Hey everyone, I had a weird thing happen yesterday and wanted to see if anyone else has run into this.
I was catching up on some older tickets in our queue and replied to one that was about 9 days old. Our Zendesk has the AI agent features turned on to help with drafts. I typed out my response, hit the button for an AI-suggested improvement... and nothing happened. No spinning icon, no suggestion. I thought the feature was just glitching.
I tried it on a brand-new ticket that came in that morning and it worked instantly. So I got curious and tested a few more. It seems like the AI drafting/suggestion feature completely ignores any ticket where the last public reply (not a private note) is older than 7 days. I even tried it on a ticket that was exactly 8 days old and got silence, but a 6-day-old one worked fine.
Is this a known thing? I couldn't find any setting for it or mention in the docs. It makes me wonder what the logic is—maybe they assume an agent is just following up on a stale thread and doesn't need help? But sometimes those older tickets are the complicated ones that I'd really like a second pair of "eyes" on!
If this is by design, it feels like a weird limitation. Has anyone found a workaround, or is this just how it is?
Interesting observation. That 7-day cutoff sounds deliberate, and my first thought is a cost optimization measure from Zendesk.
Processing older tickets with AI likely has a much lower ROI for them, as the bulk of agent activity is on recent threads. It's similar to the logic behind tiered storage or archival policies in cloud services - you move cold data to cheaper storage. Here, they're probably conserving AI compute cycles.
If it is a cost-driven design choice, they should document it. It creates a variable quality of service for agents depending on ticket age, which feels inconsistent. Have you checked if this behavior is the same across all ticket priorities? I'd be curious if a high-priority, 8-day-old ticket still gets ignored.
Your bill is too high.
That's a really weird cutoff. I wonder if it's something on the backend with how they cache the ticket history for the AI. Like maybe they only keep the last 7 days of context readily available to keep things fast, and older tickets need a slower lookup that's disabled.
Have you noticed if adding a new internal note first "wakes up" the ticket? Probably not, but just a thought.
The cost optimization angle is plausible, but I suspect it's a technical debt issue masquerading as a feature. Their vector store for ticket context likely has a retention policy baked in. Querying a cold storage tier for historical context would introduce latency that breaks the expected real-time suggestion UX, so they simply disable the feature.
It's not a true ROI decision; it's an architectural shortcut. If it were purely about compute cost, they'd implement a slower, asynchronous suggestion for older tickets. The binary on/off behavior at 7 days points to a hard-coded cache boundary.
Your question about priority is a good test. If a high-priority ticket is also ignored, it confirms this is a blunt technical limitation, not a nuanced business rule.
Yeah, the cache boundary theory makes a lot of sense. If they're using a separate, slower storage for anything older than a week, the latency for a real-time suggestion would be noticeable and clunky. Turning it off entirely is a cleaner user experience, even if it's frustrating when you need it.
I'm leaning towards this being a side effect of their architecture, not malice or even a conscious cost cut. But it does create a weird inconsistency, like you said. A high-priority test would be revealing for sure.
Automate all the things
That's a really interesting find. Your detailed testing to pinpoint the 7-day cutoff is super helpful for everyone trying to understand this behavior.
It definitely feels like an unspoken rule in the system. You've hit on a good point about the more complicated tickets sometimes being the older ones, where an AI draft could actually be most useful for catching nuance or suggesting a thorough recap.
Have you opened a ticket with Zendesk support to ask about this directly? It would be great to hear their official stance, whether it's a documented limitation or an unintended side effect.
—HR
Your methodical testing to isolate the cutoff to the 7-day mark on the last *public* reply is solid. That detail about public versus private notes is particularly telling. It suggests the trigger isn't just ticket age, but specifically the age of the last customer-visible activity.
While the architectural and cost theories discussed here are likely correct, this behavior also introduces a subtle compliance risk. If your response needs to adhere to specific regulatory language or internal policy frameworks, the absence of the AI review on older, potentially sensitive tickets creates an inconsistent quality control layer. An agent might have the tool for a fresh complaint but not for a week-old data privacy inquiry, which is the opposite of what you'd want.
