In 2018 I published a post asking whether AI would eventually replace contact centre agents entirely. I called it “Is the Terminator Our Eventual Fate?” — which tells you everything about where the conversation was at the time.
It’s now 2026. I’ve spent eight years watching this space evolve, implementing AI in support organizations at two different companies, and hiring agents into an industry that looks meaningfully different from the one I described back then. This post is my honest retrospective: what I got right, what I got wrong, and what the reality of AI in support operations actually looks like from the seat of someone who’s been running these teams.
I’ve kept the core arguments from the original post and annotated them in real time. You’ll see where my thinking has held up and where it needed updating.
What I Said in 2018
“AI will change the way contact centres work. Chatbots can resolve 10–35% of customer queries without a human agent, and this number is expected to rise. Customer-facing AI will handle routine tasks and free up agents to handle more complex queries.”
What actually happened:
The deflection numbers were right directionally but the timeline was optimistic. By 2022, most contact centres I knew were running chatbots that deflected somewhere between 15–25% of volume — meaningful but not transformative. The bigger shift came with LLMs from 2023 onwards. The nature of what AI could handle changed from “answer simple FAQs” to “draft a complete response to a complex billing query for an agent to review and send.” That’s a different category of capability entirely.
The prediction about freeing agents for complex work also came true — but with a nuance nobody discussed in 2018: what “complex” means has shifted upward. Queries that were considered complex in 2018 are now handled by AI. The complex queries of today require genuine judgment, empathy, and situational awareness that still sits firmly in the human domain.
What I Underestimated
1. How quickly hiring criteria would need to change
In 2018, a good support agent needed product knowledge, typing speed, and patience. By 2024, I was hiring for something different: the ability to work alongside AI, to exercise judgment in edge cases the AI couldn’t handle, and to bring emotional intelligence to conversations the AI had already partially managed.
This shift happened faster than I expected. If you’re still hiring support agents using the same criteria you used in 2019, you’re optimising for a job that no longer exists in its original form. I wrote more about this in How to Land a CX or Support Leadership Role — specifically the section on what I now look for in interviews.
2. The implementation gap
In 2018, most writing about AI in support read like vendor marketing — here are the capabilities, here are the benefits. What nobody talked about was the implementation reality. At Q4, when we implemented AI triage as part of our routing redesign, the technology worked as advertised. What we hadn’t fully planned for was the change management: agents who felt their judgment was being bypassed, managers who didn’t trust the deflection data, and customers who escalated immediately the moment they knew they were talking to an AI.
The technology was ready before the organization was. That’s almost always how it goes. If you’re planning an AI implementation, budget as much time for the human side as you do for the technical side. Our skill-based routing redesign is a good example of what a well-executed implementation actually involves.
3. The measurement problem
In 2018 I confidently cited chatbot deflection rates as a metric. What I didn’t think through was how contested that metric would become. Does a deflection count if the customer comes back 24 hours later with the same issue? Does it count if they deflect to a different channel? Different vendors measure deflection differently, which makes benchmarking almost meaningless.
The same problem applies to AI-assisted resolution rates. The headline numbers your vendor shows you in their demo are not the numbers you’ll see in production. I’d recommend building your own measurement framework before you implement — not borrowing the vendor’s. KPIs and the Importance of Measurement covers the broader principle of why measurement frameworks need to be built from first principles rather than borrowed.
What I Got Right
1. AI would not replace agents — it would change what agents do
This was the contrarian position in 2018. Most of the breathless coverage was about automation eliminating jobs. My view was that AI would shift the composition of work, not eliminate the need for human judgment in support.
Eight years later, contact centre employment hasn’t collapsed. What’s changed is the mix: fewer agents handling simple, repetitive queries, more agents handling the complex and emotionally charged situations that AI can’t navigate well. The total headcount picture in most organizations I know has stayed roughly flat, but the skill profile required has moved significantly upward. Hiring has definitely been impacted but more in terms of less net new roles vs. mass terminations in the contact center space.
