Pack A Go Bag

Every business I've worked with has one vendor they can't fire. Usually it's the CRM. Sometimes it's the guy who built the website in 2017 and still holds the domain login. You don't notice the trap until the invoice goes up, and then you notice it all at once.

For a lot of you, AI is quietly turning into that vendor.

On October 26, Anthropic made its S-1 public. It plans to list on Nasdaq under "ANT," with a debut targeted for Thanksgiving week. The numbers are a lot to take in. Revenue for the first half of 2026 came in at $16.3 billion. The net loss was $80.7 billion, though $72.4 billion of that is a non cash accounting charge tied to convertible notes. Operating cash flow was a positive $3.9 billion for the six months, compared to negative $1.5 billion a year earlier. And nearly half of revenue flows through cloud marketplaces run by Amazon, Google and Microsoft.

I'm not here to tell you what to think of the stock. I don't give stock advice and you shouldn't take it from a newsletter. What I care about is one line Fortune pulled from the filing: a drop in demand, including from pricing pressure, would have a "direct and disproportionate impact" on revenue.

Translation for an owner: once a company answers to public shareholders every quarter, pricing and packaging become levers. Plans get reshuffled. Features move up a tier. Limits tighten. I'm not picking on Anthropic here. Every public software company does this, and OpenAI and Google are under the same pressure.

You don't have to wait for the IPO to see the direction.

On the same day, Apple's newsroom post on Siri said certain Apple Intelligence features that rely on server side models "are subject to daily usage limits," and that "expanded access to such features will be available for a fee in the future." That's Apple, the company that sells AI as a free perk of buying a phone, telling you the perk has a ceiling.

Microsoft is further along. Its Copilot Credits documentation describes credits as "a common currency" for services with usage based billing, and lists things like Cowork, Code and advanced work in SharePoint and OneDrive. A Copilot license is the entry ticket. The heavy agent work runs on a meter. And one detail every Microsoft shop should catch: spending policies have an "Auto apply new services" setting that's enabled by default, so new metered services can land under a policy you set months ago.

Your bill probably won't double next month. I'd just expect today's deal to look different by this time next year. So here's the question for this audit.

If your main AI vendor raised prices 30 percent or changed the plan you depend on, could you move in a week?

Most people say yes without thinking about it. Then they actually try it and find out the answer is no.

You've built more inside these tools than you think. Custom instructions you tweaked over six months. A project with forty uploaded files. Memory that knows your pricing, your clients and the fact that you hate the word "synergy." A custom GPT your front desk uses every day. A Make scenario with a model call buried at step four.

That's where your real investment is, and most of it lives on somebody else's servers.

Open a spreadsheet and call it AI Go Bag. One row per asset, with these columns:

Then walk every tool. Open each project, each GPT, each Gem, each saved prompt library. Open every automation in Make.com and note each module that calls a model, and which vendor's key it uses. If you run Go High Level, check Conversation AI and Voice AI settings and write down what's configured there, because that's a model you're renting through another vendor.

Ask your team one question in Slack: "What AI tool do you use every week that I might not know about?" Give it a day. Somebody always names one.

Sort by hours to rebuild, biggest first. That's your priority list.

Do this now, while everybody still likes you. Nobody thinks about exports until the week they need one.

Per OpenAI's help article, go to your profile menu, then Settings, then Data controls, then Export data, then Confirm export. You get an email with a ZIP of your chat history and "other relevant account data." The link expires 24 hours after it arrives, and exports can take up to seven days, so don't start this the morning you need it.

Two catches. First, this self serve export isn't available for Business or Enterprise workspaces. If that's you, the workspace owner handles it. Second, the article doesn't promise that memory, custom instructions, custom GPT setups or project files come along. Assume they don't. Open your saved memories and your custom instructions in settings and copy them into a document by hand. Open each custom GPT you built and copy its instructions and list its knowledge files.

Per Anthropic's support page, on the web app or desktop go to Settings, then Privacy, then Export data. You can't start it from the phone apps. The link lands in your email and also expires after 24 hours.

On Team and Enterprise, members only see the button if an owner turned on "Allow members to export their own data." Member exports include your chats but not the contents of files you uploaded. The org's primary owner exports everything from Organization settings, then Data and privacy. Same rule as ChatGPT: the page doesn't say projects or memory are included, so copy your project instructions and keep the original files somewhere you control.

