CMiC announced NexusAI on 2 November at its CONNECT 2026 customer conference in Orlando. The feature list is the expected one for an ERP vendor in 2026 — ask a question in plain language, get an agent to do the work. One line in the release is doing far more work than the rest, and it is about permissions.
What Was Announced
NexusAI is described as an AI Launchpad inside NEXUS, CMiC’s AI platform. It lets users launch AI agents and surface ERP content by asking in plain language rather than navigating screens; it shows real-time KPIs and alerts; it offers guided workflows started either by uploading a document or by typing a request; and it lets users create jobs, budgets, invoices “and more” through a request, review and confirmation sequence. Steve Cangiano, CMiC’s chief product officer, said the new NEXUS release builds on the first by improving job costing and payroll, and framed the goal as delivering “real operational impact.”
Then the plumbing: NexusAI is powered by the CMiC MCP Server, and the release states that it runs every request within existing security settings and user permissions. The release gives no further technical detail about that server, and does not state an availability date for NexusAI.
CONNECT 2026 is the company’s 28th annual customer conference, running 1–4 November at the JW Marriott Grande Lakes in Orlando. The release puts attendance at roughly 900 customers, partners, industry leaders and product specialists, across more than 30 sessions with 30-plus speakers, four new experiences and 24 sponsors and exhibitors. Guest keynotes come from Zack Kass, a global AI advisor the release says previously led go-to-market at OpenAI, and Steve Jones of Dodge Construction Network. Shirin Ali, CMiC’s head of marketing, called it “our boldest conference yet, and that is not something we say lightly.”
A note on how CMiC positions itself: its own homepage headline reads “First AI-Powered Construction ERP,” and the release’s boilerplate describes the company as “the leading provider of AI-powered construction ERP.” Both are the company’s characterisations of itself, and this publication is reporting them as such rather than adopting either.
Permission Inheritance Is the Actual Product Decision
Read “runs every request within existing security settings and user permissions” next to what this publication published yesterday. Autodesk’s own documentation lists three Revit MCP servers — Read, Write and Experimental — and warns that two of them can modify a model. The governance gap there is not that the capability exists; it is that an agent with write access to a building model is an actor whose authority is defined by the server it is connected to, and a firm accumulating AI-built tools without named owners has no way to say who authorised what.
CMiC’s answer to the same structural problem is to refuse to create a new authority surface at all. If the agent can only do what the signed-in user could already do through the interface, then the permission model that the finance department spent years configuring is still the permission model — and an agent that creates an invoice is subject to the same approval routing as a human who creates an invoice. That is an unexciting design choice and very likely the correct one for a system of record that holds job costing and payroll, where the blast radius of an over-permissioned agent is not a corrupted model file but a posted journal entry.
Whether it holds in practice is a separate question the release does not answer. “Within existing security settings” is a claim about scope, not about the integrity of the confirmation step, and the request–review–confirm sequence is where the real risk sits: a user who approves a plain-language instruction to “create the budget” is approving something they did not themselves compose. The release describes the sequence but not what the review step actually shows. This publication has not tested the product, and CMiC has not published the detail.
The MCP Thread Is Now Industry-Wide
NexusAI arrives into a pattern this publication has been tracking all autumn, most of it among design, model and field tools — though not all of it. Graneet’s MCP interface already lets customers of a French SME construction ERP query their data from ChatGPT or Claude, and Access Evo extended an agentic layer across Access Coins, its finance and commercial ERP, on 15 October. Revizto shipped an MCP server, a properties API and a developer portal to let contractors point Claude, ChatGPT or Copilot at live model data. Nemetschek gathered Archicad Max, Allplan Max and Bluebeam Max under one premium AI tier with MCP servers as the part that travels between them. EarthCam added an MCP connector over pre-analysed jobsite imagery. Build Paperless exposed 37 tools over MCP to Australian builders. Arcadis and Autodesk built the enterprise version of the same idea.
What distinguishes CMiC’s is the breadth of what sits beneath it. The design and model servers expose data that is rich and consequential but recoverable. An ERP holds the money: jobs, budgets, invoices, payroll. Put an MCP server under all four and an agent’s mistake is a posted transaction rather than a document — which is a reasonable explanation for why the permissions sentence is in the release at all.
Nor is the permissions answer itself new, and this publication should say so, because it made the same point in October: reporting on Access Evo, we wrote that “inheriting the host system’s permission model rather than building a second one is the correct answer, and it is the specific risk that an AI layer over an ERP introduces.” CMiC has reached the same conclusion and put it in a press release, which is a reasonable sign the argument is becoming settled rather than contested. The closest comparison on the other side is NetSuite’s construction AI agents, estimate import and payment application, which reach for the same ledger without the protocol layer.
What to Watch
The availability date, which the release does not give. CONNECT is a customer conference, and conference announcements run from shipped to roadmap with no obligation to say which. Cangiano’s framing — a second release improving job costing and payroll — reads like software that exists, but the release does not distinguish NexusAI’s status from the broader NEXUS release it sits in.
And whether the MCP server is open to anyone but CMiC. The release presents it as the engine beneath NexusAI, not as an interface a customer’s own agent can call. Those are very different products: one is a vendor using a protocol internally, the other is the anti-walled-garden move Revizto made. CMiC maintains a developer site; the release says nothing about whether the MCP server appears on it.