2. Speech recognition was ready
I cited Google’s 95% speech recognition accuracy in 2018 and said the technology was ready for the market. This held up. Speech recognition has become table stakes in contact centres — the interesting developments have moved to what happens after transcription: real-time agent assist, sentiment analysis during live calls, and automatic summarization of call notes.
3. AI would create a data advantage
I wrote about AI converting unstructured data into structured data for sentiment analysis and trend identification. This has been one of the most valuable applications in practice. The ability to analyze thousands of customer interactions and identify patterns — product issues, pricing friction, onboarding failures — and feed those signals upstream to product and commercial teams is genuinely transformative. A Voice of the Customer program that uses AI-assisted analysis is a completely different beast from one built on manual survey sampling.
What AI in Support Looks Like in 2026
The framing has shifted. In 2018, the question was “will AI replace agents?” By 2024, the question became “which tasks should AI own and which should humans own?” In 2026, the leading organizations are asking a third question: “how do we design a support operation where AI and humans each do what they’re genuinely best at?”
Here’s how I think about the division now:
AI handles well:
- First contact triage and routing — categorizing incoming contacts and directing them to the right queue or resource
- FAQ deflection and knowledge base surfacing — answering clearly defined questions with established answers
- Real-time agent assist — surfacing relevant knowledge articles, macros, and next-best-action suggestions during live interactions
- Post-interaction summarization — generating call notes, tagging contacts, updating records
- Quality monitoring — flagging interactions for human review based on sentiment, compliance keywords, or resolution signals
Humans still own:
- Complex escalations with ambiguous resolution paths
- High-emotion interactions — complaints, cancellations, sensitive account situations
- Judgment calls in edge cases not covered by training data
- Relationship-based interactions with key accounts or VIP customers
- Anything where the consequences of a wrong answer are significant
The line between these two columns is moving — slowly but consistently — in AI’s direction. The support leaders who will navigate this well are the ones who track that movement deliberately rather than reacting to it after the fact.
Three Things to Do Right Now If You Haven’t Already
1. Audit what AI could handle in your current queue
Pull a sample of 200 tickets from the last 30 days. Tag each one: could this have been fully resolved by AI with no human involvement? Partially assisted? Requires human judgment? You’ll probably find the distribution sits around 30–40% fully automatable, 30–40% assistable, and 20–40% human-only. That audit tells you where to focus your implementation investment.
2. Update your hiring criteria
If you’re still hiring primarily for product knowledge and typing speed, you’re optimising for a job that AI is progressively taking over. The agents who will thrive in an AI-augmented support environment are the ones with strong judgment, emotional intelligence, and the ability to handle ambiguity. Start hiring for those things now — before the market for that profile gets competitive.
3. Build your measurement framework before you implement
Agree internally on how you’ll define success before you go live with any AI capability. What counts as a deflection? How do you measure AI-assisted resolution versus AI-handoff to human? What’s your baseline? Having clean definitions before implementation is the difference between knowing whether your AI is working and just hoping it is.
The Terminator Question, Revisited
In 2018, I was asking whether AI would eventually replace all contact centre agents. Eight years on, I think that was the wrong question.
The right question is what kind of support organization you want to build — one that uses AI to drive down costs and headcount as fast as possible, or one that uses AI to raise the quality of every human interaction by removing the repetitive work that was degrading your agents’ engagement and judgment.
Both are legitimate strategies. They produce different organizations with different cultures and different customer experiences. The leaders who will do this well are the ones who make that choice deliberately rather than by default.
If you’re thinking through how AI fits into your support strategy — and specifically how it affects how you structure, staff, and lead your team — the CX & Operations hub has everything I’ve written on this topic in one place.
Hutch Morzaria is a CX and Support Leadership professional with 19 years of experience building and leading support organizations across SaaS, Fintech, and enterprise technology. He has held Director-level roles at Q4 Inc, AudienceView, Johnson Controls, and others, and holds ITIL Expert certification across V3 and V4.
