A useful way to evaluate AI is to stop asking which jobs it can perform and start asking which minutes inside those jobs should never have required a person in the first place.
An engineer may spend an hour finding the latest approved standard before spending ten minutes making an engineering decision. A project manager may spend Friday afternoon turning information that already exists in five systems into a status report. A support specialist may read a ticket, locate the room, retrieve device telemetry, find the last incident, and only then begin troubleshooting. None of those examples requires replacing the engineer, project manager, or support specialist. They require removing the machinery around the work.
Decompose the job before you automate it
Break a role into activities: retrieve, interpret, decide, create, communicate, approve, and execute. AI is strongest today in portions of retrieve, interpret, create, and communicate. Deterministic automation remains better for many execute steps. Human judgment remains critical where context, accountability, negotiation, safety, or ambiguous tradeoffs matter.
This decomposition prevents the common mistake of trying to automate an entire role with one “agent.” A better architecture may use retrieval to assemble context, an LLM to summarize it, business rules to validate required fields, an API to create the record, and a person to approve the exception.
A real example: technical submittal review
Consider an engineering team reviewing equipment submittals against a design package. The valuable work is deciding whether a substitution meets the design intent. The expensive administrative work is finding the relevant specification, extracting model numbers, comparing technical characteristics, identifying missing documents, and formatting comments.
An AI-assisted workflow can ingest the submittal, identify the proposed equipment, retrieve the applicable specification and approved standard, produce a structured comparison, flag discrepancies, and cite the source pages. The engineer still makes the approval decision. But a 45-minute review may become a focused 10-minute engineering decision.
The important control is citation and traceability. The system should show exactly which source material produced each comparison. Without that, speed is gained at the expense of confidence.
A second example: portfolio status reporting
A global technology program may have hundreds of active projects. Status lives in project systems, commissioning reports, email, vendor trackers, and issue logs. Senior staff often spend hours assembling a management view.
Instead, an AI layer can retrieve the current records, summarize changes since the previous reporting period, identify projects with schedule movement, group recurring blockers, and draft the narrative. The output is not published automatically. A program lead reviews it, corrects context the system cannot know, and approves it.
The value is not the generated paragraph. The value is converting hundreds of fragmented updates into an exception-oriented review process.
Design for confidence, not magic
Production AI needs more than a prompt. Define the authoritative data sources, permissions, retrieval boundaries, retention policy, model behavior, human-review points, logging, and failure mode. Decide what happens when the model is uncertain or the required source is missing.
For high-consequence workflows, require source citations and deterministic validation. For sensitive data, enforce access at retrieval time rather than assuming the model will politely ignore information the user should not see. For repetitive tasks, measure false positives and false negatives, not just whether the demonstration looked impressive.
The business case is time returned to valuable work
A practical AI business case can be built from frequency × minutes saved × loaded labor cost, adjusted for adoption and review time. Then add quality benefits such as reduced rework, faster response, or better consistency.
A ten-minute improvement on a task performed 20,000 times a year returns more than 3,300 hours. That is tangible capacity. The organization can use it to absorb growth, improve service, shorten delivery, or reduce overtime without pretending that an entire profession disappeared.
The practical objective is not more technology. It is a better-performing operation with clearer ownership, less friction, and technology that can be supported over its full lifecycle.