For any advice firm that’s been experimenting carefully with AI and wondering when the real decisions start, this week is a reasonable answer: they’ve started.
On 2 and 3 June, OpenAI and Microsoft both made significant announcements about enterprise AI, and they point the same way. The AI Daily Brief described it as a move from a “subsidy era” (cheap, plentiful compute and exploratory use) to a “scarcity era” (constrained infrastructure, rising costs and pressure to deploy carefully rather than broadly).[1] For professional services firms that distinction matters more than it first looks.
AI is becoming a cost decision, not an experiment
The practical effect of the shift is that AI starts to look like infrastructure. Choosing which tools to run, on which workflows and at what cost, is becoming closer to a capital allocation decision than a trial.
SK Hynix, one of the main makers of the high-bandwidth memory AI servers need, plans to double manufacturing capacity by the end of the decade, and its chairman has said chip shortages could last until 2030.[2] That’s a structural constraint on the price and availability of compute, not a blip. For firms choosing tools or scaling agent workflows, cost per task becomes the number to watch, not cost per query.
Microsoft’s case for task-tuned models
Mustafa Suleyman, CEO of Microsoft AI, said at Microsoft Build this week that Microsoft’s MAI model, fine-tuned for McKinsey’s workflows, produced better output than GPT-5.5 on those tasks at a tenth of the inference cost.[3]
McKinsey’s workflows almost certainly don’t look like yours, but the principle carries over. A general-purpose frontier model is expensive at volume. A model tuned to one well-defined task can match it for a fraction of the price. Advice firms have plenty of standardised work (annual review letters, meeting summaries, research notes, suitability drafts for human review), and that cost gap is likely to show up as a real line in the technology budget within eighteen months.
That doesn’t mean you should start fine-tuning models yourself. It means the vendor you choose, and whether it offers workflow-specific tuning rather than raw frontier access, matters more as a selection criterion than it did six months ago.
OpenAI says the bottleneck is coordination
OpenAI’s report “The Next Era of Knowledge Work” cites McKinsey data that the average knowledge worker spends more than 25 per cent of the week on email and around 20 per cent looking for internal information.[4] OpenAI calls this “strange abundance”: technology made documents, dashboards and messages cheap to produce, so there are far more of them, and the bottleneck moved from creating work to reading, coordinating and signing it off.
Anyone running a firm of more than about ten people will recognise that. The compliance file nobody can find. The email thread that holds the actual decision. The paraplanner spending Friday afternoon chasing approvals instead of drafting.
OpenAI’s pitch is that Codex, its knowledge work platform, reduces that coordination load rather than generating more content. According to the same report, Codex has five million weekly active users, and non-developers make up 20 per cent of them and are growing three times faster than developers.[4] These tools are no longer mainly for technical teams.
“Knowledge work is still waiting for its factory redesign.” OpenAI, The Next Era of Knowledge Work[4]
People are running AI tasks in parallel
OpenAI also reports that 50 per cent of Codex users now run several tasks at once, up from fewer than a third in mid-April 2026.[4] People are running dataset reviews, script drafts and report assembly side by side, supervising several pieces of AI work rather than doing one thing at a time.
The way people use these tools is changing with it. “Ask the AI a question” becomes “give the AI several jobs and check what comes back”. For a senior adviser or practice manager, the better comparison is a capable junior who can work on several things at once, as long as someone reviews the output before it goes anywhere.
In a regulated firm that review isn’t optional. Any AI output that touches a regulated document, a client communication or a compliance decision needs a person to check it before it’s used. Working in parallel doesn’t change that. It changes how much one person can sensibly supervise.
UK regulators are hardening their expectations
The FCA, the Bank of England and HM Treasury said in a joint statement in May 2026 that frontier AI models now go beyond baseline cyber resilience capabilities, which moves AI governance from internal good practice towards a formal expectation for UK financial services firms.[5]
That’s the backdrop for this week’s announcements. The tools are maturing quickly, the costs are moving, and the regulators are firming up. Taken together, “we’re still working out our AI strategy” is harder to defend than it was a year ago.
CrowdStrike’s 2026 Global Threat Report records an 89 per cent year-on-year rise in attacks by AI-enabled adversaries.[6] The same capability that speeds up your internal workflows is making outside attacks more sophisticated. A firm that treats AI purely as a productivity tool, without thinking about security and governance, is building up a risk it probably hasn’t priced.
The US executive order and what it signals
The US administration’s executive order on AI this year shortened the pre-release notification window for frontier models and kept safety testing voluntary, with the stated aim of keeping pace on innovation.[7] Some commentators have argued that its classified review thresholds and loose “meaningful step change” test are quietly laying the groundwork for a future model licensing regime.[1]
It’s a US policy story, but it matters here for two reasons. Most of the AI tools UK firms use are built and regulated in the US at the infrastructure level, so US decisions shape what you can deploy and when. The direction of travel is the same everywhere: regulators are moving from hands-off to structured oversight. The FCA’s joint statement is the UK version of that.
My read on where this goes
Cheap, general-purpose AI for everything was a phase, and it’s ending. Within eighteen months I expect the firms that come out ahead to be the ones that matched each workflow to the cheapest tool that does it well, put the AI where the coordination friction is rather than where the drafting is, and can show an auditor who reviewed each output. Firms still paying frontier prices for every task, with no record of who checked what, will be paying more for less and carrying the regulatory risk on top.
This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.
Sources
[1] AI Daily Brief, ‘The Next Wave of Enterprise AI’ (June 2026). Podcast episode.
[2] Bloomberg, SK Hynix Plans to Double Capacity to Easy Memory Chip Crunch (June 2026). URL: https://www.bloomberg.com/news/articles/2026-06-02/sk-hynix-to-double-wafer-capacity-to-ease-memory-chip-crunch
[3] Microsoft, Microsoft AI CEO unveils 7 new AI models | Mustafa Suleyman at Microsoft Build 2026 (June 2026) Youtube Video. URL: https://www.youtube.com/watch?v=OvLIae4HCeM
[4] OpenAI, The Next Era of Knowledge Work (2026). Cited for knowledge worker share of Codex users (20%), growth rate (3x faster than developers), and the statistic that 50% of Codex users now run multiple tasks simultaneously. URL: https://cdn.openai.com/pdf/the-next-era-of-knowledge-work.pdf
[5] FCA, Bank of England, and HM Treasury, AI and Cyber Resilience: Joint Statement (May 2026). Cited for confirmation that frontier AI models now exceed baseline cyber resilience capabilities and that AI governance is a formal regulatory obligation for UK financial services firms. URL: https://www.fca.org.uk/news/statements/fca-boe-treasury-joint-statement-frontier-ai-models-cyber-resilience
[6] CrowdStrike, 2026 Global Threat Report (2026). Cited for 89% year-on-year increase in AI-enabled adversary attacks. URL: https://www.crowdstrike.com/global-threat-report/
[7] The White House, Executive Order on Artificial Intelligence (2026). Cited for reduction of the pre-release notification window for frontier AI models and voluntary safety testing provisions. URL: https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/