AI Generative Routing Untangles MEP Coordination in 2026

Walk onto any busy job site and look up before the ceiling goes in. That gap between the structural slab and the finished ceiling is some of the most contested real estate in the building. Ductwork, chilled-water pipes, sprinkler mains, cable tray, conduit, gas lines — every discipline wants the same few inches, and every one of them was drawn by a different engineer in a different model. Somebody has to make it all fit. For decades that somebody was a human, pushing pipes around in Revit at 9pm and hoping the clash report came back clean.

Generative MEP routing changes who does that first pass. Instead of an engineer hand-drawing every duct run, a solver takes the building geometry and a set of rules, then proposes complete routing layouts — sometimes hundreds of them — in the time it takes to get coffee. It’s one of the few places in AEC where “AI” isn’t a slide-deck promise. The results are real, buildable pipe.

What generative routing actually does

Strip away the marketing and the idea is simple. You tell the software where things start and where they end: the air handler here, the diffusers there, the panel in this closet, the fixtures in those bathrooms. You give it the space it’s allowed to use — usually the plenum, minus whatever the structural and architectural models already claim. You hand it the rules: minimum clearances, bend radii, slope requirements for drainage, keep-out zones around access panels. Then you let it search.

The search is the part that matters. A single duct run from a shaft to a far corner of a floor has an astronomical number of possible paths once you account for elbows, transitions, and everything else fighting for the same corridor. A person picks one path that works. A solver evaluates thousands and ranks them — shortest total length, fewest fittings, least material, best maintenance access — depending on what you told it to care about. The output isn’t a fuzzy suggestion. It’s a coordinated, connected system with real sizes and real fittings, ready to drop back into the BIM model.

Where it slots into a real workflow

Nobody rips out their whole toolchain for this. The typical flow starts in Revit or an IFC export. Architects and structural have already placed their walls, slabs, and beams. The MEP engineer loads that context, marks the equipment and connection points, and defines the envelope the systems can live in. That setup is the actual work — and it’s where good engineers earn their keep, because garbage constraints produce garbage routes.

Once the run finishes, you’re not done, you’re reviewing. You scroll through the ranked options, throw out the ones that look clever but would be a nightmare to install, and pull the strong candidates back into the model as native families. From there it’s the normal process: coordination meetings, clash checks in Navisworks, sign-off. The difference is that the first fully-coordinated draft showed up in an afternoon instead of three weeks, and it came in already deconflicted against the other trades instead of needing a dozen rounds of “you moved into my duct again.”

The constraints are the whole game

Here’s the part vendors underplay. A generative router is only as good as the rules you feed it, and MEP rules are brutally specific. Sprinkler lines answer to NFPA. Ductwork has to hold velocity and pressure targets or the space is loud and drafty. Drainage needs consistent fall — you can’t route waste pipe like it’s a data cable. Electrical conduit has fill limits and bend counts that decide whether anyone can actually pull the wire later.

Good tools bake a lot of this in. They know a chilled-water pipe needs insulation clearance, that a valve needs a person-sized gap around it, that you don’t run a hot flue next to a temperature-sensitive line. But every firm has house standards on top of code — preferred fittings, a fabrication shop that hates a certain transition, a facilities team that wants everything reachable from a lift. The setup work is teaching the software your standards. Skip it and you’ll get routes that pass code and fail your shop.

The tools doing this now

Augmenta is the name most often attached to this problem. Their system takes a building model and generates a full 3D MEP response — routing electrical, and increasingly mechanical and plumbing, against real code and clearance rules, often turning a floor around overnight. It started in electrical conduit, which is a nasty routing problem because of fill and bend limits, and grew outward from there.

Autodesk has been folding generative and automation features into its BIM stack, so more of this shows up inside the tools engineers already sit in. Hypar comes at it from the space-and-systems side, letting teams script building logic and generate layouts programmatically. And a long tail of smaller AEC startups — you’ll find hundreds in any current tools directory — are chipping at slices of the same problem: load calcs, equipment selection, duct sizing, prefab spooling. Not all of them will survive. The ones that win will be the ones that produce output a foreman trusts.

What it’s genuinely good at — and what it isn’t

It’s very good at the tedious, combinatorial middle of a project. Routing a repetitive hotel floor, coordinating a parking-garage ceiling packed with services, generating three cost-versus-clearance options for a client meeting — this is exactly the kind of high-effort, low-creativity work that burns out engineers and eats fee. Hand it off.

It’s weaker where judgment lives. It won’t tell you the mechanical concept is wrong, that the shafts are in dumb places, or that the client’s real problem is the plant size, not the pipe route. It optimizes inside the box you draw; it won’t question the box. And on a gnarly renovation with as-built conditions that don’t match any model, a sharp engineer who’s walked the building still beats the solver. The tech is a force multiplier for good MEP thinking, not a replacement for it. Treat it as an eager junior who works fast and never questions the brief.

The real payoff is downstream, in the shop

The routing itself grabs attention, but the money is often in what comes after. Because generative output is precise and consistent, it feeds prefabrication cleanly. Ducts and pipe racks get spooled, cut, and assembled off-site, then craned in as modules. That only works when the model is trustworthy down to the fitting — and a solver that respects fabrication rules produces exactly that kind of model.

Fewer surprises in the field is the quiet win. When the first coordinated model is generated against real clearances instead of stitched together from four separate discipline models, the ugly clashes surface in the office, where fixing them costs an email — not on site, where fixing them costs a crew standing around and a change order. That shift, from finding problems late to designing them out early, is worth more than the hours saved drawing.

Bringing it into a studio without breaking your standards

Start narrow. Pick one repetitive building type — a lab floor, a hotel key module, a retail box — where the payoff is obvious and the risk is low. Get your house rules encoded before you trust a single output: your fittings, your clearances, your fabrication quirks. Run the tool alongside a senior engineer for the first few projects and compare its routes to what they’d have drawn. You’re not just checking the software; you’re calibrating your own team’s trust in it.

Keep a human on the sign-off, always. The failure mode isn’t the AI drawing something wildly wrong — it’s a plausible-looking route that violates a standard nobody encoded, sliding through because it looked fine at a glance. Someone who knows the building and the code has to own the final layout. Do that, and generative routing stops being a gimmick and becomes the thing that gets the boring 60% of the model done so your engineers can spend their hours on the 40% that actually needs a brain.

Questions engineers keep asking

Does this replace MEP engineers?

No, and the framing misses the point. It replaces the hours engineers spend hand-drawing routes that any competent solver can generate faster. The engineer’s real value — sizing the system, judging the concept, catching the thing the model can’t see — becomes more important, not less. You’re trading drafting time for judgment time.

Can I trust the output enough to build from it?

Only after review, and only if your constraints were right going in. The output is buildable geometry, not a rumor — but it inherits every gap in your rule set. Treat the first coordinated model as a very strong draft that a senior engineer signs off on, not as a finished deliverable that skips the coordination meeting.

How much setup does it take before I see value?

More than the demos suggest. The routing runs in minutes; encoding your firm’s standards, fittings, and fabrication preferences takes real effort up front. That investment pays back across every project afterward, which is why it makes sense to start with one repetitive building type and reuse the setup rather than reinventing it each time.

What happens on messy renovation projects?

That’s the hardest case. When as-built conditions don’t match any model — old buildings, undocumented changes, surprises above the ceiling — the solver is working from bad data and it shows. Reality capture and a good scan help, but a renovation with real unknowns is still where an experienced engineer who’s walked the site outperforms the machine.

Want to see AI turn a plain-language brief into finished visuals and design directions? Explore what Pixintellect is building for design teams.

Images: AI-Designed