Week of 21 September 2026
If AI slows down, who gets ahead?
A slower frontier could offer breathing room. Unequal AI support and rising client expectations determine who benefits.
Executive thesis
Pacing AI
Slower frontier development could create breathing room while competitors and customers keep advancing with capabilities your organisation has yet to use.
Assess performance fairly
Factor differences in AI capability and access into regional comparisons before attributing stronger output entirely to talent, initiative or leadership.
Experience customer capabilities
Try ChatGPT, Codex and Claude in permitted settings; customers may use stronger tools and expect much faster service from you.
Why frontier leaders are debating a slowdown
“The frontier” is shorthand for the most capable AI in existence at any moment, built by a handful of laboratories. In September the head of one of them asked the rest to slow down.
One reason for the urgency is recursive self-improvement: AI helps researchers build better AI, which can then contribute more to the next generation. The prospect is compelling because progress could accelerate through this feedback loop. It does not mean systems already develop their successors entirely without people.
The prisoner’s dilemma: slowing alone can cost a lead
Each lab may prefer a safer collective pace while fearing that rivals will gain if it slows alone. That tension makes coordination difficult. Support for pacing is not an implemented industry pause, and pacing does not mean halting all progress. For your business, any breathing room depends on what you do with the time.
Why we may not feel the same urgency
Your everyday AI experience may show only a fraction of what is available. Model quality, computer-use capability, connected data and permissions all affect what work a tool can complete. If your staff mainly use a restricted assistant for drafting, the frontier debate can feel remote even while a substantial capability gap already exists.
Your approved tool can limit your view of AI
You may judge all AI by the limited capabilities your staff can access.
Your approved tool shapes the strategy you can see. If people can only draft with AI, they will not propose delegating work to it. Licence uptake cannot show how much capability the organisation has left unused.
Stronger AI support can look like stronger talent
Imagine the same junior analyst preparing the same Power BI dashboard, with identical data, skills and deadline. In one setting they have a limited chat assistant. In another they can use a stronger model equipped to operate software. Their underlying ability stays constant; the support around them changes.
The first analyst spends the deadline assembling a dashboard. The second starts with one and can challenge the numbers, investigate an exception or prepare the client recommendation. In a performance review, that difference can be mistaken for ability.
The video shows the type of task at issue. It is not a measurement of two analysts’ output.
Uneven access is already real in Hong Kong
Reports from April and June describe restrictions on the same product in Hong Kong.
Read these precisely
These reports concern particular access arrangements, not a universal ban on advanced AI. They show why a global tools policy can produce different working conditions across regions. They do not establish that either bank’s affected teams performed worse.
Account for AI capability when comparing performance
A performance review sees the delivered work. It does not see what the tools around the person were able to carry out.
Before moving responsibilities, investment or headcount between regions, investigate the AI support behind the result. Talent, local knowledge and market conditions still matter. Global leaders need to account for unequal support; local leaders need to make those constraints visible. Otherwise an equipment advantage can influence career opportunities as though it were purely a talent advantage.
Design services for clients who can delegate execution
Your compliance boundaries do not define the tools a customer can use. Computer-using agents can prepare requests and carry out supported steps, beyond gathering information. Even when authentication or confirmation prevents full execution, the client may already be ready to proceed while your process is still waiting for an employee.
In July, Radisson launched a ChatGPT app covering more than 1,000 hotels. In March, CNBC reported that OpenAI scaled back Instant Checkout after only about 30 Shopify merchants went live, shifting toward retailer-built apps.
A relationship manager should add advice, negotiation or an exception. Asking a client to repeat information their agent has already assembled makes the firm look slower than the technology the client brought.
Three questions before the next review
Breathing room becomes useful when you recognise the capability gap, account for it in performance decisions and experience the tools shaping your customers’ expectations.
If frontier development slows, what would still change in our market?
Assume competitors and customers keep improving their use of today's tools. A pause in new models would give you time, but it would not preserve your relative position.
Are we comparing people with materially different AI support?
Before allocating work, budgets or headcount across regions, record the models, data access and execution capabilities behind the results. Investigate their contribution alongside expertise and market conditions.
Have we experienced the AI our customers can use?
Have service leaders try a relevant journey with capable ChatGPT, Codex or Claude tools using public or synthetic data in a permitted environment. Identify one handoff that adds delay without advice, protection or resolution.