Zendesk removes legacy AI agents on 10 December. Migrate in place, unless you were already unhappy.

Our verdict: migrate in place. Take the upgrade to Zendesk's new AI agent experience, and use the deadline to delete rather than to replatform. Leave Zendesk only if you had already decided the deflection was not working before this announcement existed.

That is a less exciting recommendation than "here are five Zendesk alternatives," and we make it deliberately. Here is the reasoning, because a verdict without one is just an opinion:

• A vendor deadline is a terrible trigger for a vendor decision. You have 101 days, and a helpdesk replatform touches macros, views, SLAs, routing rules, reporting history, and every integration you have wired into tickets. Deciding to move your whole support stack under time pressure, with the removal date as the forcing function, is how teams end up mid-migration in December with two systems live and a queue nobody is watching.
• The upgrade is genuinely an upgrade, not a lateral move. Zendesk's own announcement says the new AI agent experience includes agentic capabilities that were previously available only through the AI agents Advanced add-on. Teams on the Essential tier are being pushed onto a more capable product. That is an unusual shape for a forced migration and it is worth taking at face value.
• Most of what is being removed should not be rebuilt anywhere. This is the part nobody says out loud. Intent-and-answer bots built in 2023 and 2024 tend to be a graveyard of half-finished flows. The honest output of this project is usually a much shorter list of automations than you started with, which is a good outcome that a replatform would have hidden behind migration busywork.

The one thing we will not tell you to do is nothing. Maintenance mode started on 31 August 2026, which means Zendesk has stopped developing this functionality and is fixing only critical bugs. A bot you leave alone is now sitting on frozen code with a hard stop date attached.

The migration is not the hard part. Knowing what is load bearing is the hard part, and it is the same failure mode we see in every dated migration: the team rebuilds the three things it remembers and gets ambushed by the long tail in week six. Capture these columns for every legacy bot, answer, and intent, including the ones you are certain are dead:

1. What it does, in one plain sentence. If nobody can write that sentence, that is your answer about whether to rebuild it.
2. Which channel it fires on. Web widget, messaging, email, or a social channel. Channel is what determines who notices when it stops.
3. Volume over the last 90 days. The single most useful column. A bot that has handled a trivial number of conversations in a quarter is a delete, not a migrate.
4. Resolution or deflection outcome. Did the conversation end, or did it become a ticket anyway? A bot that deflects nothing is costing you customer patience for no benefit.
5. What runs downstream. Triggers, automations, routing rules, tags, and any integration that keys off what the bot sets. This is where the December surprises live.
6. A named business owner. Someone who would notice and care if it stopped. No owner is a strong signal for the decision column.
7. Keep, kill, or merge. Filled in last, once the whole list is visible.

Fill in that last column only after the table is complete. Every dated migration we have worked through produces the same surprise once the list is in one place: a meaningful share of the automations get killed rather than rebuilt, and several near-identical intents collapse into one better-scoped agent. That is the real return on this project.

The same discipline applies whether you stay or go. If your help centre content is thin, no AI agent on any platform will deflect well, because they are all answering from your articles. Fixing the source of truth before the tool is the argument we make in Foundation, and it is more relevant here than the choice of vendor.

• Intents do not survive as intents. The legacy model was: match an utterance to an intent, then serve the mapped answer. The new experience is agentic, which means it reasons over your content instead of matching to a list. Porting a hundred intents across one by one carries a 2023 design into a 2026 tool. Rebuild the outcomes, not the intent list.
• Answers were quietly load bearing for routing. Many teams used answers to set a tag or a field, and then built triggers on top of that. When the answer goes, the tag stops being set, and the trigger silently stops firing. Nothing errors. Tickets just start landing in the wrong queue.
• Autoreplies with articles are invisible until they are gone. Nobody monitors a feature that works. Check whether it is on, check what it deflects, and decide deliberately rather than discovering the answer from your December ticket volume.
• Your help centre content is the real dependency. Every route in the table answers from your articles. If they are stale, contradictory, or missing, the new agent will be confidently wrong in a way the old keyword bot was not. Auditing content is unglamorous and it is the highest-value work in this project.
• Two bots live at once means two replies. While you test the new agent alongside the legacy one, guard against both answering the same conversation. Point the new one at a test channel or an internal queue until you trust it. This is the single most common way a support migration embarrasses you in front of a customer.

The pattern behind all five is the same: the visible thing being removed is a bot, but the thing that actually breaks is something downstream that nobody documented. That is why the inventory column asking "what runs downstream" matters more than the one asking what the bot says.