What Actually Changed in AI This Year

A year of announcements, most of which changed nothing for a company between five and fifty million.

Four things did. Here they are, with the evidence, and what each one should change about what you do in the first quarter.

One: the cost of a small judgement collapsed

Stanford’s AI Index for 2025 reported that the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024.

That is the most consequential number in this article and it gets almost no attention, because it sounds technical. It is not. It is the price of asking a machine to make one small decision, and it fell by more than two orders of magnitude.

What it changes: the arithmetic on small, repetitive tasks. Automating something that consumes forty minutes a day used to fail the business case because the per operation cost was real. It no longer is. Work that was correctly left to a person three years ago should be reconsidered, and the list of candidates in most businesses is long.

Two: adoption doubled in Canada, and is still low

Statistics Canada, in a release dated 16 June 2025, found that 12.2% of Canadian businesses had used AI to produce goods or deliver services in the preceding twelve months, against 6.1% a year earlier.

Doubling in a year, and still roughly one business in eight. Professional and scientific services led at 31.7%, information and cultural industries at 35.6%, finance and insurance at 30.6%. Construction, transport, retail and the trades sit far below.

What it changes: the window is real but it is closing at a measurable rate. If you are outside the top three sectors, most of your competitors are still not doing this, and that will not be true in two years.

Three: the failure rate got measured

MIT’s Project NANDA published The GenAI Divide in July 2025, reviewing over 300 publicly disclosed AI initiatives with 52 interviews and 153 survey responses. Against an estimated $30 to $40 billion of enterprise investment, roughly 95% of generative AI projects produced no measurable return.

The honest caveats matter here. It studies large organisations, not businesses your size, and no measurable return frequently means nobody built a way to measure. Both make the number less damning and more useful.

What it changes: it tells you the failure is in selection and follow through, not in the technology. The projects that die are pilots that never became systems. That is a solvable problem and it is mostly solved before anything gets built.

Four: buyers decide earlier than they used to

6sense published its B2B Buyer Experience Report on 12 November 2025, from nearly 4,000 buyer responses. The point of first contact moved from 69% of the buying journey to 61%. Buyers initiated 79% of engagements, and four out of five deals go to the pre-contact favourite.

Their median purchase is $200,000 to $300,000 in technology and services, so the numbers will not transfer exactly to a trades or construction business. The direction will.

What it changes: the period where you can influence a decision has shortened by roughly six to seven weeks, and most of it happens before anybody speaks to you. What a prospect can read on their own is now a larger part of the sale than what a salesperson says.

What did not change

Worth saying, because a year of announcements implies more movement than occurred.

Accountability did not move. A person still signs, and in regulated work that is not negotiable. Knowledge that was never written down is still unusable. A process nobody has agreed on still cannot be automated. And a system nobody owns still degrades quietly, which remains the most common cause of abandoned automation.

Those are the same four limits as last year, and they will be the same next year. They are scope rather than failure, and knowing where the boundary sits is what lets you spend confidently inside it.

How to read next year’s numbers

You will see a great many adoption figures over the next twelve months and most of them will be quoted without saying what they measured. Three questions make them readable.

Who was surveyed. Stanford’s 78% and Statistics Canada’s 12.2% look contradictory and are not. One asks large organisations whether AI is used anywhere in the business. The other asks Canadian businesses of every size whether it is used to produce goods or deliver services. Different populations, different questions, both correct.

Use or result. Almost every headline figure counts activity. Very few measure whether anything improved, which is why the MIT number lands so hard against the adoption numbers.

What period the data covers. A report published in 2025 frequently describes 2024. In a field moving this quickly that gap matters, and it is usually buried in a methodology note.

Ask those three and most of the noise resolves into a small number of facts you can actually use.

What to do in the first quarter

• Revisit one thing you decided against. Something that failed the business case two or three years ago on cost per operation. That calculation has changed by two orders of magnitude.

• Pick a process, not a tool. Daily, repetitive, currently waiting on a person being available. Write down one number before you touch anything.

• Refuse the pilot framing. Build the narrow version, put it in the path of real work, and switch the old route off. The 95% mostly stopped at the trial.

• Write the two or three pieces a prospect could read before meeting you. That period is shorter than it was and it is where the decision now happens.

None of that requires a strategy or a budget cycle. It requires one process, one owner and one number, and it is roughly a quarter of work.

Sources

• Stanford Institute for Human-Centered Artificial Intelligence, Artificial Intelligence Index Report 2025.

• Statistics Canada, Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025, released 16 June 2025.

• Challapally, Pease, Raskar and Chari, The GenAI Divide: State of AI in Business 2025, MIT Project NANDA, July 2025.

• 6sense, The B2B Buyer Experience Report 2025, announced 12 November 2025.

AI Optimize builds the narrow version that goes into real work, and tells you in week three when the honest answer is that a process should be fixed rather than automated. Start at Workflow Automation.