Your AI Tool Can Be Shut Off Overnight. Build the Replacement Map First.

Estimated read time: 11 minutes

Somewhere in your business right now there is a tool doing real work that you would struggle to replace by Friday. Maybe it drafts your customer emails. Maybe it transcribes your sales calls and the summaries feed your CRM. Maybe it is the thing your part-time contractor uses to produce the deliverable you sell.

Now ask a question most owners never ask: what happens if that company sends an email tomorrow saying service ends in sixty days?

This is not paranoia. It is the observed behavior of the AI tooling market. Products get acquired and folded in. Free tiers vanish. Model providers cut off access to a customer they have decided is now a competitor. Companies pivot away from the feature you depend on while keeping the brand alive, which is arguably worse because it looks like nothing happened. Tome shut down its slides product in 2025 while the company continued. Amazon retired Mechanical Turk, a service businesses had built workflows around for the better part of two decades.

None of that is a reason to avoid AI tools. It is a reason to know, in advance and in writing, what you would do. The exercise takes about ninety minutes and you will only ever do it once properly.

TL;DR

• List every AI tool that touches revenue or customers. Most small businesses find between six and fifteen and are surprised by several of them.
• Tier them by blast radius: business stops, business limps, business shrugs.
• For every tool in the first tier, name a specific replacement and confirm you can export your data today.
• Prefer tools that sit on a swappable model layer, keep prompts and workflows in your own files, and avoid annual prepay on anything under two years old.
• Revisit the map twice a year. It goes stale fast.

Why AI tools are riskier than normal software

Your email provider and your accounting software are boring, and boring is a feature. They have been around for years, they make money on every customer, and their business model does not depend on anyone winning a race.

AI tooling is structurally different in four ways that all point the same direction.

Most of them do not own their engine. A large share of AI products are a thoughtful interface wrapped around a model somebody else operates. That is a perfectly good business, but it means the product can be damaged by a decision made two companies upstream: a price change, a rate limit, a policy update, or a supplier deciding the customer has become a rival.

The unit economics are frequently unproven. Inference costs real money. Plenty of AI products are priced for growth rather than profit. Every one of those has a repricing event in its future, and the repricing is sometimes indistinguishable from a shutdown for a small customer.

The talent is worth more than the product. Acquisitions in this space regularly buy the team and retire the software. From the outside it looks like success. From your side it looks like a sunset notice.

Capability moves faster than contracts. A feature that justified a whole company in January can become a checkbox inside a platform you already pay for by September. That is good for you in the long run and disruptive in the short one.

We looked at the broader version of this question, which software categories AI is quietly eating, in our SaaS tools AI disruption audit. This piece is the defensive companion to it.

The five ways you actually lose a tool

“The company went under” is the least common of these. Plan for the other four.

1. The clean shutdown. You get notice, a date, and usually an export tool. Annoying, survivable, and the scenario everyone imagines.
2. The acquisition fold-in. The product joins a larger platform. Your workflow half survives, your integrations break, your price goes up, and the migration is on your schedule but on their terms.
3. The supply cutoff. The model provider underneath cuts the product off, sometimes because it now competes with them. The interface still loads. The thing it does stops working or gets noticeably worse.
4. The quiet degradation. Nothing is announced. Rate limits tighten, the free tier shrinks, the good model moves behind a higher plan, quality drifts. This one is dangerous precisely because there is no moment that forces you to act.
5. The pivot. The company is fine and healthy and no longer interested in your use case. The feature you built around gets deprecated while the logo stays up.

Four of these five give you weeks or months of warning if you are paying attention. The reason they hurt is not speed, it is that nobody had thought about the answer in advance and the thinking happens under deadline.

Building the replacement map

One spreadsheet. Seven columns. Do not overbuild it.

• Tool. The product name.
• What it actually does. Written as the job, not the category. Not “AI writing tool” but “turns our call notes into the follow-up email we send within an hour.” Jobs are replaceable. Categories are not.
• Who depends on it. Name the person. If the answer is only one person, that is a second risk sitting next to the first.
• Blast radius. Stops, limps, or shrugs. Defined below.
• Named replacement. A specific product, not “we would find something.” If you cannot name one, that is the finding.
• Data export status. Have you actually exported once, or do you assume you could?
• Switching cost in days. Honest estimate of the work, including retraining whoever uses it.

The column that produces the most value is the second one. Writing down the job rather than the tool is what reveals that two of your subscriptions overlap, that one of them is not really doing anything, and that the one you assumed was critical is actually a convenience.

Tiering by blast radius

Three tiers. Resist the urge to add a fourth.

Tier 1: the business stops. If this disappears, you cannot deliver what customers paid for, or you cannot take money. For most small businesses this is a very short list, often one or two items, and sometimes zero. These get a named replacement, a tested export, and a documented manual fallback, however ugly, that would carry you for two weeks.

Tier 2: the business limps. Output slows, quality drops, somebody works late, but revenue keeps arriving. This is where most AI tools actually sit once you are honest. These need a named replacement and a rough switching estimate. They do not need a rehearsed plan.

Tier 3: the business shrugs. Nice to have. If it vanished you would notice in a week. These need nothing except an annual look at whether you are still paying for them, which is a question worth asking regardless.

Two things reliably surprise people doing this the first time. The first is how many tools are Tier 3. The second is that the genuine Tier 1 dependency is usually not an AI tool at all. It is the payment processor, the booking system, or the single spreadsheet that runs scheduling. The AI layer is newer and louder, but it is rarely the load-bearing wall.

Getting your data out before you need it

The export button that exists in the documentation and the export button that works are not always the same button. Find out while you have time and leverage rather than during a sunset window when support is overwhelmed.

For each Tier 1 and Tier 2 tool, actually run an export this quarter and answer three questions.