Gemini goes through Google Takeout. Hit Deselect all, check Gemini to get your Gems, then under My Activity include only Gemini Apps for chats, generated media and uploads. Pick a delivery method and Create export. Google says it can take hours to days.

Make lets you export a scenario blueprint as a file from the scenario editor. Do it for every scenario that touches a model. If you've built skills for Claude or anything else that lives as files, you already have them. Just make sure they live somewhere other than one laptop.

This next part makes everything else easier. Your work should live in your own folder, and the AI tool is just where you go to run it.

The work itself goes in one vendor neutral folder. Notion, Google Drive, a plain folder on a shared drive. I don't care which. Pick whichever one you already use and won't lose track of. It just has to be yours.

Structure it like this:

Then make one house rule: if you change a prompt in the tool, you change it in the folder the same day. Otherwise the folder rots in a month and you're back where you started.

If you want help writing that About us document, have the tool you use most draft it. Paste this in.

Read every line. Fix the [CHECK] items. Save it in the folder.

Most people find out their backup doesn't work on the day they need it. Let's not do that.

Back in September we covered building a fallback so a scenario survives an outage. This is different. This is proving the backup produces work you'd actually ship, before you need it.

Pick your top three assets from the spreadsheet. For each one, take the instructions out of your folder, paste them into a second vendor's tool cold, with no memory and no history, and run five real inputs from last month. Compare the output side by side with what your main tool produces.

Grade each one with three questions:

Question three is the one that pays off. Every answer is something living in your vendor's memory instead of your folder. Write it into the folder.

For automations, do the same in Make. Duplicate the scenario, swap the model module to a second provider, and run your saved test inputs through both. If the output is structured, like a category label or JSON for the next step, check that the second model returns exactly the same format. In my experience, formatting breaks a swap far more often than writing quality does.

If an asset fails the drill badly, don't panic. Fix the instructions now, on a calm Tuesday, instead of on the worst day of your quarter.

Now put a number on it.

Add three columns to the go bag spreadsheet: monthly cost today, cost at plus 30 percent, and cost to switch. Cost to switch is the rebuild hours times what an hour of your time is worth, plus any new subscription.

Then look at each vendor and answer one question: at plus 30 percent, would I stay or go?

Most of the time the honest answer is "stay," and I've got no problem with that. The point is that now you'd be staying because you ran the numbers, and you'd know what leaving costs.

Two things to watch while you're in there. Seat based plans and metered plans fail differently. A seat price that goes up hits you once, on a known date. A meter goes up quietly with usage, and you find out on the invoice. Flag every metered line in the sheet and put a budget alert on it wherever the vendor offers one. And if you're on Microsoft 365, go check that "Auto apply new services" setting on your spending policies and decide on purpose whether you want it on.

If you're on any business or team plan, send your rep or support three questions this week. Short email, nothing dramatic.

Save the answers in the folder. Pay attention to the vague ones. A vendor that dodges these questions is the one I'd watch closest.

Picture a six person bookkeeping firm. Your go bag spreadsheet comes back with fourteen rows. Ten are minor. Four matter: the client onboarding project in Claude, the custom GPT that drafts monthly close emails, the Make scenario that sorts incoming receipts, and the memory that knows every client's quirks.

The fire drill shows the close email GPT moves cleanly. The receipt scenario breaks, because the second model wraps its category in a sentence instead of returning one word. Twenty minutes of tightening the instructions fixes it. The client quirks were never written down anywhere except in ChatGPT's memory, so you spend an hour writing them into the folder.

Total time, maybe six hours across a week. And the next time a vendor raises prices, you open the spreadsheet and make the call.

Start the AI Go Bag spreadsheet and list every asset, with your honest rebuild hours.

Request exports from ChatGPT, Claude and Gemini today, since links expire in 24 hours and exports can take days to arrive.

Copy your memory, custom instructions and project instructions into one folder you own.

Run the fire drill on your top three assets with a second vendor and write down what the instructions assumed.

Add the plus 30 percent column and decide, vendor by vendor, whether you'd stay or go.

Send the three question email to every AI vendor you pay on a business plan.

Jordan

The AI Newsroom | Practical AI for people with a business to run.