ISO 19650 set out to make project information a governed asset. In practice, most of that rigour landed on the model. The 2026 draft revision moves from a focus on modeling to “information management”, widening the scope to the contracts, permits, records and certificates that surround a project’s lifecycle. What it leaves untouched is a historical assumption: that a competent human reads before acting. AI agents bring no such judgement, and will act on a superseded document as readily as a current one. Here’s what that means for capital projects, and what it asks of the record you hold today.
Building a hyperscale data centre, a new process unit at a refinery or a municipal water treatment plant takes dozens of organisations, each sharing and working on thousands of project documents. Each of them produces information the others depend on. The design changes, equipment gets substituted, the permit comes back with conditions attached. Somewhere in that traffic, a decision can get made on a document that was superseded three weeks earlier.
ISO 19650 aims to manage such risks coherently: it is the international standard for managing information across the life cycle of a built asset, while treating information as an asset in its own right. The client states what it needs and when, and the delivery teams agree who produces what. Everything then moves through a single defined place of record, and nothing is used until its status says it can be.
That was the intent. In practice, what happened was narrower: users focused primarily on Building Information Modeling (BIM).
It is the reason a revision was needed. The revised standard shifts the focus to “information management”, merges the delivery and operational phases of a capital project into a single whole-life process, and widens the record it governs beyond the model. The draft went out for consultation in March 2026, and publication is expected in 2027.
The revision fixes the scope of the problem by expanding the definitions. It does not touch the assumption sitting underneath it.
There is much more to a project than the model
For most of the past decade, the discipline concentrated on the model, and for good reasons. This was the one thing every discipline agreed was worth governing, and it rewarded the effort. A BIM model is structured data, not a picture: each element carries its tag number, its specification reference, its duty rating and maintenance interval. Authoring responsibility was defined and revisions were controlled. Clash detection saved real money and time.
Around and outside that model sat everything else. Method statements and factory acceptance test reports. Weld maps, inspection records, punch lists. Variation orders, permits and the correspondence that came with them, vendor data books, warranties.
On a large capital project this runs to hundreds of thousands of documents, and it is where a great deal of the risk lives. A permit condition that nobody traced to the affected work package is a potential blocker. A commissioning certificate that cannot be found at handover is a delay, and sometimes a dispute.
That material was rarely ungoverned in the absolute sense of the term. It was governed by the hard manual work done by experts going over folder structures, naming conventions, a shared drive and a document controller with a long memory. That holds up as long as the people involved stay put and pay attention. The flip side is that manual work involves the risk of human error.
The standard never asked for this split. Its unit of currency has always been the information container: any named set of information, which includes a contract as readily as a model file.
The narrow reading came from the industry, helped by the word “BIM” sitting in the title. The revision is closing the gap between what the standard said and what most enterprises did.
The model is not going anywhere; it becomes part of a bigger whole.
Scope: Where the gap shows
Handover, usually, because that is where somebody has to live with the result.
A water treatment plant runs for thirty years or more. The people operating it in 2050 will not be the people who built it, who will have likely moved on or retired. And the contracts that produced the information will have expired decades earlier.
When the model arrives in good order, and the surrounding record does not, the operator inherits a plant it cannot fully account for. Maintenance planning starts from an incomplete asset register, and a regulatory audit turns into an exercise in digging out historical facts.
The same pattern shows up on data centre campuses built out in phases, where the team commissioning hall four needs to know exactly what was approved in hall two.
Most owner-operators have a version of this story. What is new is who reads the record next.
AI: An over-confident reader raising the stakes
The controls the standard defines are signals to a trained reader. Suitability codes, revision status, approval states: an engineer who opens a drawing marked for information rather than for construction knows not to build from it, and hesitates when the status looks unclear. That hesitation is doing more work than anyone ever wrote down.
Agents are now being pointed at exactly this material. But unlike humans, AI-powered chatbots and assistants do not express doubt when confronted with incomplete documentation.
Consider an agent asked to confirm the commissioning status of a set of equipment on a large energy project. It searches, finds a signed certificate, and reports the system as complete. That certificate was superseded after a retest, and the later document sits in a different folder under a different naming convention because it came from a different contractor.
Nothing in the record told the agent which one governed. A document controller who knew about the retest would have caught it, or at least paused. The agent answers with the same confidence either way, in seconds, across thousands of documents, in a form that reads like settled fact.
The weakness was always there. Human judgement was covering for it. Take the reader out of the loop, and it becomes an operating risk, at a speed no manual check can catch up with.
This bites hardest outside the model, which is the second half of the problem. Model environments carry version control by design. Contracts, permits and certificates often do not, and those are exactly the documents an agent gets asked about when someone wants to know whether work can start, or what was warranted and by whom.
What the revision and rise of AI agents change together
A new landscape is emerging quickly. ISO 19650 is widening its scope to the information and the processes around a project rather than the model at the centre of it. At the same time, agents that cannot judge the standing of a project-critical document are being connected to project records.
So what can you do to anticipate and act?
Status has to bind, not advise. A naming convention is a request. The record itself has to carry the state of a document in a form an agent cannot misread, with the superseded version demonstrably retired rather than merely older, and without the overhead that makes heavyweight project systems unusable for ordinary documents.
The scope has to match the record. Governing the model well and the rest loosely leaves the exposure where it already is. A unified environment across design, construction, commissioning and operations does what the revision does to the standard: it removes the cracks where information falls through.
AI has to work inside the governance. Most AI available on the market indexes what it can reach and answers from it, inheriting every gap in the underlying record and hiding it behind a polished answer. The useful version starts from the control model, respects permissions and approval states, and treats a superseded document as unusable rather than as one result among many.
So where does that leave managers of large infrastructure and capital projects?
You already know that while the asset outlives the project, the information has to outlive both. Now, it also has to be legible to AI readers who bring no judgement of their own. The new revision will not be in force until 2027, but the agents are already being connected, and the record they read is the one you have today.
The long and short of it:
The 2026 ISO 19650 revision replaces “BIM” with “information management”, merges the delivery and operational phases into a single whole-life process, and is expected to be published in 2027. To be well prepared, ensure you control content – all of it – to be ready for all the challenges it entails: a larger-than-ever scope, a focus on information management, and the rising importance of efficient AI tools that lack human judgement.
Frequently asked questions
When does the ISO 19650 revision take effect?
The draft went out for consultation in March 2026 and publication is expected in 2027. The existing parts remain in force until then.
Does the ISO 19650 revision replace BIM?
No. The language broadens from “BIM” to “information management”, while the model stays central. The standard now expands to address the entire record.
What does ISO 19650 cover beyond the model?
Any named set of project information: permits, variation orders, inspection records, commissioning certificates, vendor data books and more. On a large project that runs to hundreds of thousands of documents.
Why does the revision matter for AI agents?
The standard’s controls assume a trained person reads them before acting. An agent brings no such judgement, and will treat a superseded document as authoritative unless the record says otherwise.
Find out more about AODocs AI-native Document Controls for Engineering, Construction and Energy projects.
