Hello Project AI enthusiasts,
If your firm pays for Claude, Copilot or Gemini, this is the week the bill got a name on it. Accenture now funds $29 a month per person and charges the rest to project cost codes. JPMorgan has locked outside agents out of customer accounts, while Amex built a gate for them. Google opened Gemini 4 Argon to paying users and, more usefully, published what it did when left alone for 11 hours. Meta's Muse Code subscriptions train on your work by default. And Tesla hopped a car over a row of parked cars, which raises a question for anyone who designs streets.
Featured This Week
Your firm's AI budget now has your name on it: Accenture funds just $29 of Claude a month
404 Media saw the internal memo: from 1 October, Accenture centrally funds $29 of Claude per person per month for most staff, with anything above charged to a Work Breakdown Structure code. Premium-plan users get $250 of token usage a week. Some consultants had been spending $500 to $1,000 a month in Claude Code and are scrambling to keep access; one has lost it. Deloitte data published this month, from a May 2026 study, shows where the overflow goes: 49% of workers using AI at work already use unapproved tools, and senior leaders are 32% more likely than staff to do it.
We recently covered Microsoft cutting its internal Claude spending plans by a third. This week the cap has reached the individual consultant, and our industry runs on exactly the WBS codes and fee budgets Accenture is now using. Expect the next appointment letter to ask whether AI tooling sits inside the fee or is a recoverable cost. Expect the cheap workaround, a personal account with client data in it, to be the default wherever nobody has agreed what the tool is for.
Editor's Picks
This week's must-read stories on AI, project delivery and infrastructure:
On JPMorgan's Q3 call Jamie Dimon said the bank will not let outside agents act inside customer accounts because it is "too risky for clients and for JPMorgan". The same day American Express published an Intent Intelligence Framework that records what a customer authorised before an agent may spend. Your client portals, procurement systems and payment approvals are smaller versions of the same choke point, and an agent with a valid login will arrive at them within the year.
Gemini 4 Argon is open to paying users, and Google's own test shows the catch
Argon reached Pro and Ultra subscribers and the API on 13 October at $2/$10 per million tokens until 30 November. The document worth reading is Google's write-up of an 11-hour unattended run on a 205,000-line codebase: round one quietly changed a public interface and ran 216 of 16,892 tests; round two, with house rules written down and the goal pinned, ran everything and proposed six changes, of which the team shipped two. If Argon is now the engine in your Workspace tenant, the operating rule matters more than the benchmark.
Meta's Muse Wonder arrives priced like Opus, and trains on your code by default
Meta's first frontier models landed on 12 October with Active Memory that carries context across conversations. Wonder is priced at $4/$20 per million tokens, the same as Opus 5.5; Wonder Full is only available through Muse Code subscriptions. Meta's own documentation says those subscriptions default to a ‘Contributor’ mode that trains on what you submit, and its terms say not to submit confidential information while it is on. Check the setting before anyone in your firm pastes a tender into it.
What a hopping Tesla asks of the built environment
Tesla's next-generation Roadster used compressed-air thrusters to launch off a half-pipe and clear a row of parked cars at a closed site in Texas, with production pencilled for 2028 and the flying trim not approved for public roads. A Smithsonian curator called it hopping rather than flight. For planners, highways engineers and car park operators the useful part is the shape of the question: a hop has a launch point, an arc and a landing point, and all three happen within the footprint of ordinary streets.
Tool of the Week
Google Deep Review: two free AI reviews a week of any PDF
Google Research opened Deep Review to the public on 14 October after 16,000 paper reviews for STOC, ICML, NeurIPS and ICLR. Upload a PDF and a four-stage agent pipeline segments it, checks logical consistency, grounds claims and flags ambiguous terms; Google names legal and financial documents as targets. Two reviews a week are free, and documents are not used for training. Try it on a tender report or a method statement before it goes to the client, starting with something non-sensitive.
Quick Win
QS Quick Win: reconcile a contractor's application against your valuation in 20 minutes
Export the application and your valuation as two tables, paste both into ChatGPT, Claude or Copilot and ask it to sort every unmatched line into four buckets: missing on one side, duplicated, wrong amount, or wrong period. Ask for many-to-one groupings with a tolerance of 1% or £50 and some date slack, then check row counts and the unmatched total rather than the grand total, and keep a status column so the reconciliation is auditable. You end up with a list of specific lines to query rather than a figure to defend.
