If I had ten minutes with a CEO about to sign off next year's AI budget, this is what I would say about the last three months: the technology stopped being the reason to wait. In a single quarter it became more capable, cheaper and easier to buy. What decides the result now sits inside the company.
📣 📣 Before you read on: we are running a Europe-focused AI adoption benchmark study for PE-backed companies, publishing in Q4.
Most studies count who has started with AI. Ours measures how much of the business actually runs on it, and compares companies across Europe.
As a participant, you’ll get (at no cost):
✓ Your company's level on a five-level scale, with notes on your sector
✓ The full findings before your competitors see them, while 2027 budgets are still open
Only 50 places, closing 31 October.
Plenty of warnings about AI risk, little action to match
In September the heads of Anthropic, OpenAI and xAI publicly agreed that AI development should slow down, and Anthropic's CEO set out the risks that worry him. Apart from pledges by Anthropic and OpenAI to let outside evaluators test their models, nothing concrete followed. Ten days later both released new models.
Our read: from the outside, nobody can judge how close these risks really are. The people raising the alarm are the same people shipping the next model. So treat the headlines as noise for your planning and a signal for your governance: decide now what AI in your company may do without a person checking.
The economics of AI tokens are changing fast
Cost used to come up in almost every AI conversation. In the last few weeks it dropped sharply. On 22 September Anthropic cut the price of its top Opus model by 20%, and OpenAI cut its GPT Sol model by 50%. Open models you can run yourself are now good enough to matter too.
The cost of reaching the same AI performance has fallen about 47% a quarter since 2023, faster than electricity, batteries or computing ever did (trend estimate). Chart: Epoch AI, Sept 2026,
What this means for your budget: cost still matters, but it may soon stop being the biggest variable. Weigh it against the value of the work and the effort to adopt, and recheck as prices fall.
New kinds of AI models are arriving too. Jev, from a start-up founded by one of ChatGPT's engineers, writes nothing. It only makes quick decisions inside software, like routing or scoring, and its maker claims it is up to 445 times cheaper than a top model. Not every task needs a frontier LLM.
AI agents are getting more mainstream
This was the quarter agents moved from something IT teams build to something anyone can download and start using. They connect to your systems and take action for you. Meta's Muse, launched on 8 September, books trips, sends emails, fills in forms and makes purchases for the user. It was downloaded faster in its first two weeks than ChatGPT was at launch, and Meta's shares jumped more than 11% in a single day.
What this means for your team: your people will not wait for a policy. The decision that matters now is where an agent may act on its own and where a person signs off first. Even among large US companies, only about half have written that down.
China is setting the floor on price
Chinese labs such as DeepSeek, Alibaba and Moonshot now release models close to the best American ones for a fraction of the price.
For a European company this cuts both ways: an open model on European servers keeps your data close, but still needs a security review, and in some sectors a Chinese provider is a board conversation of its own
EU AI Act: what applies now
Since 2 August, Europe's AI transparency rules apply. People must be able to tell when they are talking to an AI, and realistic AI-generated images, video or audio must be labelled.
What this means: any AI your company puts in front of customers, from a website chatbot to an AI-made campaign visual, now needs a clear disclosure, and breaches can cost up to 3% of global turnover.
What we are seeing in our deployment work
Here is something I did not expect a year ago. The smarter the models get, the more companies ask us for help deploying them. You would think better AI needs less support. In practice it is the other way round.
A more capable model can do more, so the questions get bigger: which process should it run, what data can it see, who checks the output, and who owns it when something changes. None of that comes in the box.
AI literacy has gone from nice to have to urgent. With agents a download away, your people will use their own tools anyway, and training them on your data, processes and rules is what turns that into company capability.
The question to take into your 2027 AI strategy meeting
If every AI tool you wanted were available, affordable and allowed tomorrow, what would still hold you back?
This quarter moved all three in your favour. What's left usually sits inside the company: your data, your processes, who owns what, and how well your people use the tools. That's where next year's plans will do the most.
Want to know where your company stands? One 30-minute call before 31 October gets you your level and the findings, while there is still time to shape your 2027 budget. Take part in the study →
Until next time,
Pooja
PS: If you found this useful, please share it with a colleague who is working on next year's AI strategy & budget.