Where MEP Contractors Actually Use AI: 90 Days of Usage Data

An issued-for-construction set lands in your inbox and the answers you need are scattered somewhere inside it: the shutdown restrictions, what a vendor quietly left out of a quote, the scope that still has to be priced. None of it is hidden. It is just spread across 900 pages, and finding it costs an estimator's or a project manager's hour.

The short version: we mapped 90 days of real contractor usage onto the six SOP steps where those answers hide. Three steps carry most of the work: document control, project acquisition, and detailing. Each is run by more than nine in ten contractors. Underneath them sit seven recurring flows, every one run by over half the contractors in the product. This is trade-wide procedure, not one company's habits.

There is also an interactive version of this analysis on the Pelles Unblog, where you can place your own bet on the busiest steps before the data is revealed, click through the heat map step by step, and sort the flows yourself.

How the map was drawn

Nothing here came from a survey. Both signals are things contractors already do, repeatedly, without being asked.

The first signal is repeated questions. Over 90 days, 55% of every readable-document question turned out to belong to an identical question set, re-run project after project. A question list retyped on every new package is not really a question. It is a specification the customer wrote for us.

The second signal is workflow shape. Sorting every workflow run by what it was actually doing returned seven recurring flows, and each one is run by more than half the contractors in the product. The most common is run by 95%.

The six SOP steps, lit by usage

Here are the six steps in the order a job moves through them, with the share of active contractors who run each one:

SOP step	Phase	Contractors running it
Project Acquisition	Pursuit	95%
Job Cost Budget	Preconstruction	59%
Document Control	Execution	91%
Detailing & Engineering	Execution	91%
Major Equipment Buyout	Execution	59%
Submittals & Material Matrix	Execution	68%

Document control is the busiest step by volume, and it is the only one that runs the length of the job rather than sitting at a single point in it. It is also the step before almost every other flow on this map, which is why so much work passes through it.

Project Acquisition: 95% of contractors

The widest reach of any step. It splits into two questions asked in sequence: are these documents complete enough to bid, and what commercial terms do I need before I price? Estimators answer both by retyping the same list into a new project, package after package. One electrical contractor re-runs a 24-question checklist on every package that lands. A mechanical contractor pricing a schematic-design set asks what the drawings might not show that still has to be carried in the number.

Job Cost Budget: 59% of contractors

Two very different flows share this step. One is arithmetic: labor hours and material cost, broken out by trade. The other is a 40-question scope sweep, re-run project after project, looking for work you own but might not have priced: who furnishes, installs, wires, starts up and commissions each piece of equipment; where the tie-ins to existing systems sit; what shutdown windows apply. One leader put it more bluntly and simply asked which items sit inside their scope but never made it into the estimate. That is the same failure mode we wrote about in why good estimates fall apart once the job starts.

Document Control: 91% of contractors

Comparing revisions is the single most-run flow in the product. A June drawing set against the August set. Two versions of a subcontract, looking for what moved between them. Splitting a set is a substantial flow in its own right: a finishing subcontractor asks for only the drawings belonging to their trade, the matching specification sections separated out, and confirmation that nothing in their scope was left behind.

Detailing & Engineering: 91% of contractors

The most cross-trade pattern in the data: one flow wearing different clothes per trade. Electrical pulls panel schedules and breaker counts. Mechanical pulls equipment. Sheet metal pulls gauges and pressure classes. Flooring pulls finishes. The tell is in the wording: people are already asking for the output as a spreadsheet, not as prose.

Major Equipment Buyout: 59% of contractors

The real work here is not picking the cheapest quote. A quote is often cheapest because it excluded startup, or freight, or carries a shorter warranty, and the contractor absorbs the difference months later. Leveling means reading each quote against the specification, not against the other quotes.

Submittals & Material Matrix: 68% of contractors

The step where contractors have gone furthest. It carries the widest repeat use of any standing agent in the product: instead of walking the specification section by section themselves, project managers ask the agent where a section stands and get the answer back with the source attached. That shift from asking on demand to letting something run in the background is the one we described in going from on-demand to always-on.

The seven recurring flows underneath

Six steps, but seven flows, because some steps carry more than one. Volume is measured against the most-run flow; the contractor share is the more reliable figure.

Recurring flow	SOP step	Share of volume	Contractors running it
Compare and find changes	Document Control	100%	91%
Scope narrative and completeness	Project Acquisition	91%	95%
Extract to table / BOM	Detailing & Engineering	76%	91%
Submittal package assembly	Submittals & Material Matrix	51%	68%
Split and isolate documents	Document Control	45%	59%
Quote leveling	Major Equipment Buyout	34%	59%
Budget and labor hours	Job Cost Budget	29%	59%

One caveat on reading this table: the volume categories are keyword-matched and they overlap, so treat those figures as sizing rather than precise measurement. The contractor spread is the number to trust, and what it says is that these are not one customer's preferences. They are the trade's procedure.

The specifications contractors wrote without knowing it

The clearest signal in the data is not a feature request. It is the same list of things to check, worked again against a new project. Three examples, summarized rather than quoted:

• Electrical estimator, 24 questions, re-run on every new package. Whether prevailing wage applies, whether the project is taxable, what the liquidated-damages exposure looks like, how change-order overhead and profit are handled, and which bonds the package requires.
• Mechanical project manager, 40 questions, re-run project after project. Who furnishes, installs, wires, starts up and commissions each piece of equipment; every tie-in to existing systems and utilities with the temporary services each one drags along; every shutdown restriction and allowable outage window.
• Sheet metal detailer, a standing fabrication list, used at two shops. Which metals and gauges the ductwork calls for, what pressure classes apply, whether any welded ductwork is required, and how much flex is allowed at the grilles.

None of these people filed a feature request. They wrote a procedure and ran it by hand, over and over, which is a far stronger signal than anything a survey would have returned.

What every busy step has in common

The map is not random. Every step that lights up shares the same three properties, and together they are what turn a repeated question into something an AI agent can own:

1. The input is a document set. Drawings, specs, subcontracts, quotes. If the answer lives in a 900-page PDF rather than in a database, someone is reading it manually today.
2. The output is a table, not an essay. A bill of materials, a panel schedule, a leveled quote comparison, a list of missing scope. Contractors are already asking for Excel explicitly, so the deliverable is the file.
3. The question set never changes. Only the project does. That stability is what makes it an agent rather than a chat. Package after package, it is still the same list.

Where all three hold, the work is repeatable enough to hand off. Where they do not, it stays a judgment call, and it should.

What this means for your SOP

AI is not spread evenly across a contractor's SOP, and it should not be. The value concentrates on the six steps that run from the day a package lands to the day the job is bought out: acquisition, budget, document control, detailing, buyout, and submittals. Those are the steps where the input is a document set and the procedure repeats, and where every hour saved is an estimator's or a project manager's hour, not an hour of data entry.

If you want a practical starting point, look at your own team's version of these seven flows. Whichever one you re-run most often, with the same list of questions every time, is the one worth handing to an agent first.

Book a demo to see these flows running against your own document sets, or explore the interactive version of this analysis on the Unblog.