Qonic has launched Qonic AI, a set of agents built into its cloud BIM modelling platform, alongside an MCP server that lets a user point their own assistant at the same models. AEC Magazine reported the launch on 5 November, timed 9:34am, and that account is the source for the product detail below.
The framing in the magazine’s account is that the agents work inside the model rather than alongside it. That distinction has been doing a lot of work in AEC marketing lately, so it is worth being precise about what it means here. A user types a plain-language request into a chatbot called Ask Qonic, and the agents answer it by driving the platform’s own tools — the same find, check, edit and report functions a person would click. The output is not a suggestion pasted into a chat window. It is a change to the model, made the way the application makes changes.
The Demonstrations, and the Two Guardrails
The capabilities the magazine describes fall into four buckets, and they escalate.
The first is retrieval and display: isolating all furniture in a model and colouring it by level, in one step. The second is data completion — finding every window missing a fire rating, selecting them, and writing a value (F60) into the correct property. The third is reporting: generating a concrete quality report grouped by building and by concrete type into the platform’s Report panel, which the user can then adjust. The fourth is geometry. The magazine’s example is widening selected windows by 0.5 metres to the right, with the direction taken from the user’s viewpoint and the change executed with the stretch tool, and will create a building or a facade from prompts and reference images.
Writing a fire rating into a property field and stretching a window are different categories of act, and the fourth bucket is the one that makes this a governance story rather than a search story. Two guardrails are named. Qonic AI can only change models the user already has permission to edit, and its changes can be undone like any manual edit.
Both of those are the right answers, and neither is novel — which is the point. CMiC made permission inheritance the headline feature of the MCP server it put under its ERP on 3 November, on the same logic: inheriting the host system’s permission model beats building a second one beside it. And the undo guarantee is the quieter of the two but may be the more useful, because it puts an agent’s edit in the same bucket as a human’s mistake — recoverable by the mechanism everyone in the room already knows.
The magazine’s account describes permissions and undo; it does not describe an audit trail — a log of which agent changed what, when, and on whose instruction. Qonic’s own home page goes further than the magazine does, saying Qonic AI “understands the model, compares what changed and fixes it directly — with a receipt for every edit.” That is the company’s wording rather than a specification, and what a receipt records, how long it is kept and who can read it are not stated on that page. The record itself is the thing this publication argued three days ago the sector is not announcing work on while it builds the connectors, so a vendor claiming one in its own marketing copy is worth noting — and worth asking about.
The User Picks the Model
Qonic AI lets the user choose which model answers a request — Sonnet, Opus or Astra — depending on how much reasoning the task needs.
That is a small line in a product announcement and a significant one in this archive. Across the coverage this newsroom has published since June, vendors have shipped MCP servers, agents and assistants in volume; a search of that archive turns up no previous case of an AEC vendor exposing the model choice to the user inside its own interface as a per-request decision. Nemetschek named Claude, GitHub Copilot CLI and AnythingLLM as supported MCP clients on the Bluebeam Max leg, which is the adjacent thing and not the same thing: that is a choice of which assistant drives the software from outside. The nearer miss is Snaptrude’s design agent, which selects a model per task between Claude and GPT — but there the vendor picks, not the user. Qonic’s choice is made inside the application, by the user, per request. The claim here is scoped to what this publication has covered, not to the industry.
Why it matters is a cost argument more than a capability one. If the user picks the model and the AI is metered, the user is choosing how much to spend on each request. A vendor that bundles a single model into a flat subscription absorbs that decision — and the margin risk. Qonic has pushed both to the customer.
The Pricing Is the Argument
Qonic AI runs on tokens. Paid plans include a monthly token allowance; free-plan users can buy a one-off token pack starting at €50; purchased tokens do not expire, and the balance is visible at any time. The company’s pricing page is more specific about what consumes them: natural-language queries through Ask Qonic, and uploading and converting Revit files through the dashboard. Imports through Qonic’s own Revit plugin do not use tokens. Prices on that page load dynamically and were not rendered in the version fetched for this article, so the figures for the platform itself are not reported here.
