SITELIVE What the three companies announced
On 25 August, Fujitsu, Tokyu Construction and Kitano Construction said they are running a field trial of what they call an AI for construction process management support, on the redevelopment of Fujitsu Technology Park — Fujitsu’s own site in Nakahara-ku, Kawasaki, being rebuilt in a programme with Kawasaki City that runs from 2025 to 2035. The trial period is 3 August to 25 December 2026. Tokyu Construction and Kitano Construction are each carrying out different works on the site concurrently, and Fujitsu is verifying the system separately with each of them. Fujitsu built the system and analyses the site data; the two contractors supply site expertise, the verification data, and the evaluation from a construction-operations point of view.
The stated job is specific. The system reads what accumulates in shared site files — drawings, construction schedules, daily work reports — together with records of procedures, approvals, inspections and safety instructions. Against the laws, ordinances and regulations set by national and municipal authorities, it flags candidate omissions in required applications, material ordering, heavy-equipment arrangements, inspections and coordination with subcontractors, one to two months ahead of the work. It presents delay risks with the supporting rationale, and it analyses discrepancies between the programme and the daily reports to surface likely impediments. Fujitsu’s published screens show two of those outputs: recommended confirmation items, and the permits and applications required for the location of the works.
Why this is a more interesting experiment than a schedule predictor
Most construction AI announcements point at a model that forecasts a completion date. This one points at the gap between the programme and the administration that has to happen for the programme to be legal and supplied. Delay on a job like this rarely arrives as a mis-estimated duration; it arrives because a road-occupancy application went in a fortnight late, an inspection was not booked before the pour, a crane was not confirmed for a week that two trades both assumed they had, or a long-lead item was never ordered because the person who knew it was needed was on another site. Those are lookahead failures, and the record that would have caught them exists — scattered across a schedule, a diary, an approval chain and an inspection log that nobody reads together.
The evaluation criteria are the tell that this is engineering rather than marketing. Fujitsu says the trial measures the accuracy of detecting planning omissions and delay risks, the usefulness of the underlying information the system cites, and the effect on site managers’ workload — and, separately, whether the system can cope with the different data formats and operating processes of two different contractors on the same site. That last test is the one worth watching. Anything can be made to work on one company’s templates. Two builders, two sets of drawings, programmes and daily report conventions, one engine, is a real generalisation test, and it is the reason the same idea usually dies between the pilot and the second customer.
The insistence on rationale matters just as much. A flag that says “delay risk: high” is unusable on a site, because a site manager cannot act on a score — they can only act on “the scaffold inspection for level 3 is not booked and the pour is on the 14th”. Fujitsu describes the system as presenting candidate omissions and risks with the supporting reasons, and the trial explicitly scores whether those reasons are useful. That is the correct thing to measure, and it is measured far too rarely.
The honest limits
There are no results. The trial started on 3 August and runs to 25 December 2026, so every capability in the announcement is an aim, not an outcome: no detection rate, no false-positive rate, no count of omissions found before they bit, no measured reduction in a site manager’s hours, no delay avoided. Nor is there a baseline — nothing states how many arrangement or application omissions these sites currently experience, which is the number any later percentage would have to be read against. Fujitsu names no product, no availability date and no price, and its Kozuchi AI platform and Digital Annealer are described as things it plans to use to advance the verification, not as what is running today. Integration with photo data, point clouds, drawings, BIM and company knowledge is stated as future work.
The setting is also unusually favourable, and that cuts both ways. This is the owner’s own site, with the technology vendor as the client, two contractors who have agreed to hand over their site data, and a redevelopment programme with a decade to run. Data access on that basis is not what a subcontractor on somebody else’s job gets, and the willingness of the participants is part of the result. The regulatory reasoning does not travel either: the value in “which applications does this location require” comes from encoding Japanese national and municipal requirements. In Australia the same function would mean the relevant state and local instruments — road-occupancy and traffic-management permits, dilapidation and out-of-hours approvals, hold points under an inspection and test plan, notifications to the regulator — encoded per jurisdiction. Nothing about proving it in Kawasaki proves it in Parramatta.
And a system that reasons over the site record inherits the quality of that record. If the daily report is a photograph of a whiteboard, if approvals live in one person’s inbox, if the programme on the wall was last updated in June, there is nothing for a model to cross-check and the failure will look like an AI problem when it is a records problem. That is not a criticism of the trial — Fujitsu is testing against real contractor data formats, which is exactly the right exposure — but it does determine who could use the result.
What it means for operators
The transferable lesson is available now, before any of this ships. Every one of the failures this system is hunting is invisible only because the evidence is fragmented: the lookahead in one tool, what actually happened in a diary, the permit in an email, the inspection in a spreadsheet, the procurement in someone’s head. Bring those into one record that reconciles — programme against actuals, arrangements confirmed or not, approvals with dates, inspections with hold points — and a competent site manager finds most of the omissions themselves, three weeks out, without a model. That reconciliation is also the precondition for any of this class of tool ever being useful to you.
When a vendor does bring one to your tender, ask the questions this trial is asking. What is the detection rate, and against what ground truth — someone had to go back and count the omissions that actually occurred. What is the false-positive rate, because a lookahead that raises forty flags a week will be ignored by week three. Does it cite the record it reasoned from, so the flag can be checked in seconds. Which jurisdiction’s permits and inspection requirements are encoded, and who maintains them when they change. And what happens to your data: what it reads, where it is processed, who else’s site it improves. A four-month trial with two contractors and no numbers yet is not a product — but it is a reasonable specification for what to demand from anything sold to you as one.