Asite announced five AI agents on 10 November, from a London dateline. The release is titled “Less Checking, Faster Setup, Fewer Errors: Asite AI Agents Take on Construction’s Information Workload,” and it says the agents are “available today to Asite Common Data Environment users.”
The interesting thing about the set is where it has been placed. Three of the five agents act at the moment a piece of project information arrives — on upload, before a request is submitted, before an award is made — rather than at the moment somebody later goes looking for it. Neither of the other two is described as checking anything; they build the configuration the checks run inside.
The Five
AI Document Metadata Reviewer. Checks each uploaded document for the right code, status, revision and reference. The release’s tagline for it is “clean data from the first upload.”
AI Form Template Generator. Builds a complete form template from a natural-language description, and can also recreate a form from a PDF, screenshot or image as an editable Asite template. Tagline: “from idea to working form in minutes.”
AI Workflow Designer. Takes a plain-English description of a process and builds a workflow on the Workflow Designer canvas using the project’s real roles and statuses, which the user then refines and publishes. Tagline: “turn business processes into auditable workflows in minutes.”
AI Work Inspection Request (WIR) Validator. Before a request is submitted, checks for missing fields, missing attachments and mismatched descriptions, compares the request against past approvals, and returns a readiness score with suggested fixes. Tagline: “inspections approved the first time.”
AI Bid Evaluator. Reads bid responses against project requirements and flags scope gaps, missing information and non-compliance, creating a record behind each award. Tagline: “tender decisions that stand up.”
Why the Placement Is the Story
Asite has been announcing AI steadily for about eighteen months, and the centre of gravity in those announcements has been moving.
It opened with retrieval. The Cognitive CDE, released on 27 May 2025, introduced AI-powered search across the common data environment: the pitch was that a user could surface insight from files and records without combing through legacy history. It assumed the information was already there and the problem was finding it.
The pattern held into 2026. Both of the releases that follow are datelined day and month without a year; 2026 is this publication’s reading, on the basis that the June one describes itself as coming a year after the May 2025 Cognitive CDE announcement and points at Digital Construction Week 2026. On 17 March Asite released what it described as eight AI agents. The six the release names are mostly query tools — DeepDive and eDiscovery for natural-language questions across project data, Image Search, agents that answer from inside Microsoft Teams and run bulk queries from Excel. The exception is a 3D Data Validation Agent that checks BIM models against standards and contract requirements — the only one of the six named that checks rather than answers. The release does not name the other two.
Then on 1 June came Asite Virtual Coworkers, framed around personas rather than functions: a Document Controller, a Building Safety Act Specialist, a Project Controls Advisor. Dr Jozef Doboš, Asite’s chief digital officer, put the shift as the industry “moving from reactive chatbots into proactive task execution.” Among the capabilities named there is checking uploaded documents against project requirements and compliance obligations — the same job the Metadata Reviewer now does, but attached to a role rather than to a moment.
Retrieval has not gone away alongside this. On 3 September 2026 Asite launched an AI platform for architectural practices with a Royal Institute of British Architects partnership, and the release describes it in finding terms: ask a question in plain English across current and past projects, get back the relevant drawings, revision histories and issue records. The two tracks run in parallel.
November’s release is where the checking work gets attached to moments instead of to roles. A metadata reviewer that runs on upload, a validator that runs before a WIR is submitted and an evaluator that runs before an award is made are not personas you consult; they are checks that sit at named gates in a workflow and fire whether or not anyone thought to ask. Mahmoud Hleihel, Asite’s chief strategy officer, makes the logic explicit: the agents are meant to catch problems at the moment they are created, “using each project’s own rules and data.”
That is a meaningful distinction in a common data environment specifically. Asite’s own solutions page says its CDE lets teams “manage data according to ISO 19650-specified standard procedures and governance” — and ISO 19650 is in large part a discipline about metadata: codes, statuses, revisions, and the suitability of a document for a given purpose. A CDE in which those fields are wrong is a CDE whose search results are wrong, however good the search is. Fixing the fields at upload is therefore the more structural of the two interventions, and the less visible one.
This publication has watched the same move elsewhere in the market. Structured AI stopped reviewing drawings and started reviewing the model they came from in October, on the same reasoning: the artefact is downstream of the thing that produced the error. Pirros extended its Revit agent from managing content to reviewing a live project. UpCodes published results from running its reviewer over 1,300 plan sets. The common thread is a shift from answering questions about a record to policing the record.
The Authoring Agents Are a Different Animal
The remaining two are worth separating out. The Form Template Generator and the Workflow Designer are authoring tools: they produce configuration, not checks. Their value proposition is setup time, which is why “Faster Setup” sits in the release’s own headline between “Less Checking” and “Fewer Errors.”
That matters more in a CDE than it sounds. Configuring forms and approval workflows for a large project is specialist, slow work, and it is one of the reasons CDE deployments take as long as they do. An agent that turns a described process into an auditable workflow using the project’s actual roles and statuses is attacking implementation cost rather than operating cost — and the release’s “auditable” is doing real work in that sentence, since a generated approval chain that nobody can trace is worse than a hand-built one.
Recreating a form from a PDF or a screenshot is the most quietly practical capability in the set. Plenty of contractors and owners have paper and spreadsheet forms that never made it into the system because rebuilding them by hand was not worth anyone’s week.
Flag and Draft, Not Decide
The release puts the boundary plainly: the agents “flag and draft while leaving human judgement to make the calls,” and Hleihel’s line is that “The decisions stay with the people accountable for them.” The release names who those people are — controllers set the rules, engineers approve inspections, panels award contracts.
That is now the standard shape of the sentence in this market, and this publication has recorded it arriving platform by platform. CMiC made permissions the headline feature of its ERP’s MCP server. ProjectTeam named the four actions its AI connection keeps in-platform. Trimble, the day before this release, attached standardised safety labels to the agents in Trimble Connect and said six of ten capabilities were available today.
Asite’s version of the boundary is less formal than Trimble’s — the release carries no label, cites no standard for the agents themselves, and does not describe what an agent is permitted to change. What there is instead is a role map: five agents, each named against the human role that retains the decision. For a CDE, whose whole claim is to be the auditable record, that is the relevant assurance, and it is also the one most easily checked by a buyer reading a workflow.
What the Release Does Not Carry
The release names no customers. No accuracy figure, false-positive rate or benchmark appears anywhere in the document. The only quantities it gives are loose ones — a project produces thousands of documents, setup takes days, checking takes hours every week. For the WIR Validator in particular, which returns a “readiness score,” the absence of any stated accuracy is the gap a buyer would probe first: a readiness score is a number, and a number invites a question about calibration.
No pricing is given in the release. The agents are described as available today to CDE users, which reads as included, but the release does not say so and the page routes interested readers to sales.
What the release does say about assurance is about the platform rather than the models. Asite states that its information security management system is independently certified by BSI against ISO/IEC 27001, that the platform is accredited to store and process UK Ministry of Defence data up to Official-Sensitive, and that it operates regional data centres in the UK, Europe, North America, the Middle East and Australia-Pacific. Those are Asite’s statements, reported as such. They are also the reason a release like this one matters beyond its feature list: Asite’s homepage describes the company as having been defining the CDE category since 2001, and a platform of that vintage adding agents at the upload gate is a different event from a startup shipping the same feature. The install base comes with the deployment.