The structure around the tokens is the interesting part. The platform is sold either per user or on portfolio size — the total capacity of active and passive projects, with passive projects counted at 20% of their original size — and the portfolio option carries unlimited users. Coordination, Drawings and Planning are add-on modules. The free tier covers an active portfolio up to 5,000 m², one full user, unlimited viewers, one active project and five models per project. All prices exclude VAT.
So the shape is: pay for the estate, invite whoever you like to look at it, and pay separately for the thinking. Set that beside the two other answers to the same question in this archive and a genuine divergence appears. Motif priced its agent-native BIM tool per AI user rather than per seat — there, the AI is the unit, and it is a licence. nCircle Tech’s nVibe, also reported this week, is explicit that it charges a flat rate rather than per token so that costs do not rise with use — there, the AI is a fixed cost. Qonic’s portfolio plan carries unlimited users and meters the AI separately. Three vendors, three answers. The comparison is this publication’s, drawn from each company’s own published pricing, and none of the three has been running long enough to say which survives contact with a procurement department.
The MCP Server Is the Other Half
Alongside the in-app agents, Qonic shipped an MCP server so that a user can connect their own assistant — the magazine names Claude, Codex and Mistral Le Chat as examples. Through it, an assistant can look up projects and models, query elements, calculate quantities, export to IFC, change elements and manage library items.
Read that list carefully. “Change elements” and “manage library items” are write operations, reached from outside the application by whatever assistant the user has configured. The in-app agents are constrained by the permissions of the signed-in user and by the undo stack; whether the same two guardrails apply over the MCP path is not stated in the account this article is based on, and it is the first question a firm’s IT function should ask.
This is now a well-populated pattern rather than a novelty. Autom8Labs gave away 22 AEC skills on the argument that the connector was never the hard part; Snaptrude shipped a design agent and a benchmark to grade it. Qonic’s contribution to it is that the platform underneath is IFC-native and cloud-resident, which makes an external assistant’s write path shorter than it would be into a desktop authoring tool — a design advantage and an exposure, in the same sentence.
Who Qonic Is
Qonic NV is based in Ghent, Belgium. AEC Magazine’s 2022 profile of the company names five co-founders — Erik De Keyser, Mark Van Den Bergh, Tiemen Strobbe, Tjerk Gauderis and Sander Scheiris — all previously at Bricsys, the developer of BricsCAD. De Keyser founded Bricsys, which Hexagon acquired in 2018. That profile describes the company then as a “re-start-up” of about twenty people, building a direct-modelling BIM kernel aimed at making quantities reliable enough to take off the model. The platform today imports Revit and Rhino work through IFC and its own API and runs in a browser.
The product’s own site presents Qonic AI as an existing capability — classifying elements, checking models, updating data — and claims “98%+ classification accuracy based on more than 8 million real-world BIM objects.” That number is the company’s and is reported here as such. The page names the corpus the figure was drawn from; it does not say what the classifications were scored against, and this article has not independently tested it.
Qonic’s funding history is not reported here. One commercial database lists a single accelerator-stage entry and no amount; that is a single aggregator with no corroborating announcement, which is not a standard this publication publishes a company’s capital structure on.
What to Watch
The agents are the demo and the pricing is the news. An AEC vendor that meters AI consumption separately from access has made a bet that customers will tolerate a variable line item in a category that has long been sold on predictable per-seat licensing. The bet may be right — it is how the underlying models are sold to Qonic — but it transfers forecasting risk to the buyer, and buyers in this sector have historically paid a premium to avoid exactly that.
The second thing to watch is narrower and more immediate. Two write paths now reach the same models: the in-app agents, with stated permission and undo guarantees, and the MCP server, with an unstated relationship to both. A platform that is cloud-native and IFC-native is unusually well placed to govern that. Whether it does is the question the next release notes should answer.
