Most of what gets written about AI is written for people who do not run anything. It is written for investors, journalists, and the kind of executive who reads the headline and forwards the article without finishing it. The result is that operators, the people actually responsible for whether the work gets done, end up making decisions based on a picture of AI that has almost no relationship to the one their team is going to encounter on a Tuesday morning.
This is the short version we wish someone had written for us at the start of the year. What changed in the last twelve months that an operator should care about, and what did not, separated cleanly so the next planning meeting starts from something solid.
What actually changed
1. The conversation about agents got honest
A year ago, every vendor was selling autonomous agents that would run entire functions of a business. The demos were impressive. The production reality was not. Twelve months later, the people who run real systems have stopped talking about agents that do everything and started talking about agents that do one thing reliably. The hype has narrowed into something useful.
If a vendor is still pitching a single agent that will replace a department, the pitch is two years out of date. The companies getting real value are running small, specific agents on narrow workflows, with humans clearly in the loop. That is the standard to hold any proposal to.
2. The hard question moved from "does it work" to "who owns it"
For a while, the hard question was whether AI could do the job at all. That question is mostly settled for a wide set of common business tasks. The new hard question is what happens after the AI is doing the job. Who watches it. Who fixes it when it drifts. Who decides when to turn it off.
There is a term for what happens when nobody asks these questions in advance. It is called agent sprawl, and it looks like the shadow software problem from ten years ago, except the software is now making decisions instead of just storing files. This is the single most common operational problem we are now asked to come in and clean up.
3. Pricing models started to shift
Software used to be priced by the seat. You paid per user, per month. AI is breaking that model, because the AI is the user and it does not need a seat. Vendors are now charging by the task, the workflow, or the outcome. Some are charging by how much the AI thinks.
The budgeting muscle your finance team has developed over the last decade is the wrong muscle for the next decade. A tool that costs a predictable amount per user per month is becoming a tool that costs a variable amount depending on how much work it does. Forecasting usage is going to matter more than forecasting headcount, and the people doing the forecast will need to understand the work, not just the spend.
4. Model cost stopped being the bottleneck
Two years ago, a serious AI feature cost serious money to run. Today, the same feature costs a fraction of what it did, and the trajectory is still going down. The bottleneck has moved. It is no longer the model. It is the work of connecting the model to your actual business, your actual data, and your actual people.
This is a quiet but important shift. It means the right people to invest in are not the ones with the cheapest model access. They are the ones who know how to do the integration work, the data work, and the change management work. The model is a commodity now. The system around the model is not. That is most of what we do.
5. Operators started saying no more often, and it started to matter
The most interesting change this year is harder to measure. A year ago, the social pressure inside most companies was to say yes to any AI proposal, because saying no looked like falling behind. That pressure has eased. The number of operators who have lived through a failed pilot is now large enough that "we tried that, here is why we stopped" has become a respectable sentence in a board meeting.
This is good news. The next round of AI work is going to be done by people who have learned something, which is a higher quality of buyer than the market had last year. If you are one of those operators, your skepticism is an asset. Use it.
What did not change, despite the headlines
1. AGI did not arrive, and the timeline is not the point
A lot of column inches were spent this year on whether artificial general intelligence is two years away or twenty. The honest answer is that nobody knows, and the more important answer is that it does not matter to your operational planning. The AI you can deploy in your business this quarter is the same AI you could deploy last quarter, plus or minus some improvements at the margin. Plan for what exists, not for what was promised at a keynote.
2. Quantum did not change anything in your business
Quantum computing is interesting. It is also, for an operator running a real company, irrelevant to any decision you will make in the next three years. If a vendor mentions quantum in a pitch about something other than cryptography research, they are reaching for a headline. Ignore the slide.
3. The AI workforce replacement did not happen
The companies that announced large layoffs and attributed them to AI mostly did the layoffs for other reasons, and the AI line was good press. Where AI has actually changed headcount, it has done so quietly and at the margin, usually in roles where the work was already in question. The "AI replaced our team" story has not been written honestly very often, because the honest version is not as dramatic. Your hiring plan does not need to be torn up. It might need to be adjusted. Those are different things.
4. The model wars did not produce a winner that matters to you
A great deal of attention went to which company has the best model. From an operator's standpoint, the differences between the leading models are now smaller than the differences between a well-designed system and a poorly-designed one. Picking a model is not a strategic decision. Picking a way to work is.
What to do with all of this
If we had to compress this into a single suggestion for the planning meeting, it would be this: spend less time on the question of what AI can do, and more time on the question of who in your business will own what it does. The first question is mostly settled. The second one is where the work is.
That is most of what we end up doing inside a business. Not picking models, not vetting demos, but figuring out who owns which agent, how it gets watched, and how the operating plan changes around it. The technology is the easy half. The half that is harder, slower, and more valuable is the operational layer that sits on top of it.
We will publish this piece again next year, with the same structure and a new set of answers. The hope is that it stays short, because the things that actually change in a year are usually fewer than the headlines suggest.
