What the intelligence layer actually does, which workflows it runs today, and exactly where a person has to sign before anything leaves the building.
The workflow
Ten stages, drawn as one line so the hand-off points are visible. The amber nodes are people — the brief at the start and the sign-off at the end. Everything between them is machinery that reports to them.
Worked example
One sentence from an engineer. Here is every step between that sentence and a report they can act on — including the two places the system stops and asks.
"Check the chilled water system on R04." The scope, the model and the standard being applied are established by a person who knows why it matters.
Reads the federated model, identifies the chilled water network, and separates it from adjacent services by system assignment rather than by colour.
Equipment, terminals, pipe segments, diameters, design flows, insulation and valve positions — off the elements, not off a spreadsheet.
Sizing bands against flow, velocity limits, index-run pressure drop, clearance and access to plant, isolation and drain-down provision.
A DN150 branch carrying flow sized for DN200 upstream. Four segments with no design flow parameter. A failing index run. Each finding carries its element ID and the rule that fired.
Findings grouped by cause and severity, with the parameter values that produced them, and an explicit list of what could not be determined from the model.
Confirms the real findings, discards the false ones, and decides what changes. False positives go back into the rule set — that feedback is the actual product.
Why this is stronger than "AI makes engineering faster." Every step above names a real artefact: a system assignment, a parameter, a rule, an element ID. A claim that names artefacts can be checked. A claim that names a feeling cannot.
What the system does not do. It does not resize the pipe. It does not close the finding. It does not update the model or the drawing. It does not decide whether a marginal velocity is acceptable on this project — that is an engineering judgement with a name attached to it.
The corresponding scripted run is shown in the console demonstration on the homepage. That console is simulated; this description is of the workflow as designed. Model reading and rule application are in development and used internally; the engineering review step is how we already work.
Engineering agents
An agent is a workflow with a named input, an explicit rule set, a reasoning step and an output an engineer signs. It is not a personality and it is not a chat window.
01 INPUT
Structured model data plus the engineering rule set that governs it.
02 REASON
Resolve relationships, test against rules, detect what does not fit.
03 VALIDATE
Findings checked against discipline criteria, not just geometry.
04 AUTOMATE
Validated actions run as routines. Drafts, never silent model edits.
05 APPROVE
A named engineer accepts, amends or rejects. Nothing issues without this.
Reads a model and answers structured questions about what is in it — element counts by system, parameter completeness, where a given family is used and how systems connect.
Tests the federated model against clearance, access and zoning rules, then classifies what it finds by discipline, severity and who has to move.
Assists with sizing, load and system-performance checks by reading design parameters off the model instead of re-entering them into a spreadsheet.
Runs the project rule set over the model before issue — naming, classification, parameter completeness, view and sheet organisation, status codes.
Handles the repetitive parts of drawing and schedule production that are fully determined by the model, so engineers spend their time on the parts that are not.
Extracts and classifies quantities directly from model geometry and parameters, with the measurement rule visible for each line.
Evaluates design alternatives against stated engineering criteria — routing options, equipment selections, spatial strategies — and shows the trade-offs.
Turns model and project data into decisions a project manager can act on — where change is concentrating, which systems are unstable, what is not converging.
Read the badges. Four of these are in development and used internally on our own delivery. Four are roadmap — architecturally coherent, not built. None run unattended, and none are offered for licence.
What we automate
The standards pass that has to happen before every issue, run as a rule set instead of a person opening views one at a time.
OutputQA report with an element ID against every finding, and failures clustered so you fix the template rather than 241 elements.
Clash detection produces a number. Clash intelligence produces a decision — what actually conflicts, why, and whose model has to move.
OutputIssue list grouped by cause and responsible discipline, with clearance and access failures separated from true hard clashes.
Sizing and load calculations driven from model parameters rather than re-keyed into a spreadsheet that then drifts from the model.
OutputCalculation output reconciled against the as-modelled network, so the two cannot silently disagree.
Drawings and schedules derived from model state, so a model change and a drawing change are the same event.
OutputDraft sheets and schedules ready for engineering review, with the model as the single source.
Quantities measured from geometry and parameters, classified, and traceable back to the elements they came from.
OutputStructured quantity export where every line can be traced to the elements behind it.
Generating model content from engineering intent and rules — routing, supports, penetrations and repeated typologies.
OutputGenerated content presented for engineering acceptance. In development; used on our own delivery, not offered as a product.
Accountability
A short, deliberately boring list. It is the part of an AI engineering proposition that most sites leave out, and the part a technical buyer actually needs.
Never
Generated or amended content is produced in a reviewable state for an engineer to accept. No routine commits geometry to a shared model on its own authority.
Never
Findings are raised, classified and prioritised. Closing one is a decision with a consequence, so a person makes it and their name is against it.
Never
Nothing reaches a common data environment with a suitability code without a named engineer issuing it.
Never
Automation reads suitability and revision state. It does not promote a container from work-in-progress to shared.
Never
Whether a marginal result is acceptable on this project, with this client, under this standard, is not a computation. It is the job.
Never
Client model data is not used to train anything. It is worked on for the appointment and not retained beyond it without written instruction.
The trade this implies. These constraints make our workflows slower than a fully autonomous system would be. They also make the output defensible, which is the only state in which an engineering deliverable has value. We are not planning to relax them.
Engineering enquiry
Send a scope, a drawing set, an information requirement or a workflow you are tired of doing by hand. You will get a technical response on approach, disciplines, deliverables and what is realistically automatable — from an engineer, not a sales desk.