AI can advise on your facility handover.

This April, Tokyo Electric Power Company (TEPCO) restarted Kashiwazaki-Kariwa Unit 6. The world's largest nuclear station ended fourteen years of silence. Its handover documentation runs to terabytes of structured information. Every object passes through a designated information reviewer.

TEPCO assigned humans to that role. Hinkley Point C, now deep in electromechanical fit-out, assigned humans. Sizewell C, financed at £14.2bn ($18bn, €17bn) in June 2025, will assign humans. They could have assigned software since August 2022.

That month, the British Standards Institution (BSI) published BS EN ISO 19650-4:2022. Clause 3.2.3 defines the information reviewer. The clause permits two kinds of reviewer. The first is an artificial intelligence (AI) agent. The second is an automated rules-based system. That single Note authorised software to do work most operators still assign to humans.

The standards press covered it. The construction press missed it. Three years on, the clause remains unread across reactor builds, liquefied natural gas (LNG) terminals, and hyperscale data centres. The Emirates Nuclear Energy Corporation (ENEC) signed off Barakah Unit 4 in September 2024. Shin Hanul 3 follows. None has appointed an AI reviewer.

Manual review will not scale. The standard already says so.

ISO 19650-4:2022 is the fourth part of the international family on information management for built assets. Parts 1 to 3 set the framework. Part 4 specifies the exchange itself: provider, receiver, reviewer.

The reviewer role matters because it gates every handover. Information moves from contractor to operator only after the reviewer accepts it. The standard lists six review criteria: conformance, continuity, communication, consistency, completeness, and technical quality. Clause 7 sets the test.

What the construction press missed was small but decisive. The reviewer does not have to be a person. Note 3 to Clause 3.2.3 puts an AI agent or an automated rules-based system on equal footing with a human reviewer.

The Netherlands saw this coming. USPI-NL is the Dutch oil and gas information standards body. It refined the Capital Facilities Information Handover Specification (CFIHOS) from 2012 onwards. CFIHOS structured handover data so machines could parse it. Governance moved to the International Association of Oil and Gas Producers (IOGP) in January 2020. By the time ISO 19650-4 published, that data layer already existed in process industries.

buildingSMART Netherlands and buildingSMART Spain were aligning national templates simultaneously. BSI Group Japan certified Fujita Corporation for the BIM Kitemark against ISO 19650 in 2024. Korea adopted the series through KS standards. The standards world moved. The construction trade press did not follow.

Three failures break human review at modern project scale.

First, volume. A reactor or hyperscale campus generates millions of structured information objects. A 1.4 GW pressurised water reactor (PWR) handover package can exceed 8 terabytes of native files. Reading every object once would take a senior engineer years.

Second, continuity. Reviews happen in batches. Human reviewers fatigue, drift, and lose context across handovers spaced months apart. The same clash, the same metadata gap, the same wrong classification slips through twice. Construction projects globally pay for these errors at the operations stage.

Third, economics. Specialist reviewer day rates in the UK run to £1,200 ($1,524, €1,440). A full information model review for Sizewell C, the £14.2bn programme above, would consume hundreds of thousands of hours. The numbers do not work.

A compliant AI reviewer does four things, vendor agnostically.

It checks information against the Exchange Information Requirements (EIR). The EIR specifies the data, format, and level of need. The AI parses the delivery, compares it to the requirement, and flags deviations in seconds.

It tests the six ISO 19650-4 criteria. Conformance against schemas. Continuity across version histories. Communication of metadata. Consistency between models. Completeness of attribute fields. Technical quality of geometry.

It produces an audit trail. Every check, every deviation, every accepted exception is logged with timestamp and rule reference. The operator inherits a reviewable record, not a signed certificate that hides the work.

It defers to humans on judgement calls. The standard does not require AI to replace the reviewer. It permits AI to perform the role. A qualified human keeps sign-off; the AI surfaces exceptions.

This pattern is now appearing in production. Singapore's Building and Construction Authority and Dubai's regulators run automated compliance checks on building submissions. Finland and Estonia are following. The UK BIM Framework added an AI capability to its guidance interface in 2024. The infrastructure to do this exists.

The opportunity is temporal. The clause is three years old. Most operators have not read it. Some vendors are selling it back as innovation. The arbitrage sits between what the standard already permits and what the industry believes it must wait for.

Compare regions. The Netherlands and the UK formalised the data layer first. Japan and Korea moved through national certification. Gulf operators are different. TRANSCO in Abu Dhabi and the project teams behind Saudi Vision 2030 are starting handovers fresh. They can embed AI review from day one.

The temporal gap is wider in nuclear and LNG than in commercial buildings. Nuclear handovers run for decades. The Hinkley Point C operating life will exceed 60 years. Information handed over in 2030 must remain intelligible to operators in 2090. An open standard gives that information a longer half life than any single contractor's review process.

The Office for Nuclear Regulation (ONR) references ISO 19650 in its information management expectations. The Nuclear Decommissioning Authority (NDA) does the same. The International Atomic Energy Agency (IAEA) safety guides increasingly cite information modelling principles. None of these regulators objects to AI reviewers when human accountability is preserved.

ISO/TC 59/SC 13 holds the secretariat for the 19650 series. Parts 1, 2, and 3 are under review. Amendments are expected in 2026. The direction of travel is clear: more automation, more open data, fewer manual gates. Early movers will already have the audit trails to prove their reviews held up.

Three rules separate compliant adoption from theatre.

Use open formats. The standard works because it presumes machine-readable structured information. Proprietary file formats break automated review. Industry Foundation Classes (IFC) and CFIHOS-aligned data structures hold up.

Keep a human accountable. The clause permits AI to act as reviewer. It does not absolve the appointing party of accountability. Sign-off remains a human signature.

Log everything. The audit trail is the artefact. An AI reviewer that does not produce a queryable, exportable record is not a reviewer; it is a black box.

For nuclear, LNG, and data centre investors, the implication is operational and economic.

Operationally, projects designed for automated review hand over faster and with cleaner data. Operators inherit a model they can use, not a static document set they must reverse engineer. The handover gap shrinks measurably.

Economically, review cost compresses. A £1.2m ($1.5m, €1.4m) human review programme drops to a fraction when AI handles volume work. Humans then handle exceptions. Capital that funded low-judgement labour reallocates to engineering decisions.

Strategically, projects that embed Clause 3.2.3 Note 3 from procurement onwards establish a defensible information standard. Late adopters will pay for retrofits. The early adopters in Abu Dhabi, the Netherlands, and Korea are setting the practical precedent now.

A standards body in London authorised AI agents to act as information reviewers in August 2022. The construction press never wrote about it. Three years later, operators are still appointing human-only review teams for handover packages too large to read. The arbitrage is sitting in BS EN ISO 19650-4:2022, Clause 3.2.3, Note 3. Operators who read the clause will compress review timelines, lower cost, and inherit cleaner asset information. The clause is public. The opportunity is not waiting.

Next week: Microsoft owns the reactor now. Amazon leases one. Which model wins?

Agents are desperate for ephemeral databases.

They spin up projects, fork environments, test ideas, and tear them down. Over and over. But every database on the market was designed for humans who provision once and stick around. Agents don't work that way.

Ghost is a database built for agents. Unlimited databases, unlimited forks, 1 TB of storage, and 100 compute hours per month. All free. Try it here.

Install the ghost MCP