--- Title: "The hardest tickets never needed a faster agent. They needed a different memory. " Url: "https://devrev.ai/blog/why-complex-support-tickets-still-stall" Published: "2026-07-20" Last Updated: "2026-07-30" Author: "Daniel Rojo" Excerpt: "Complex support cases still start from zero while AI handles the easy stuff. Let's dive deeper into this with our Head of Global Customer Outcomes." Reading Time: 5 --- # The hardest tickets never needed a faster agent. They needed a different memory. **Summary** - The first wave of AI cleared your easy tickets, and that felt like progress. But the hard cases – the ones that actually cost you – never got any easier. - Those cases still start from zero. Every time, an agent rebuilds a story that four other people already knew in fragments. - Watch one play out: a single complex ticket, six open tabs, context stitched together by hand. The answer was never sitting in the helpdesk. - Meanwhile new agents take months to get good, because the knowledge that matters lives in the heads of your three most senior people. - And your dashboard says things are fine – because deflection rate looks great while quietly hiding where the real cost lives. - Here's what it all points to: the volume was never the problem. The problem is everything your team knows that your systems forget. - So the question worth sitting with is simple. **_What does your most experienced person know that your system forgets the moment they log off?_** I have spent the better part of twenty years inside support. Service desks, escalation bridges, SLA reviews at 2 a.m., the quiet dread of a major incident before anyone has the full picture. So when people tell me AI has "changed customer support," I tend to ask them what, exactly, they think has changed. Because from where I sit, most of what's on the market hasn't touched the part that actually hurts. Here's the thing we don't say out loud enough. For most support organisations, the volume was never really the problem. The problem was that every hard case started from zero. ## **The part AI was supposed to fix – and mostly didn't** We all bought the first wave. Deflect the easy questions, auto-answer the FAQs, take the repetitive load off the queue. And to be fair, that worked. The simple tickets got simpler. But the simple tickets were never what kept your best people up at night. It was the L3 case that touched billing, and identity, and a bug that engineering hadn't documented yet. The one where the agent opens six tabs, copies context from one system into another, and pieces together a story that four other people already knew in fragments. The first generation of support AI couldn't help there, because it could only see the helpdesk. It answered from past tickets. It had no idea _why_ the thing actually broke – that truth lived in a Jira issue, a Slack thread, a telemetry spike, a release note. So on the cases that mattered most, the AI went quiet, and the human went back to swivel-chairing. > [!INFO] > **That's the roadblock. Not "we need faster answers." We need answers that are _connected to reality_.** ## **What the teams getting ahead actually did differently** The organisations I watch pulling ahead didn't buy a better chatbot. They changed the substrate underneath the work. Three moves keep showing up. **One – they stopped letting knowledge die at resolution.** Every solved case used to vanish into a closed ticket. Now, the moment something is resolved, it becomes something the next person – or the next automated answer – can actually reuse. The knowledge base stops rotting because it repairs itself from real resolutions, not from someone's good intentions to "document it later." **Two – they connected the answer to the whole company, not just the support stack.** The context an agent needs was never all in the helpdesk. So they gave their people one place where the ticket, the customer's real history, the engineering status, and the product signal all sit together – assembled _before_ the agent starts, not hunted for after. **Three – they kept the human firmly in the loop, on purpose.** This is the part I care about most, and the part that gets lost in the hype. The teams doing this well are not trying to remove their agents. They draft, they suggest, they assemble – and a person still approves what goes out. Trust is earned case by case, not demanded on day one. ## **What came out the other side** I'll be careful here, because I've sat through enough vendor decks to distrust a tidy number. But the direction of travel is consistent, and it's real. Complex cases that used to take the best part of an hour start resolving in minutes, because nobody is reassembling context by hand. New agents stop needing months to become useful, because the institutional memory no longer lives only in the heads of your three most senior people. And those senior people – the ones you can't afford to burn out – get their attention back for the genuinely hard, genuinely human conversations. Notice what none of that is. It isn't "we replaced the team." It's "the team finally spent its time where the team is irreplaceable." That's a very different outcome, and it's the one worth chasing. > [!INFO] > ## **The lessons, if you're deciding where to start** > > If I were advising a peer sitting where I've sat, I'd offer four things I wish I'd internalised sooner. > > - **Stop measuring the wrong win.** Deflection rate feels good and tells you almost nothing about your hardest cases. Look at your L3/L4 handle time and your ramp time for new agents. That's where the cost – and the suffering – actually is. > - **Follow the tabs.** Sit with an agent on a complex case and count how many systems they open to close one ticket. That number _is_ your problem, drawn in plain sight. Any AI that can't see across those systems is answering a different question than the one you're asking. > - **Fix the memory before you chase the automation.** Automation on top of scattered, stale knowledge just makes wrong answers faster. Get your knowledge connected and self-maintaining first; the automation becomes almost boring after that. > - **Earn trust, don't mandate it.** Start with the AI drafting and the human approving. Adoption follows usefulness, never the other way around. If your agents don't trust it on the small cases, they'll never let it near the big ones. ## **Where this is really going** The future I actually believe in isn't "AI answers your tickets." It's the end of ticket-passing. No more fragmented ownership, no more handing the customer between silos, no more starting every hard problem from a blank page. Support and success stop being a relay race and start being one continuous, informed relationship. We're not all the way there. But the gap between the teams who've started and the teams still buying faster chatbots is widening – and it compounds. The knowledge you connect this quarter is the ramp time you save next quarter, and the burnout you avoid the quarter after. **So the question I'd leave you with isn't "which AI should we buy." It's simpler, and harder: _what does your most experienced person know that your system forgets the moment they log off?_** Answer that honestly, and you'll know exactly where to start.