Skip to main content
Guide9 min read·Updated July 17, 2026
🤖

How to Tell If an AI Tool Is Dead or Dying (2026)

B

A. Frans

Published July 17, 2026

ai toolsvendor due diligencetool evaluationdata portabilityprocurement

The vendor always tells you, just never loudly

While checking facts for this piece I went looking for proof that Casetext had shut down. The first page of results was wall-to-wall posts from companies selling Casetext alternatives. Paxton, NexLaw, StrongSuit, LegesGPT. Every snippet led with the same date and the same call to action: switch to us.

A dozen competitors repeating a date isn't a source.

The proof was sitting on Thomson Reuters' own support site, in a page called "CoCounsel 2.0 Migration Information." Verbatim: "Beginning on February 1, 2025, the Casetext.com website will redirect to a new CoCounsel page." Further down, customers keep access to the old platform "through March 31, 2025." Thomson Reuters had bought Casetext for $650 million in cash, announced August 17, 2023. Roughly nineteen months later, the wind-down schedule went up in a help doc that nobody outside the existing customer base would ever think to open.

That gap is where you get hurt. The vendor posts the truth in a support article. The directories keep the old pricing card up. You pick the tool off the card.

So here's the check I run before committing to anything. It takes about ten minutes and it has saved me from two migrations already this year.

The signals, and where they live

None of these is conclusive alone. Three of them together is a pattern.

SignalWhere to checkWhat it means
Changelog gone quiet/changelog, /releases, /whats-new90+ days of silence on a product that shipped monthly is the loudest signal there is
Last commitGitHub repo, default branchDependency bumps only, no feature work = maintenance mode
Package publish datenpm, PyPI, crates.ioThe SDK is the tell. Nobody sunsets the SDK and keeps the product
Docs last-updatedFooter or page metadataDocs describing features the UI no longer has means the writers left first
Status pagestatus.<domain>An incident from March still open in July means nobody's on call
Pricing pageVendor's own /pricingPublic tiers replaced by "contact sales" often precedes enterprise-only retreat or a sale
Support channelDiscord, community forum, XFounder used to answer in a day, now it's a bot and a backlog
Acquisition newsVendor newsroom, acquirer's newsroomStart the clock. Most absorbed products get 12 to 24 months
Domain behaviourType the URL, watch the barA redirect to the acquirer's marketing page means the product is already gone
Job postingsCareers page, LinkedInRoles pulled and never reposted. Hiring is the cheapest honest signal a company emits
The careers page deserves a note. Companies lie in blog posts and they hedge in support docs, but they don't spend money hiring for a product they've decided to kill. When the engineering reqs for a product vanish and don't come back, that decision has already been made upstairs.

Dead, dying, absorbed: pick the right word

These three get flattened into "shut down" and they call for completely different responses.

Dead means the endpoints return errors and your data is gone or going. There's no negotiation. You needed to be out yesterday.

Dying is the dangerous one, because dying tools still work. The login page loads, the API answers, the invoice clears. What's stopped is the future: no new features, no security patches after a while, no roadmap. A dying tool will keep serving you competently for another year and then hand you an emergency migration on someone else's schedule. If your security posture depends on the vendor patching things, a dying tool is a live risk, which is the same reasoning behind the vendor-viability checks in our guide to AI tools for cybersecurity professionals.

Absorbed is a different animal, and the industry keeps miscategorising it. LexisNexis announced general availability of Lexis+ with Protégé on February 24, 2026, and said in its own press release that the platform "replaces Lexis+ AI." Replaced, not killed. By May 7, 2026 they'd shipped another expansion onto it with agentic drafting and larger document vaults. That product line is unambiguously alive. The name on the door changed.

Absorbed still costs you, though. Casetext is the case study. Thomson Reuters didn't delete the technology, they folded it into CoCounsel and Westlaw. The customers who got hurt weren't hurt by deletion. They were hurt by repricing and repackaging: the cheap standalone product they'd built a practice around stopped existing as a thing you could buy on its own terms.

So when you see an acquisition, don't ask whether the product survives. Ask whether your tier survives.

The ten-minute check

Run this before you commit, in this order. It's ordered by signal strength per second spent.

1. Type the domain into the address bar. Watch what happens. A redirect to a parent company ends the evaluation right there. (30 seconds) 2. Open the changelog and read the top entry's date. Under 60 days, healthy. 60 to 120, watch it. Over 120 days on a product that used to ship monthly, treat as dying until proven otherwise. (1 minute) 3. Search the vendor's own site for "migration", "sunset", "end of life", "deprecat". This is the step that found the Casetext answer while the whole first page of Google was selling me something. Support docs are written by people who have to be accurate. (2 minutes) 4. Check the SDK on npm or PyPI. Last publish date, download trend. (1 minute) 5. Check the careers page for engineering roles on that product. (1 minute) 6. Check the newsroom of the acquirer if there's been an acquisition, not the acquired company's blog. The acquired blog goes quiet; the acquirer keeps publishing. (2 minutes) 7. Open the pricing page and screenshot it. You'll want the receipt when it changes. (30 seconds) 8. Find the export button before you find the import button. More on this next. (2 minutes)

Steps 1 through 6 are the diagnosis; 7 and 8 you do regardless, because you'll want the pricing receipt and the export route whether the tool is healthy or not. If the first six come back clean, adopt. If two or more come back ugly, either walk away or adopt with a documented exit plan and a calendar reminder to re-check in 90 days.

Check the export path while you still don't need it

This is the part people skip, and it's the part that costs money.

Go back to that Thomson Reuters migration page, because it says something remarkable in plain corporate English: "we have made the decision not to migrate chats or documents loaded via chat into the new platform." Databases could migrate, in two waves, starting February 1 and March 1, 2025, taking up to two weeks. Chats and chat-loaded documents did not. Users had until March 31, 2025 to go get them.

A $650 million acquisition by one of the largest information companies on earth, a fully funded migration project, professional program management, and the work product still didn't come across. Not because anyone was careless. Because migrating conversational history was judged not worth it.

Your vendor will make the same judgement. Every AI tool you use is accumulating exactly this kind of asset: prompt libraries, chat threads, fine-tunes, evaluation sets, annotations, the accumulated context that makes the tool useful to you specifically. Ask the boring question on day one. Can I get all of it out, in a format something else can read, without opening a support ticket?

Test the export on day one, not on the day you need it. An export button that produces a JSON blob nothing else can ingest is decoration. If the answer is "contact support for a data request," you don't have an export path, you have a hope. Where this gets formal is procurement: the enterprise security and compliance tooling worth buying treats data portability as a contract term, not a feature request.

My rule now: no tool holds data I can't retrieve unattended in under an hour.

About directories, including this one

We list AI tools here. Some of those listings are stale right now, as you read this. I'd bet money on it.

That's mechanics. A directory is a snapshot of a market that moves weekly. Vendors don't email directories when they raise prices, and they definitely don't email when they're winding down. The incentive runs the other way: a dead listing still earns affiliate clicks and still ranks. Casetext's shutdown was documented on Thomson Reuters' own help pages, and listings elsewhere carried its old standalone pricing for months afterward.

So use directories for what they're good at. Discovery, category maps, finding the six candidates you didn't know existed. Then leave and go verify on the vendor's own site before you spend a dollar or an hour.

Any directory that tells you its listings are always current is telling you it doesn't understand its own product.

FAQ

How long should a changelog be quiet before I worry?

Measure against that tool's own history, not a universal number. A stable CLI that ships twice a year is fine at 150 days. An AI product that shipped every three weeks all last year and has been silent since April is telling you something. Compare the tool to its past self.

A tool I depend on just got acquired. What do I do this week?

Export everything now, while the old team still runs support and the old export still works. Then find the acquirer's product page for your tool and read how they describe it. If your tier isn't mentioned, assume it's going away. You typically have 12 to 24 months, and Casetext's customers had roughly nineteen from the August 2023 acquisition announcement to the March 31, 2025 access cutoff.

Is open source safer?

Safer against disappearance, not against abandonment. Nobody can switch off a repo you've cloned, and that's a real guarantee. But an unmaintained model wrapper with a dead dependency tree and no security patches will hurt you on a slower timeline. Check commit recency and whether more than one person has merge rights.

The pricing page went to "contact us." Is that fatal?

No, but log it. It usually means one of three things: they're moving upmarket, they're testing what enterprises will pay, or they're dressing up for an acquirer. Only the last is your problem, and it rarely arrives alone. Pair it with the careers page. Pricing going private while engineering reqs disappear is a much darker combination than either signal by itself.

Viability deserves half the evaluation

Most tool evaluations are 90% features and 10% viability, and the ratio should be closer to even. Feature gaps are visible on day one and you can plan around them. Viability failures arrive on the vendor's schedule, during a quarter you'd planned to spend on something else.

Ten minutes on the vendor's own domain beats an afternoon of reading reviews written by their competitors.

Share this article

📬

Get More AI Tool Guides

New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.