The phrase "AI-powered customer service" used to mean a chatbot that deflected tickets to human agents. That's changed. By late 2026, the standard expectation is full resolution: an AI system that can reason through a multi-step request, access backend databases, process a refund, and close the ticket without human involvement.
The market calls these "agentic AI" platforms. They're not chatbots with better scripts—they're digital workers designed to handle complete workflows. And with the call centre AI market projected to hit $30 billion by 2035, every vendor is racing to position themselves in this space.
The problem? Most evaluation frameworks haven't caught up. Operations directors are still asking questions suited to 2023-era chatbots, not 2026-era autonomous agents.
What Actually Matters Now
Forget containment rate. That metric—what percentage of interactions the bot handles without escalating—made sense when the goal was deflection. It doesn't capture whether the customer's issue was actually resolved.
The metric that matters is automated resolution rate: what percentage of customer issues does the AI completely resolve, verified and closed, without any human touch? Leading platforms now claim 80%+ resolution rates. Aissist.io reports 83% automated resolution with a 4.8/5.0 customer satisfaction score. Those numbers would have seemed implausible two years ago.
When evaluating any platform, ask for their resolution rate methodology. How do they define "resolved"? Is it self-reported by the AI, validated by customer feedback, or verified by a human QA process? The answers reveal a lot about vendor maturity.
The Pricing Shift
Traditional contact centre software charges per seat or per interaction. The new model is outcome-based: you pay per resolved issue.
Typical pricing sits between £0.20 and £0.60 per resolution, varying by channel (voice costs more than chat). This sounds attractive—you only pay when the AI actually works—but the maths requires scrutiny.
Calculate your current cost-per-resolution including agent salaries, training, platform fees, and overhead. For most UK contact centres, that lands somewhere between £3 and £8 per resolved interaction. Even at £0.60 per AI resolution, the economics look compelling. But watch for hidden costs: setup fees, integration charges, minimum commitments, and what happens when the AI escalates to a human (does that count as a resolution or not?).
Integration Complexity
Here's where evaluations often go wrong. The demo looks brilliant—the AI chats smoothly, accesses customer records, processes changes. Then you discover it took six weeks of custom integration work to connect your twenty-year-old CRM.
The honest questions to ask:
- What's your average deployment time for companies similar to ours?
- Which systems can you connect to out-of-the-box versus requiring custom API work?
- What happens when our backend systems have downtime?
- Can we see a reference customer with a similar tech stack?
Enterprise platforms like Decagon have dedicated teams to map complex corporate environments and connect legacy databases without modern APIs. Mid-market platforms like Aissist.io position themselves as "operational layers" over existing helpdesks, claiming ten-minute deployments. The right choice depends on your infrastructure reality, not your aspirations.
Multi-Agent Architecture
The interesting technical development in 2026 is coordinated multi-agent systems. Rather than one monolithic bot trying to do everything, these platforms deploy specialised agents that hand off to each other.
One agent gathers context. Another checks policy. A third executes the action. A fourth handles quality assurance. Think of it like a well-organised team where each member has a defined role.
This matters because specialist agents can be optimised independently. If your refund resolution rate drops, you tune the refund agent without touching everything else. The architecture also handles complex cases better—an issue requiring both a billing check and a technical lookup gets routed through both specialist agents rather than relying on one general-purpose bot to know everything.
Ask vendors to explain their architecture. Are they running a single model end-to-end, or coordinating multiple specialised components? Neither approach is inherently better, but you should understand what you're buying.
Compliance and Data Handling
Regulated industries face additional scrutiny. If you're in financial services, healthcare, or handle payment card data, the AI platform becomes part of your compliance scope.
The minimum certifications to look for: SOC 2 Type II and ISO 27001. For healthcare, HIPAA compliance is non-negotiable. For payments, PCI-DSS matters.
Beyond certifications, understand the data flow. Does customer data leave your environment? Does the AI vendor use your conversations to train their models? Where are the servers located—relevant for GDPR if you're serving EU customers? Can the system automatically redact sensitive information before it reaches the AI model?
Fini, which focuses on regulated industries, advertises "always-on redaction" that strips personal and financial information before processing. That's the kind of specific, technical answer you want—not vague assurances about taking security seriously.
The Human Handoff
No AI platform resolves 100% of issues. The 15-20% that require human intervention need graceful handoffs, and the quality of that handoff dramatically affects customer experience.
A good handoff includes:
- Complete conversation history so the customer doesn't repeat themselves
- AI-generated summary of the issue and attempted resolution steps
- Suggested next actions for the human agent
- Sentiment analysis flagging if the customer is already frustrated
A bad handoff dumps the customer into a queue with a ticket number and no context.
Test the handoff flow specifically during your evaluation. Have someone pose as a difficult edge case and see what the human agent receives when the AI escalates.
Making the Decision
The agentic AI market is genuinely transformative, but it's also genuinely fragmented. Vendors target different segments—Zendesk AI and Intercom's Fin work best if you're already in their ecosystems; Sierra and Decagon target large enterprises with custom needs; Aissist.io and others aim at the mid-market with faster deployment and simpler pricing.
Start with your actual requirements:
- What percentage of your current interactions are candidates for full automation?
- What systems does the AI need to access?
- What's your compliance environment?
- What's your realistic timeline and budget for deployment?
Then shortlist two or three platforms that match those requirements and run genuine pilots—not demos, pilots with real customers and real edge cases.
The platforms that survive this scrutiny will deliver genuine value. The ones that don't weren't ready for production anyway.
At Hostcomm, we help contact centres evaluate and implement AI solutions that match their operational reality. Our CXCortex platform provides the analytics foundation for measuring AI performance, while Persona handles voice interactions for organisations ready to automate phone-based customer service. Talk to us about building an AI strategy that actually works for your environment.