Customer service was the first function where AI got deployed widely, and it is also where the reputational damage from getting it wrong is most immediate. Every business leader has been on the customer side of a bad deflection experience, looping through options that do not apply while looking for a way to reach a person.
That experience is not an argument against AI for customer service. It is an argument against measuring the wrong thing. Containment tells you whether the conversation ended. It says nothing about whether the customer’s problem ended. Those two numbers move in the same direction when a system works and in opposite directions when it does not, and most dashboards only report the first one.
At iSpark we work with service leaders who need the cost benefit without the customer backlash. This article covers where AI powered customer support genuinely works, the single constraint that determines your ceiling, a documented example from a US bank with nearly a decade of public data behind it, the mistakes that damage trust, and how to start.
The Knowledge Base Is the Ceiling on Everything
This is the least discussed and most important point. An AI support system answers from your documentation. If your knowledge base is incomplete, contradictory or two years out of date, no model will fix that. It will simply deliver your bad documentation faster and with more confidence than a human agent would have.
The implication is uncomfortable but useful. Most organisations that want AI customer service automation should spend the first phase auditing and repairing their knowledge base, not selecting a platform. Our work on AI for customer service teams treats that audit as the project’s foundation, because the alternative is discovering the gap through customer complaints.
Where AI Customer Service Tools Deliver
| Application | What it does well | What to watch |
|---|---|---|
| Routine query handling | Balances, status, policy and account questions instantly | