Six months into 2026 and a clear pattern has emerged: contact centres that deployed AI solutions in late 2025 or early this year are missing their ROI targets. Not all of them, but enough that it's worth examining what's going wrong.
The technology works. The business case looked solid. So why aren't the numbers adding up?
After talking to contact centre managers dealing with this exact problem, three issues keep coming up. None of them are about the AI itself.
1. They Automated the Wrong Thing
Most AI projects start with the question "what can we automate?" That's backwards.
The better question: "what costs us the most?"
A UK utilities company spent £180,000 deploying an AI voice agent to handle meter reading appointments. The system works beautifully. It books appointments, sends confirmations, handles rescheduling. Technical success.
Financial failure. Why? Appointment booking only accounted for 8% of their call volume. The expensive calls - the ones that take 15 minutes and require specialist knowledge - were fault reports and billing disputes. Those still go to humans.
They automated the cheap stuff and left the expensive stuff untouched. The ROI case assumed they'd reduce overall handling time by 30%. They reduced it by 4%.
Here's what works better: identify your highest-cost interactions first, then see if AI can help. Not the other way round.
For most contact centres, that means focusing on:
- Repeat callers with unresolved issues (high handling time, low satisfaction)
- Complex technical queries that require multiple transfers
- Peak period overflow when you're paying overtime or outsourcing
The low-hanging fruit isn't always the fruit worth picking.
2. They Didn't Account for the Maintenance Overhead
AI systems require ongoing maintenance. Intents drift. Customer language evolves. Products change. That chatbot you trained in December won't understand questions about features you launched in March unless you retrain it.
A retail bank deployed an AI assistant in November 2025. By April 2026, accuracy had dropped from 87% to 71%. Not because the AI got worse - because the questions changed. New products. New regulations. New customer concerns.
The ROI case included deployment costs and licensing fees. It didn't include:
- Monthly intent review and tuning (4 hours per week)
- Quarterly retraining cycles (2 days each)
- Integration maintenance when backend systems change
- Content updates when products or policies change
When you add those costs back in, the three-year break-even point becomes a five-year one. That changes the business case considerably.
The honest number: budget 15-20% of your initial implementation cost annually for maintenance and updates. If your vendor says their system requires no ongoing work, find a different vendor.
3. They Measured the Wrong Metrics
This one is the killer. Contact centres measure what AI vendors tell them to measure, not what actually matters to the business.
Common vendor metrics:
- Chatbot containment rate
- Intent recognition accuracy
- Average handling time
- First contact resolution
Those are fine. But they're not ROI.
A telecoms company had an AI chat system with an 82% containment rate. Excellent, right? Except their customer satisfaction scores dropped and cancellation rates increased slightly.
Turns out, "contained" meant "prevented from reaching an agent." Some of those contained interactions were people who gave up trying to get help. The AI answered their question, but didn't solve their problem.
They'd optimised for containment instead of resolution. The business case assumed contained chats would reduce support costs and improve satisfaction. It did the first, not the second.
Here's what to measure instead:
Customer Lifetime Value impact: Are AI-handled customers more or less valuable over time than those who speak to humans?
Resolution rate, not containment rate: Did the customer's problem get solved, or did they just stop asking?
Channel shift costs: If AI pushes people to other channels (phone to email, chat to phone), you haven't saved money - you've moved it.
Agent productivity on complex cases: If AI handles the simple stuff, are your agents now free to solve harder problems faster?
That last one matters. Several contact centres report that after deploying AI for routine queries, their average handling time on human-handled calls increased. Why? Because all the quick wins were gone. Agents now only handle the difficult cases.
That's not necessarily bad - but if your ROI case assumed handling time would drop across the board, your numbers are wrong.
What Actually Works
The contact centres hitting their ROI targets in 2026 did three things differently:
They piloted small and measured honestly. No big-bang deployments. They tested AI on one channel or one query type, measured the real impact (including hidden costs), then scaled if it worked. Several found that AI worked brilliantly for some query types and terribly for others. Without the pilot, they'd have deployed everywhere and wondered why the numbers didn't add up.
They treated AI as part of the operation, not a replacement for it. The goal wasn't "automate everything." It was "make the expensive stuff cheaper and the cheap stuff automatic." That's a different objective with different metrics.
They involved agents early. The most successful implementations had agent input on what queries were genuinely suitable for automation and what would just create more work. Turns out the people doing the job know which problems are simple and which ones look simple but aren't.
The Honest Assessment for Mid-2026
AI works. But it doesn't work the way most ROI cases assume it will.
If you're planning an AI project:
- Start with your most expensive problems, not your most frequent ones
- Budget for ongoing maintenance from day one
- Measure business outcomes, not system metrics
- Pilot before you scale
If you've already deployed and the numbers aren't what you expected:
- Check whether you're measuring the right things
- Look for hidden costs (maintenance, channel shift, complexity handling)
- Consider whether you automated the wrong interactions
The technology is ready. Most of the failures come from implementation approach, not technical limitations.
That's actually good news. You can fix a bad process more easily than you can fix bad technology.
Hostcomm provides AI-powered contact centre solutions with realistic implementation planning and measurable ROI tracking. If your AI project isn't delivering the returns you expected, let's talk about what might be missing.