Staying ahead
How to keep up with AI without chasing launches
Another week, another AI launch, and someone on LinkedIn says your business is already behind. You have a shop to run, and no time to test every new model that gets a headline.
What actually changes when a new AI model launches?
For most small businesses, very little changes on launch day. The model you already use still drafts the same replies and reads the same invoices. A launch only matters to you if it changes what your job costs, what happens to your data, or whether the AI can now do a task it couldn't before.
Everything else is somebody else's news.
Most launches don't touch your week
A new model may score higher on a test written by engineers. That tells you nothing about whether it writes a better quote for a boiler service or a kinder reply to a café customer whose order went wrong. Your jobs are narrow, and the leading models already handle most narrow jobs well. Stanford's 2025 AI Index report found the top two models were separated by just 0.7%.
Three changes that do matter
A price change on the plan you pay for. A change to how your data is stored or used on that plan. And a new ability that your job genuinely needs, such as reading a scanned delivery note it used to get wrong. If a launch brings none of these, you can let it pass.
Why the headlines feel urgent
AI news is written for people who sell AI, build AI or invest in it. Urgency suits them. A bookkeeper who switches tools every month spends that time re-checking work instead of doing it, and gains nothing a customer would notice.
- 60 daysminimum notice Anthropic gives before retiring a publicly released model
- 6 monthsminimum notice OpenAI gives before retiring a generally available model
- 0.7%the gap between the top two models in Stanford's 2025 AI Index
Source Anthropic model deprecations page; OpenAI deprecations page; Stanford HAI, 2025 AI Index report.
A 30-minute monthly routine for staying current
- 01
Keep a one-page list of your AI jobs
Write down each job you use AI for, the tool and plan it runs on, and what good output looks like. A plumber might have two lines: enquiry replies drafted in the morning, job notes written up in the evening. This page is what you check launches against.
- 02
Save one test job for each
Keep a real example you can run again: last month's awkward enquiry, a messy invoice, a set of job notes. Write down what the right answer looked like. This becomes your own benchmark, and it is the only one that measures your business. Anthropic's own guide to building tests for AI gives the same advice to developers: make the test mirror your real tasks, including the awkward cases.
- 03
Read your provider's own page, once a month
Skip the commentary. Go to the announcements or release notes from the company whose tool you pay for, and look only for the three changes that matter: price, data and a new ability your job needs. Ten minutes is usually enough.
- 04
Re-run the test before you change anything
If something looks promising, run your saved test job on the new option and compare it with the answer you wrote down. Better on your job means worth a closer look. Better on someone else's demo means nothing yet. Our guide to trialling AI on one job sets a simple pass or fail bar.
- 05
Change one thing, then write it down
If you do switch, switch one job, not all of them, and note the date and the reason on your one-page list. If it goes wrong, you know exactly what changed and you can put it back.
- 01Your job listeach AI job, its tool and plan
- 02Provider's own newsprice, data, new ability
- 03Saved test jobre-run only if something changed
- 04Decideswitch one job, or do nothing
- 05Write it downdate and reason on the list
Where should you read AI news?
Read the release notes of the tool you pay for, from the company that makes it, and very little else. Those pages tell you what changed on your plan, in plain terms, without the opinion layered on top. Bookmark one page per tool you use.
If you use Claude
Anthropic keeps release notes for the Claude apps in its help centre, listing new features by date. That is the page to scan for changes to the plan you are on.
If you use Gemini in Google Workspace
Google's official Workspace Updates blog posts new features for Workspace customers, with a weekly recap. Look for the posts that mention Gemini and the apps you use.
If you use Copilot in Microsoft 365
Microsoft publishes a roadmap of AI features for work with estimated release dates. Useful for seeing what is coming to Outlook, Word and Teams before it lands.
If someone built you a connection
If your AI connects to your own systems through a provider's API, the developer notes matter too: OpenAI keeps a developer changelog and Anthropic keeps one for its platform. Whoever built the connection should be reading those for you.
Why do old AI models get switched off?
Providers retire older models to make room for new ones, and give notice first. If you use AI through an app, the app moves you across and you rarely notice. If a developer connected AI to your systems, the connection has to be moved before the date, or it stops working.
How much notice you get
Anthropic's model deprecations page says it gives at least 60 days' notice before retiring a publicly released model. OpenAI's deprecations page promises at least 6 months for generally available models, and much less for preview models. Plenty of time, if someone is watching.
This is the one launch-related job you can't skip
A retirement is not news you can ignore. If a background job runs on a model with a retirement date, it will fail on that date. Re-run your saved test on the replacement well before, and move once it passes.