1. Does it include everything, or just the obvious part? Chat tools often export conversations but not the custom instructions, saved prompts, or knowledge base that made them useful. That configuration is the part you spent months tuning.
2. Is the format usable by anything else? A ZIP of proprietary JSON is technically an export and practically a souvenir. CSV, Markdown, and plain text are portable. Anything else, check before you rely on it.
3. How long does it take? Some exports are instant. Some are queued jobs that take days, which matters enormously when the shutdown window is thirty days and everyone else is queuing too.

The strongest single habit here is keeping your prompts, system instructions, and workflow definitions in your own files, in a folder you control, and pasting them into whatever tool you are using. It feels redundant. It also means that switching vendors is a copy-paste problem rather than a rebuild.

How to buy so this hurts less

You cannot make vendor risk disappear. You can make it cheaper. Six purchasing habits do most of the work.

• Prefer tools that let you choose the model underneath. A product that can point at more than one model provider survives an upstream cutoff. A product hard-wired to a single supplier inherits that supplier’s decisions about you.
• Think hard before prepaying annually for anything under two years old. The discount is usually ten to twenty percent. You are buying that discount with the entire remaining balance if the company sunsets. Monthly is insurance and it is cheap insurance.
• Keep the workflow outside the tool. Prompts in your own documents. Templates in your own drive. Process written down somewhere a new tool can read. Store the thinking where you control it.
• Do not let one vendor hold both the data and the only interface to it. If a tool is the sole home of your customer records or your content library, that is a Tier 1 dependency no matter what it does.
• Check who is underneath before you sign. Most AI products will tell you which models they use if you ask. If they will not, that is information too.
• Read the data ownership and deletion terms once. Specifically: who owns the outputs, what happens to your data on termination, and how long you have to retrieve it. Fifteen minutes, once, before you commit.

None of this means avoiding new tools. Early products often deliver a genuine edge, and that edge is worth some risk. The point is to take the risk knowingly and to size it, which is what the map is for. If you are evaluating that category right now, our roundup of AI agent tools for small business covers what these products currently do well and where they still fall over.

The ninety-minute version

If you do nothing else from this article, do this. Block the time this week.

1. Twenty minutes. Open your card statement and your app store receipts and list every AI or automation tool you pay for. Include the free ones you rely on, because free tiers disappear first. Include the ones your contractors use on your behalf.
2. Twenty minutes. Next to each, write the job it does in one sentence, as a job and not a category.
3. Fifteen minutes. Mark each one stops, limps, or shrugs. Be strict. Most things limp.
4. Twenty minutes. For everything marked stops or limps, name a specific alternative. Ten minutes of searching per tool is enough to name a candidate.
5. Fifteen minutes. For everything marked stops, find the export function and confirm it exists and what it produces. Run one.

Put a calendar reminder six months out to redo it. The map decays quickly in this market, and a stale map is worse than none because it produces false confidence.

Where this turns into wasted effort

Continuity planning has a failure mode of its own, which is spending more on the plan than the outage would have cost.

Do not run parallel subscriptions “just in case.” Paying for two tools that do the same job is a real, recurring cost defending against a hypothetical one. Name the alternative. Do not fund it.

Do not build custom abstraction layers. Small businesses occasionally get talked into a middleware layer so they can swap AI vendors freely. That layer is now software you maintain, and it will break more often than your vendors will disappear.

Do not avoid good tools because they are young. The compounding advantage of using a genuinely better tool for two years usually exceeds the cost of one migration. Risk is a number to weigh, not a veto.

Do not make this a quarterly ritual. Twice a year is right. More often and it becomes a task nobody does properly, which is the same as not doing it while feeling like you did.

The goal is not resilience for its own sake. It is that when the email arrives, you spend an afternoon executing a decision you already made instead of a fortnight making it badly.

Frequently asked questions

How often do AI tools actually shut down?

Outright shutdowns are less common than the other four disruption modes. Acquisitions, pivots, repricing, and quiet degradation are all more frequent, and each of them can end your use of a tool just as effectively. Planning only for bankruptcy means planning for the rarest case.

Should I only use tools from large companies?

No, and it would not protect you anyway. Large companies retire products routinely, sometimes services that businesses had depended on for years, as our coverage of the Amazon Mechanical Turk shutdown laid out. Size changes the odds of a clean wind-down with decent notice. It does not change whether the product survives.

What is the single most important thing to do?

Keep your prompts, instructions, and workflow documentation in files you own rather than inside a vendor account. That one habit converts most migrations from a rebuild into a transfer, and it costs nothing.

Is annual billing ever worth the discount on an AI tool?

For an established company with a long track record and a business model that clearly works, yes. For a product from a company founded in the last two years, the discount is rarely worth prepaying twelve months of exposure. Treat the monthly premium as insurance priced at ten to twenty percent, which is reasonable for the risk being covered.

How do I know which model a tool is built on?

Check the documentation, the security or trust page, and the subprocessor list, which many companies publish for privacy compliance and which names their infrastructure providers. If none of those answer it, email support and ask directly. A straight answer is a mild positive signal. Evasion is a stronger negative one.

What if my whole business runs on one AI tool?

Then that is your single most important business risk and it deserves more than a spreadsheet row. Work out what a manual version of the service looks like, even at reduced capacity and worse margin, and write it down. You are not planning to run that way. You are making sure that if you have to, you already know how, and that you can tell customers something more useful than that you are looking into it.

Related Coverage

• SaaS Tools AI Disruption Audit: which subscriptions AI is quietly making redundant, and what to cut.
• Best AI Agent Tools for Small Business: what to weigh before a new tool becomes a dependency.
• The Amazon Mechanical Turk Shutdown: what happens when a very large company retires a service businesses were built on.

Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what is next.