating what happens after launch, when production use exposes opportunities to improve the process and the work employees can do with it.
Prafulla Patil, Founder and CEO of OneSolve, and J.P. Mehlhaff, Partner at OneSolve, help enterprises design and operate AI and automation workflows. Patil previously served as CIO/CISO at Mapbox and held technology leadership roles at LinkedIn; Mehlhaff managed Mapbox's go-to-market technology stack for more than six years. Their implementation work has shaped a specific order for deciding how much intelligence a business process needs.
"How we look at things is integration, automation, AI. That's the sequence. If there's a problem you can solve using integration, you've got to solve it using integration because that's reliable," Patil said. That sequence makes the allocation of work an architectural decision with consequences for operating cost, consistency, and maintenance. A governed integration layer can give teams a reusable foundation as individual workflows expand across departments and systems.
Discover the process: Process mapping needs to capture what employees actually do, including the steps that never made it into a requirements document. Patil described shadowing users to uncover that work before building tools around their intentions. "Is your data clean? Is your data at the right place? Is it accessible? Then, number two, do you have processes down?" The answers establish what the implementation must account for: where information lives, how employees retrieve it, and which decisions depend on experience that hasn't been documented. They're also inputs to a useful project scope, since connecting two applications doesn't establish everything that must happen between them.
Allocate judgment selectively: A workflow can combine predictable execution with a few specific calls for interpretation. Patil offered a six-step process in which four steps follow fixed rules and two use a model to make decisions. Mehlhaff described how the team locates those boundaries: "How far can we take this before there's so many permutations that we need an agent? And then, if we were to instruct a person to do it, how would we instruct them?" That exercise produces instructions, context requirements, and a clearer account of why each AI step belongs. It also makes the cost of adding judgment visible at the point where the process is designed.
Consolidate operational control: Each additional gateway, logging system, and agent platform adds another component for IT to operate. The architectural goal is a shared control plane through which teams can oversee those dependencies and their costs. "What I see CIOs doing now is buying bits and pieces. They're buying MCP Gateway, they are putting logs in Splunk so that they have some governance and monitoring for token spend, and then they have another agentic platform. There's just so much sprawl going on," Patil said. His concern extends to employees inventing different ways to perform the same task, leaving the enterprise with inconsistent methods and little assurance that those methods are efficient.
The design decisions continue into the operating budget. Someone has to review production behavior, decide which changes are warranted, and determine whether the workflow is improving against its intended outcome. NIST's AI risk management framework calls for regular testing during operation and post-deployment monitoring that captures and evaluates user input. For CIOs, that creates a practical funding question: who owns the review cycle once the initial implementation team hands the system over?
Use production feedback: Patil described why the operating plan needs to make room for learning from users: "Your job is not done when you launch it. It gets started. We built a sales agent for one of the customers. The real gains started coming from month two. We had one month of conversation history. We saw what was working, what was not working, and then we fine-tuned those agents." After those refinements, he observed salespeople brainstorming with the agent as a second brain. That use raises an additional evaluation question for the business: How should it judge the quality of support for an employee working through a decision, beyond the completion of an assigned task?
Measure the work: Patil reported that one customer's purchase-order implementation processed 75% of its purchase orders directly through the system, against a previous service-level agreement of roughly a week. He described faster fulfillment and the opportunity to redirect staff from manual processing. A useful operational baseline lets a CIO distinguish those outcomes from the amount of AI activity recorded. Mehlhaff tied the evaluation to employees' time: "You can look at these discrete tasks that we're removing, or the time it takes to prep for a call. But I think the actual impact that you see in the organization is, now, as the salesperson or in accounting, I'm actually spending more of my time interacting with the people I work with or the people I need to be working with." Time saved can be counted; whether it improves customer contact and collaboration requires attention to how roles change.
Those measures also need an owner across the functions the workflow touches. Mehlhaff observed that AI can expose loose definitions of success and unclear accountability when an initiative spans departments. A CIO can use the process map to bring those decisions into the same conversation as architecture and cost, giving the business a shared basis for deciding what to expand, revise, or stop.
That discipline leaves room to change technologies without reconsidering the purpose of the whole workflow. "Something that'll never go out of style is cost efficiency," Mehlhaff said. "There needs to be more focus on the actual architecture so that you don't really care about the next model coming out, because you can use a more cost-efficient one for these tasks."
If this caught your attention, that’s not accidental.
The best editorial systems don’t happen by accident. Outlever builds them.









