Hello Project AI enthusiasts,
OpenAI's dots brings a persistent agent into a $100 ChatGPT plan, following Microsoft's Autopilot and Meta's Muse. Alongside it, a surveyor's £10.64m asset-strategy tool took 78 minutes to build, Google has gated Gemini 4 Argon to cyber defenders, OpenAI has shelved GPT-6.1 Astra over scope and authorisation failures, and Anthropic's research puts robot cost competitiveness at just 0.3% of physical tasks. The common thread is authority: who sets the brief, who checks the output, and who carries the consequence?
Featured This Week
OpenAI’s $100 dots can work overnight, but not for UK personal accounts
OpenAI launched dots at DevDay on 29 September: persistent agents with their own cloud computer and browser, access to more than 4,000 apps, and the ability to keep working after you close the laptop. The first dot is included with ChatGPT Pro and Business Premium. The $100 Pro tier is now the entry point, though personal Pro access excludes the UK, the European Economic Area and Switzerland. Business availability differs by supported region and workspace settings.
For a delivery firm, the important change is placement. The agent can sit near project email, documents and Teams channels, reading connected material on its own schedule. OpenAI's Custom Rules and Auto-review controls let administrators require approval, hand off sensitive actions or block them. The rules are instructions the dot tries to follow, and OpenAI says it can make mistakes.
That makes this a governance decision before it becomes a productivity experiment. Start with an archived project, narrow the connected folders, require approval for every external message and keep live cost and tender data out until someone owns the activity log. The first question is not whether a dot can draft a report. It is what the dot is allowed to infer when nobody is watching.
Editor’s Picks
This week's must-read stories on AI, project delivery, and infrastructure:
AI modelled a £10.64m sell-or-refurbish call in 78 minutes. What’s left for you?
Antony Slumbers used Codex to model a £10.64m sell-or-refurbish call in 78 minutes on a fictional office. It produced a £10.64m sell-now figure and exposed a modelling quirk that only a domain expert would challenge. For surveyors and cost consultants, the billable value is moving towards purpose, evidence, boundary testing and responsibility.
Google gates Gemini 4 Argon to cyber defenders first. There is no date for the rest of us
Google's first Gemini 4 model is rolling out initially through the Fairwind programme to vetted cyber defenders. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20. Workspace is not in the named first wave, so firms should use the waiting period to identify who owns the model switch, logs and connected data.
OpenAI paused its own training, shelved a model and dismissed three safety researchers
OpenAI paused training on its most capable models after a sandboxed agent used a DNS gap to reach an external chatbot and continued for two and a half hours after an alert. It then shelved GPT-6.1 Astra over reported deception and scope failures. The vendor's own evidence is a prompt for buyers to test authorisation, automatic approval and incident liability before deployment.
Robots: 0.3% of all job tasks
Anthropic's economics research estimates robots are technically capable of 74% of physical tasks in US jobs, yet cost-competitive for only 0.3%. At historic price declines, reaching 10% would take about 40 years. Contractors should price autonomous plant by task and unit rate, while placing the nearer-term retraining budget around office-based delivery work.
AI in AEC and Projects
Here are some intriguing reads, specifically for AEC and project delivery professionals:
The library pairs photorealistic Gaussian splats with aligned collision meshes, giving robotics teams a realistic way to train and test before a system reaches a live site. Construction is one of the categories you can vote for in the next batch.
EliseAI says its software, which automates leasing, maintenance and renewals, is used in one of six US apartments. Property operations show where narrow, repetitive workflows are scaling first in the built environment.
The power and cooling contractor raised $540m below its $20 to $24 range despite a reported $1.1bn backlog. The data-centre boom still carries supplier-side financing and valuation risk.
The Association for Project Management reports that 88% of project managers see AI capability as the most important skill for the next five years, alongside a wider warning about the UK delivery talent pipeline.
Matt Aromando's practitioner build turns property images or an address into a storyboarded listing video, showing how delivery professionals can productise domain judgement with existing models.
Curated Links
Selected stories worth reading before you get back to delivery:
NVIDIA launches an Open Agent Safety Platform. OpenShell and the Sentry watchdog put enforceable boundaries below the model, including quarantine outside the agent's control.
Perplexity Computer adds Automations for ongoing work. Scheduled and event-triggered runs remember earlier work, making permissions and review points part of the workflow design.
Cloudflare introduces Clef decision models. Typed outputs with probabilities could triage RFIs or change requests cheaply, while deferring uncertain cases to a human.
Tavus introduces Griffin. In the company's one-minute study, 26 of 54 participants thought its video avatar was human, so face-to-face verification deserves a rethink.
$16bn Hudson Tunnel project begins boring. A major rail project moves into tunnelling after funding disputes, with purpose-built machines and a 5,100-foot first section.
Tool of the Week
Microsoft Fabric IQ brings governed Power BI context into Copilot
Fabric IQ is generally available in Microsoft Copilot Chat and Cowork for Fabric and Power BI customers. It grounds answers in Power BI semantic models, metrics, relationships and business definitions, so a question such as “what moved the forecast?” can use the governed numbers behind the reporting pack. It requires Microsoft 365 Copilot and your existing Power BI access to the underlying model.
Governance
Rogue-agent risk is moving from lab incident to procurement issue
The UK AI Security Institute found that GPT-6 Astra, the version already released, completed simulated unsanctioned supply-chain attacks in 29.2% of runs, compared with 6.3% for GPT-5.6 Sol and 0% for GPT-5.5 (on a smaller set of runs). OpenAI’s cyber classifiers were switched off during the test.
Training
Google Cloud's Advent of Agents, Season 3
Google Cloud's Season 3 runs across October 2026 with short daily tutorials covering agent development, identity, permissions and security. The opening lessons include lifecycle foundations and SAIF guardrails. It is vendor-run and free to follow, yet the sequence gives delivery and IT leads a structured way to learn the controls behind the agents they may soon be asked to approve.
Also This Week
Itai Green discusses the urgency of innovation in the fast-paced world of technology [Podcast]
One More Thing
A Facebook Marketplace seller let Meta's Muse handle a listing. It accepted a lower offer, shared the pickup address and replied that it was present when the buyer arrived at about 9:15pm. The seller was not there. The lesson for project enquiries and tender correspondence is precise: “Allow Always” needs to mean the same thing to the user, the agent and the person receiving the message.
Quick Win
Compare three subcontractor quotes in a ChatGPT project
Futurepedia's walkthrough shows how to upload three quotes in PDF, email or spreadsheet form, ask ChatGPT to compare price, inclusions, exclusions and contract length, then branch into a line-item review and save the process as a reusable skill. Try it first with archived M&E or groundworks returns. Projects are on paid ChatGPT plans, and an analyst still checks scope gaps, provisional sums and qualifications before any client decision.
Event of the Week
CMAA webinar: Renovating While Operations Continue, 8 October 2026, online
The session covers renovation in occupied campuses and public buildings, focusing on safety, schedule certainty and stakeholder trust. It runs from 2:00 to 3:00pm EDT and offers 1 CMCI Recertification Point and 1 Professional Development Hour for eligible attendees. The webinar is free for CMAA members.
This week's newsletter is sponsored by:
Movar Reply is helping the construction industry make smarter decisions with AI. Their tools connect your planning systems, cost data, and risk registers into one unified data environment—then use AI to surface insights you'd never find manually. They even offer free AI tools for project reporting.
If you're curious about what AI can actually do with your project data, they're worth a look.
👉 Explore Movar Reply's AI tools: https://movar.group/
Till next time,
Project Flux
All content reflects our personal views and is not intended as professional advice or to represent any organisation.