Leadership practice / AI-enabled operations
AI-enabled operations: Improve throughput with human control
How I connect AI to the work teams already do, while keeping quality, cost and release decisions accountable.
Explore my leadership experienceChoose an operating problem worth solving
I start with a recurring delivery problem: the effort of keeping design and code aligned, reviewing project files or carrying business context into repeated content work. The team needs to understand what consumes capacity before it chooses the automation. A quick first draft is useful when it reduces the work required to reach an accepted result.
In enterprise R&D, new initiatives competed with maintenance, improvements and refactoring. I led design–development synchronisation and investigated how AI could support that operating model. The mandate was to increase useful delivery capacity while keeping consistency, quality and business responsibility clear.
Build on foundations the team can maintain
With development, we connected a human-maintained Figma UX library to a production frontend component library. Shared patterns and implemented components gave AI a basis for proposing designs and producing implementation. The design-to-code case explains how those foundations support the workflow.
This creates an ongoing leadership obligation. Patterns need owners, components need maintenance and the team needs a way to resolve discrepancies between them. I assess automation alongside those commitments. A workflow that depends on neglected foundations can repeat inconsistencies quickly and increase the work required to correct them.
Make responsibility explicit at the handoffs
Designers review and refine the proposal. Business owners approve the direction before implementation proceeds. Design and development review the resulting work. Each stage has a different purpose: experience judgment, business authorisation and implementation quality.
I want the operating model to state what AI can prepare, what a reviewer must check and who can authorise the next consequential action. That keeps the team’s judgment connected to the work. The amount of oversight should fit the consequences of a mistake and the evidence available about the result.
Count the complete cost of delivery
Model calls are one cost. Preparing context, reviewing results, repairing failures and maintaining the integration also consume capacity. For a new workflow, I would compare the full effort needed to reach the accepted outcome with the previous process. Otherwise, automation can move work from creation into review without making the job more efficient.
In Launcherry’s prompt-caching work, two measured generation-repair calls reused 3,879 of 3,932 input tokens from cache each, approximately 99% when rounded. That is a narrow observation about token reuse. It does not establish overall cost savings across the product or the development process. The lesson is to measure the part of the workflow the evidence actually covers.
Improve quality through evidence and correction
While building my independent products, I developed a broader framework for directing AI implementation and reviewing its results. QAWAI is the read-only CLI/MCP product built from that practice. It checks declared file evidence for reported findings; the broader framework also includes forms of verification outside that product’s scope.
That distinction matters when choosing quality controls. Reproducing a quote from a file supports the evidence behind a finding. It does not prove every interpretation is correct or every defect has been found. Live behaviour needs its own checks when the claim depends on how a product runs. Review should make the remaining uncertainty visible.
Give leadership an adoption decision
The enterprise case reports an estimate: comparable jobs require about one-third of the previous delivery time on average since the AI workflow was introduced. I keep that scope visible: it is an estimate from the current practice, not a controlled benchmark or a guaranteed improvement for another team. A rollout decision needs evidence from that team’s own workload and quality requirements.
My recommended starting point is one recurring job with a clear accepted result. Establish the current effort, introduce the workflow and compare the complete delivery experience, including review and repairs. Extend it when the evidence supports doing so. This is how I approach AI-enabled operations: connect the technology to business performance and retain human accountability for what ships.