Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Created time
Jul 16, 2026 10:35 PM
Title length (<60)
Author
Last optimised
Ecomm?
Image
Publish date
Jul 18, 2026
Video
Slug
100-percent-support-automation-is-coming
Featured
Type
Article
Ready to Publish
Ready to Publish
Every vendor tells you support automation plateaus somewhere short of 100%. We think they're wrong, and the ceiling is going away. The standard advice hands AI the repetitive work and keeps humans for the rest, forever. But everything a support agent does happens on a computer. Give an AI agent the same access a human has (MCP, API, whatever the format), add a system that flags what it struggles with, and the gap closes with every run. Even the empathy objection is wobbling. In a peer-reviewed study, AI replies made people feel more heard than human ones, right up until people learned the reply came from AI.
What everyone assumes
TL;DR: Vendors and analysts agree that AI should take the routine tickets while humans keep a permanent share of the queue. Zendesk and Talkdesk both say it in public, and the caution goes back to the decision-tree chatbots everyone remembers.
Every AI support demo I've sat in on this year gives the same advice: automate the repetitive tickets, and keep humans in the loop for everything else. Zendesk's own AI guide says it outright:
"AI will not replace customer service but will transform customer interactions."
The same post reports that nearly 90% of CX leaders expect AI to resolve most customer issues in the coming years (that's Zendesk's own research, so a grain of salt). So even a vendor expecting AI to resolve most issues won't say the ceiling goes away.
Talkdesk takes the same line, calling AI a great productivity tool but not a replacement for support workers anytime soon.
And the instincts behind all this are reasonable. Most of us got burned by the decision-tree chatbots of the 2010s, CSAT is on the line, and some tickets are brand new.
Our own benchmark corpus puts the field median at around 70% AI-handled today (aggregate numbers, and directional), so the consensus gets the present right. The mistake is projecting it forward as a law.
Why that's wrong
TL;DR: Support tasks all run through a computer, so an agent with real access plus a loop that fixes whatever it fumbles will close the gap over time. On empathy, a peer-reviewed study found AI replies did better at making people feel heard until the AI label showed up.
Anything we do in support, we already do on a computer. We look up orders, check accounts, issue refunds, update subscriptions, and we write the replies. Once an AI agent has the same access a human has (MCP, API, whatever the format), there's no task-level reason it can't do the job after setup.
I made the short version of this case on LinkedIn, and the obvious pushback is: what about the tickets the AI can't handle?
Those tickets are a finite list. Build a system that flags what the AI struggles with, fix those things, and the automatable share ratchets up each run.
We productized this loop as Self-Learning. When a ticket hands over to a human, it compares the AI's reply with what your agent said and drafts a new knowledge article from the difference. Every escalation becomes training material for the next run.
On the access side, our agents call APIs through Tasks & Tools, read live backend data through the User Data API ("where is my order?", "what plan am I on?"), and come with Shopify pre-built. We built it so each action either runs on its own or proposes first and waits for approval. You choose per action, and you can widen autonomy as trust builds.
That leaves the empathy objection, the one every vendor leans on. A 2024 study in PNAS (peer-reviewed) tested it. AI-generated replies made people feel more heard than human-written ones, and the AI was better at detecting emotions.
The advantage disappeared only when people learned the reply came from AI, and that label penalty was about the same size as the gain. The study wasn't run on support tickets, but feeling heard tracked the quality of the reply, whoever typed it.
In support, people want the problem solved fast, in the moment. The "people want a human" line is about those old decision-tree chatbots, the ones that trapped you in a menu (we've all rage-typed "AGENT" into one of those). Nobody misses them.
We'll get there, for some businesses within the next couple of years. From there it comes down to each company's will to do the engineering and use the right products to close the last of the gap.
Where this doesn't hold
TL;DR: Novel problems with no precedent, work that leaves the computer, and legally required human sign-off still need a person. That boundary is real today, and it's drawn in pencil.
Today's AI still struggles with problems nobody has seen before: the outage with no precedent, the edge case no document covers. Those need a person. For now, that's the boundary, though it's today's boundary rather than a permanent one.
The argument also assumes support work stays on a computer (ours does; yours might not). If your tickets end with a van, a warehouse, or an engineer on site, the automatable share is smaller. Same for regulated flows where the law requires a human sign-off.
I'd treat 100% as a direction of travel. With the field median around 70% today, there's distance left to cover.
What this means for you
TL;DR: Buy on trajectory. Ask a vendor how escalations turn into new knowledge, which actions the agent can take beyond replying, and whether customers' resolution rates keep climbing quarter over quarter.
Instead of buying for a ceiling, buy for a slope. A vendor promising a permanent 30% human remainder is telling you where their product stops.
Three things I'd ask this week:
- What happens to the tickets the AI can't handle? You want every escalation turned into new knowledge, the way our Self-Learning feature drafts new knowledge from handed-over tickets. An escalation dead-end means your rate stalls where it starts.
- What can the agent do, beyond answering? Ask about the action surface: API calls, backend lookups, order changes (ours run through Tasks & Tools and the User Data API). Answers-only agents cap out early.
- Is your resolution rate climbing each quarter? Track it against the field median (around 70% right now) and expect your number to move. Don't let "you'll never hit 100%" excuse stalling at 40%.
Echo covers the audit side on our platform. Your team can ask it why the agent gave any answer and which knowledge source it used, then watch handover rates fall in Insights as the loop closes.
FAQs
How do I automate customer support without losing the personal touch?
Make the replies better, because the personal touch lives in the reply itself. In the 2024 PNAS study, AI-generated replies made people feel more heard than human ones, and the quality of the reply did that work. Set tone and terminology rules through Guidance so the agent sounds like your team, and give customers an instant route to a human whenever they ask for one.
What resolution rate should I expect from AI customer support?
The field median is around 70%, with half of the deployments we track landing between 56% and 80%. Treat those numbers as directional. Vendors label their metrics in different ways, and the "resolution" and "automation" medians sit about 12 points apart.
Our own customer base runs at 72% on a rolling 30-day average. Wherever you start, expect the number to climb quarter over quarter.
How to automate L1 support and only escalate complex issues
Start with the known-answer L1 tickets, then review what escalates. Each handed-over ticket shows you what to fix next. Our Self-Learning feature drafts the missing knowledge, and Tasks & Tools graduate categories like refunds and account changes into automatable ones.
Set your Handover and Escalation rules in Guidance to define what still routes to a human, and expect that list to shrink every quarter.