Your set-up survives the move
When the model changes, your written instructions, your limits and your checks stay the same. That is why we put the effort there, as explained in what harness engineering is. A new model under the same set-up is an afternoon's check, not a rebuild.
A worked example: a café's monthly check
Here is the routine in practice for a café that uses AI for two jobs. It is an illustration, with example details rather than a real client's records.
The job list
Two lines. Replies to booking and allergy questions, drafted in the morning on the business plan of the AI built into their email. Weekly social posts from photos and the specials board, drafted on the same plan. Each has a saved test: last March's awkward group booking, and one week's specials.
The ten minutes of reading
The provider's update page lists a new image feature and a change to how long chats are kept. The image feature might help with the social posts. The retention change is worth a note on the list, because allergy questions can include health details.
The decision
The owner re-runs the specials test with the new image feature. The posts look better, so she switches that one job and writes the date down. The booking replies stay exactly as they were. Total time: about 30 minutes (example figure), and nothing that was working got broken.
When should you switch AI tools?
Switch when your own test job shows a clear gain you can name, such as fewer corrections, a lower bill or a task it can now do. Stay put when the only evidence is a launch video or a friend's enthusiasm. The cost of switching is real: new settings, new habits and a fresh round of checking.
Switch when the test says so
If the new option gets your saved test right and the old one didn't, that is a reason. If both get it right, the cheaper or simpler one wins, and that is often the one you already have. Our post on which AI model a small business should use covers the choice in more depth.
Watch your plan's terms more than its features
The change most worth catching is quiet: a different price, a new limit or a change to how your data is handled on the plan you use. Those affect every job at once, and they are easy to miss behind the louder feature news. Our post on free or paid AI for business lists what to look for.
Keep the set-up when you change the model
Your written instructions, the limits on what the AI may touch and the checks on its work are yours. They carry across to a new tool. That is why time spent on them pays back for longer than time spent chasing the newest model. Keep the instructions in your own document, as set out in how to write AI instructions it will follow.
Common mistakes when keeping up with AI
Falling behind is rarely the problem. Wasting time on changes that don't matter, and missing the two that do, is.
Switching everything at once
Moving every job to a new tool on the same day means that when something breaks, you don't know which change caused it. One job at a time, with the date written down.
Taking advice from people who sell the news
Feeds full of "you're already behind" are selling attention. The test of any tip is simple: does it change what one of your jobs costs, how your data is handled, or what the AI can do for you? If not, scroll on.
Having no saved test
Without a saved example and a written right answer, every comparison is a feeling. Ten minutes to save one real example per job turns every future decision into a quick check. When a test catches a slip, add a rule, as in every AI mistake should become a rule.
Forgetting the other half of keeping up
Staying current is also about whether AI assistants mention your business. That moves month to month too. The AI Visibility Monitor asks the same buyer questions every month and reports who is named, and our post on getting found in AI answers covers what moves it.
Questions
08Asked and answered
Once a month is enough for most owner-run businesses, as long as you check against your own jobs rather than the news. Half an hour covers the provider's own update page and one saved test. Check sooner only if a model you rely on gets a retirement date. Start with one job, as set out in where to start with AI.
Usually not: the leading models already handle most small-business jobs, and the set-up around them matters more than which one is newest. The gap between the top models is small and changes every few months. We explain why the set-up matters more in what harness engineering is.
Run the same saved job on both tools and compare each answer with the one you wrote down beforehand. Count the facts it got wrong and the minutes of editing each answer needed. Fewer of both is better; anything else is a feeling. Our guide to trialling AI on one job sets a simple pass or fail bar.
When a model is retired, apps usually move you to a newer one, but a connection built on the old model stops working on the retirement date. Providers give notice: Anthropic says at least 60 days for publicly released models. Its model deprecations page lists the dates, so test the replacement well before.
Check the price, the usage limits and what happens to the data you put in on your plan. Those three affect every AI job at once, so they matter more than any new feature. The differences between plans are covered in free or paid AI for business.
Not if your instructions, limits and checks are written down outside the app, because they move with you to the new tool. What breaks is anything that only lived in one app's settings. Re-run your saved test after the move. Our safety rules show the limits we write down for every job.
Yes: we set up AI jobs at fixed prices and keep them working, so watching for changes and retirements is part of the job. With Autonomous Agent Operations, the monthly fee covers adjusting jobs when your tools change. See how it works for the three steps.
This is general guidance, not legal or financial advice. For your own situation, speak to a qualified adviser.