Event of the Week
Tomorrow, 21 October, 15:00 BST: RICS webinar on its 2026 AI in Commercial Property and Construction report
One hour of CPD on the report's findings: two-thirds of construction organisations now use AI, 37% plan to spend more, nearly two-thirds report low preparedness. Report authors: Anil Sawhney, Tom Storey and Katherine Pitman. Free for RICS members, £36 plus VAT otherwise. Can't make it? Digital Construction North, Manchester Central, 18 November, is free for industry.
AI in AEC and Projects
Three reads for AEC and project delivery professionals:
AI roadside cameras at two fire stations flagged around 17,100 seatbelt and 1,920 phone-use offences in a two-week trial in March; mayor Richard Parker called it "frankly shocking". These are potential offences, and the results are under review for future enforcement.
Brookfield and NextEra will build a 1.8 GW campus on the former DOE uranium site in Kentucky with 2 GW of dedicated gas and up to 2.6 GW of storage. After Oracle's force majeure and the Texas permit freeze, capital is choosing sites that bring their own power.
The September white paper models six AI levers on a completed San Francisco multifamily job: 17 to 20% off cost, 22 to 25% off schedule. The authors call the inputs directional. Board-level framing, modelled rather than measured.
Governance
Microsoft's Brad Smith says milk is more regulated than AI, and the UK should lead
On the BBC’s Big Boss Interview, Microsoft’s president, Brad Smith called for stronger rules governing advanced AI. He wants an emergency brake, often called a kill switch, for AI systems that get out of control. He also called for mandatory incident reports, with victims informed when AI agents break into systems. Finally, he argued that independent government evaluators should vet new frontier models before release, with the UK's AI Security Institute taking a leading role. Two days earlier Satya Nadella argued firms should treat every model as a compromised insider and keep the brake outside it.
FTC opens a formal investigation into OpenAI and Anthropic
On 14 October the FTC sent civil investigative demands to both labs on safety incidents, safeguards and consumer harm and plans to compel executive testimony. If your AI policy leans on a vendor's safety assurances, this is the regulator asking whether they were ever true.
Curated Links
Selected stories worth reading before you get back to delivery:
Anthropic gives six METR staff employee-level access and the right to publish without approval, subject to narrow redactions, two days after Reuters reported labs and evaluators negotiating over what evaluators may disclose.
Around 400 scientists including Hinton and Bengio call for an international body to log AI containment failures, arguing September's voluntary White House commitments are not enough.
Anthropic says small models are months from self-replicating across cheap consumer hardware; a 9bn-parameter model stood up inference on eight of nine device types in its tests.
A Harvard working paper on 718 firms finds AI coding agents lift lines of code about 30% while each change takes 49% longer to merge. Production gets cheap; checking does not.
Microsoft patched more than 1,400 flaws in October and credits AI for finding them; Bank of America hired 1,000 contractors to fix flaws exposed by AI models, including Anthropic's Mythos.
Antony Slumbers: AI gains in real estate land on NOI and get capitalised at 20x, if the people deal is right, with a six-part people deal for any workflow redesign.
Minerva emerges with $10m, and Roger, a teleoperated humanoid for bomb disposal, hazmat and offshore work, with first paid pilots this autumn.
Also This Week
The Future of AI and Cybersecurity: Robert Maxwell discusses the evolving landscape of cybersecurity, AI's role in security, and more [Podcast]
Project Flux Pro: we explore the latest developments in AI, including OpenAI's DOTS, persistent agents, and impact on various industries. [Podcast]
One More Thing
Meta's new frontier model spent 792 turns and $96 lost in a Pokémon forest
A developer ran Muse Wonder as an agent in Pokémon Red on launch day. It reached Viridian City by turn 59, then spent 792 turns and about $96 wandering Viridian Forest without reaching the first gym. Worth remembering before anyone leaves an agent alone with a programme.
Till next time,
Project Flux
All content reflects our personal views and is not intended as professional advice or to represent any organisation.