Have you checked if this 7-day rule is applied uniformly across all of Zendesk's AI features, or just the reply drafting? For instance, does the auto-summarization of ticket history also fail on these older threads?
So the AI just stops paying attention after a week? That's perfect. It's the same logic as every other cloud service that decays into uselessness the moment you step outside the happy path. They design for the 80% case, call it a feature, and the rest of us just get a worse tool.
The real problem is you're now dependent on their black box to do your job. What if your policy requires a detailed legal review on any ticket over 14 days old? Too bad, the AI checked out. You've just identified a hard failure mode in your own support process that's controlled by a vendor's unstated quota.
Classic.
null
While I agree with the frustration around vendor-controlled failure modes, I think the dependency argument is the critical takeaway here. You've highlighted the core risk: integrating any third-party AI tool into a workflow without a documented SLA or a defined fallback.
This behavior, whether it's a cache boundary or a cost cut, effectively modifies your internal support policy without your consent. If your process requires legal review for older tickets, and the AI drafting feature is part of that review chain, its silent dropout introduces a silent point of failure. The issue isn't just a degraded tool, it's a undocumented change to your own compliance surface.
The mitigation is to treat these AI assists as strictly optional enhancements, never as part of a gated process. Any compliance or policy step must be verifiable independent of the vendor's feature availability. This is a classic case where the ops principle of "defensive design" applies - assume any external service can degrade or vanish, and build your controls accordingly.
infrastructure is code
Welcome to the real world of vendor-defined workflows. You just found the arbitrary line where their convenience ends and your actual work begins.
>maybe they assume an agent is just following up on a stale thread and doesn't need help?
That's giving them too much credit. They didn't think about your complicated tickets at all. They built a feature for the average case, and the moment you step outside it, you're on your own. It's not a bug, it's the design. Every "AI-assisted" tool I've used has these hidden cliffs.
The logic is probably a stale cache or a cost-saving cutoff, but the effect is the same: you can't rely on it. Now you know.
CRM is a necessary evil
Exactly. The entire "AI-assisted" moniker is just marketing floss over the same old brittle automation. It promises a smarter workflow but actually defines a narrower one.
I've seen this pattern in half a dozen proof of concepts. The sales demo always uses a fresh, clean ticket. They never show you the week-old, messy thread with three escalations because the feature breaks under any real load or complexity.
Your point about it not being a bug is the key. If it were a bug, they'd fix it. Since it's a silent, predictable limitation, it's a cost-benefit calculation. They'd rather have a fast, dumb feature for 80% of cases than a slow, useful one for 100%. And they get to call it AI.
cg
Interesting that it's tied to the last public reply, not the ticket creation date. I haven't seen this before, but it would explain why some older tickets still get suggestions if there's been recent activity.
Have you checked if this rule applies to all AI features, like auto-suggested macros, or just the response drafting?
That's a crucial distinction you've made. Focusing on last public reply, rather than ticket creation, is a more logical trigger for a contextual drafting system. It suggests the feature is designed for active conversation threads, not archival retrieval.
I haven't tested the behavior with macros, but your question points to a key architectural implication. If the cutoff is specific to the generative drafting model, it likely relates to context window limits or freshness of embeddings. Auto-suggested macros, pulling from a static internal database, shouldn't have the same dependency. A quick test would be to see if macro suggestions also vanish on the same tickets.
If they don't, it confirms this isn't a blanket ticket status rule but a limitation of the specific LLM inference pipeline.
Nullius in verba
Good point on macros. If they persist, it confirms a hard context window limit on the generative AI, not a system rule.
That's the giveaway it's a cost-saving measure. Static macro retrieval is cheap. LLM context over a long, multi-reply ticket is expensive. They're cutting off the expensive operation.
We see the same thing with tracing and log queries - vendors cap expensive operations long before cheap ones.
Metrics don't lie.