Most companies begin an AI conversation by asking what the technology can do. That question produces demonstrations. It rarely produces a durable business.
A better starting point is the operating expense line. Look for the work a company pays people to repeat every week: reconciling information, rebuilding context, chasing documents, checking exceptions, and moving decisions between systems. That is where useful AI often begins.
Follow the work people have learned to tolerate
The strongest opportunities are rarely described as innovation projects. They are accepted as part of doing business.
A project manager reviews the same kind of document for the hundredth time. An estimator rebuilds information that already exists somewhere else. An operations lead keeps a private spreadsheet because no single system reflects what is actually happening. Someone catches a missed change order because they happen to remember a conversation from three weeks ago.
None of these moments look dramatic in isolation. Together, they become a recurring cost: salary, delay, working capital, lost margin, and management attention.
That cost is useful because it gives an AI system something concrete to improve.
The opportunity is not “use AI in the business.” It is “remove the recurring cost of rebuilding the same judgment.”
MEP contractors make the pattern visible
Mechanical, electrical, and plumbing contractors operate inside dense flows of drawings, specifications, submittals, RFIs, schedules, field reports, and change documentation. The work moves across estimating, preconstruction, project management, procurement, and the field.
At every handoff, information can lose structure.
An assumption made during estimating may not survive the project handoff. A revision can affect a scope package without reaching the person who needs to price it. A field condition may be documented in one place while the commercial consequence lives somewhere else. Closeout becomes a final effort to reconstruct decisions that happened over months.
The obvious response is to propose a platform that connects everything. That is usually too broad. The more useful move is to find one repeated operational expense inside the flow.
For example:
- identifying scope changes across drawing revisions;
- assembling the evidence needed for a change order;
- checking submittals against specifications;
- routing coordination issues to the right owner;
- tracking closeout requirements before they become urgent.
Each is narrow enough to observe and valuable enough to measure.
Choose a wedge with an economic consequence
Repetition alone is not sufficient. The workflow needs a consequence the business cares about.
If a missed signal delays procurement, consumes project-management time, or leaves billable work unpriced, the system has a measurable reason to exist. The first version does not need to understand the entire company. It needs to improve one decision path and leave an evidence trail.
A good initial wedge has four properties:
- It happens frequently. There are enough examples to learn from and enough future volume to matter.
- The inputs already exist. The documents, messages, and system records are available, even if they are fragmented.
- A person can judge the result. There is a clear owner who can approve, correct, or reject the system’s work.
- The outcome has economic weight. Time, margin, revenue, or risk changes when the work improves.
This is the difference between an impressive prototype and an operational product.
Productization starts after the pattern repeats
The first deployment should be specific to the way one company works. Its vocabulary, exception paths, approval rules, and source systems are part of the problem—not implementation details to ignore.
As the system runs, a more general product may emerge. Similar decisions appear across projects. The same evidence is repeatedly useful. Review steps become stable. Integrations recur. What began as one focused workflow starts to reveal a repeatable operating model.
That is the moment to generalize.
The path is not custom work or product. It is custom work that earns the right to become a product.
Start with the expense
When we look for useful AI, we look for the work a business already funds but would not choose to design that way again. The budget is already present. The pain is already observable. The people closest to it already know where the exceptions live.
Start there. Find one costly loop, make it legible, and prove that changing it creates value. The technology becomes much easier to design once the work is understood.
See this approach in a worked example: reviewing a completed MEP project for missed additional services.
