Hello Project AI enthusiasts,
This week’s ‘Feature’ is an unusually direct account of how AI is already moving through surveying, including fabricated evidence credible enough to trigger a re-inspection. Around it sit four practical signals: Microsoft is bringing metered agents into the M365 workplace, frontier model prices have fallen again, an OpenAI agent has been treated as an intruder by an Australian government system, and Oracle’s Stargate campus has exposed the contractual pressure around power and permits.
Featured This Week
Litmus publishes its first report: surveying is adopting AI faster than it can explain it
The first Litmus report, Faster than we can explain, draws on a group that meets quarterly under the Chatham House Rule to discuss what senior practitioners are seeing across UK surveying. Its five findings are qualitative and uncomfortable precisely because they are practical. One contributor, Chris de Gruben, estimates that 80 per cent of the valuation process could be automated. That is a practitioner estimate, not a measured finding from the report.
Firms completed visible compliance work after the RICS Professional Standard on the responsible use of AI took effect on 9 March 2026, then lost momentum. Adoption does not track firm size. Exposure accumulates through staff, subcontractors, and clients whether a firm has made a formal decision or not. Complaints are changing too: fabricated evidence can now be credible enough to force re-inspection.
The report’s central delivery implication is about judgement. The instinct that catches an error before it becomes a claim is built through practice. If firms automate entry-level work without protecting routes into professional experience, they may spend that scarce capability faster than they can rebuild it. Litmus is preparing a broader survey for Autumn 2026.
Disclosure: Project Flux founder James Garner is a founding Litmus contributor in a personal capacity.
Editor’s Picks
This week's must-read stories on AI, project delivery, and infrastructure:
Microsoft’s Copilot super app arrives, with agents on the meter
Microsoft’s new Copilot structure brings Home, Code, and Autopilot into a more unified workplace experience. Home combines chat and Cowork, Code can build apps and trackers inside a tenant, and Autopilot is a persistent agent with its own identity and email. For delivery firms, the budget question is central: several agent and frontier-model capabilities add usage-based billing to the seat licence.
Claude Opus 5.5 and GPT-6 Sol and Luna cut frontier AI prices in one day
Anthropic released Opus 5.5 at lower pricing, and OpenAI followed with GPT-6 Sol and Luna, cutting their API prices by 50 per cent against the previous GPT-5.6 promotional pricing. The practical consequence is a cheaper route to model choice. Savings still depend on routing the right task to the right model and measuring output quality.
OpenAI agent breached an Australian government portal. Check your contracts
Australia’s prime minister said an OpenAI agent accessed public and non-public files on a Medicare statistics portal on 18 June. OpenAI said its review found no patient records were accessed and described the activity as unintended model behaviour. The delivery lesson is contractual: public-sector frameworks need clear rules for vendor agents touching client systems, data, and security boundaries.
Oracle declares force majeure on its 2.45GW Stargate campus after gas-pipeline permits fail
Oracle has sent a force majeure notice to the developer of Project Jupiter, a Stargate data-centre campus in New Mexico. The notice would allow Oracle to delay payments if the facility misses its 2028 target, while Oracle and Blue Owl both say the project remains on schedule and the financial commitments stand. The Energy Transfer gas pipeline has been delayed nearly six months to 1 February 2027, and the air-quality permit has a 23 November decision deadline.
AI in AEC and Projects
Here are some intriguing reads, specifically for AEC and project delivery professionals:
The programme makes AI project delivery a formal learning pathway. It matters, as evidence from Stanford suggests AI-adopting firms are changing the balance between junior and senior roles.
The study analyses 1.25 billion job postings and finds a smaller junior share at AI-adopting firms, raising a direct question about how AEC will preserve routes into judgement-based work.
The paper puts projected US AI investment at 3.63 per cent of GDP annually through 2032 and warns that joint ventures and private credit can make correlated project risk harder to see.
Personal Pulse, Assistant Spaces, and Assistant Builder are designed to connect project context with downstream effects. The expanded experience is planned for 2027, so treat it as a preview rather than an available capability.
Curated Links
Selected stories worth reading before you get back to delivery:
Perplexity tests network confinement for AI agents. The study found no VM-to-host escape in 108 runs, while network-policy bypasses appeared before remediation.
Databricks acquires Row Zero. Governed live spreadsheets address the familiar problem of sensitive cost and programme data leaving controlled systems.
Google prepares Project Suncatcher for orbit. The engineering challenge is a useful reminder that compute capacity depends on cooling, radiation, connectivity, and physical deployment.
Claude discovers a novel enzyme system with CRISPR-like repeats. Anthropic’s account shows how agent swarms can narrow large research searches, while human scientists still verify the result in a laboratory.
Akamai announces an $11.6 billion multi-year agreement with Anthropic. The deal illustrates how agent workloads are creating demand for CPU capacity alongside the better-known GPU race.
Tool of the Week
FastDraft: contract administration with clearer action visibility
Built Intelligence’s FastDraft supports NEC3, NEC4, JCT, FIDIC, IChemE, and bespoke contracts. Its workflows cover notices, early warnings, compensation events, approvals, reminders, and audit trails. The useful test for a QS or PM is simple: can the system make deadlines, responsibilities, and the record of decisions easier to see across a live project? It is contract-management software, rather than an AI-native product, and that is worth stating plainly.
Governance
Australia consults on standards for large AI data centres
Australia’s consultation proposes national standards covering energy, water, community engagement, costs, location, and workforce impacts for large data centres. Submissions close at 5 pm AEDT on Friday 9 October 2026. For planners, developers, utilities, and local authorities, the signal is that AI infrastructure is entering the same public-interest conversation as other major projects.
California’s ADMT rules include opt-out and risk controls
The California Privacy Protection Agency says its completed regulations cover risk assessments, annual cybersecurity audits, and consumer rights to access and opt out of certain automated decision-making technology. The page lists an effective date of 1 January 2026, so firms should check the final rule text rather than rely on the tracker’s 2027 date.
Training
Anthropic Academy offers free structured Claude courses
Anthropic Academy provides free learning content through Skilljar and does not require an Anthropic account to access the course platform. For project teams, the useful starting point is capability and verification: learn how to delegate, describe, discern, and apply diligence before introducing AI into live project information.
Also This Week
Chris de Gruben on AI, automation, AGI, the future of work, and more [Podcast]
We explore the rapid advancements in AI, its implications for jobs, project delivery, and the ethical considerations [Podcast]
One More Thing
A 1.2-metre robot sailboat crossed the Atlantic without a crew
Oshen’s C-Star PC13 completed the Microtransat crossing after 127 days at sea, covering 2,926 nautical miles without intervention. The useful reminder for project teams is that autonomy can earn trust through narrow objectives, long observation periods, and a clear operating boundary.
Quick Win
Extract a schedule of rates from a tender PDF inside Claude
Adobe has added more than 80 tools, including Acrobat tools, to its Adobe for Claude plugin. A low-risk test is to use a copy of a tender or valuation PDF, ask Claude to extract the pricing schedule into a table, flag provisional sums, and compare the result with the source pages before anyone relies on it.
Event of the Week
Built Intelligence live webinar: Effective Early Warning Notices, 30 September, online
This 60-minute session covers drafting, compliance, notification timescales, tracking, accountability, and commercial control for NEC early warning notices. It is relevant for project managers, supervisors, contract managers, and anyone responsible for contract administration.
Till next time,
Project Flux
All content reflects our personal views and is not intended as professional advice or to represent any organisation.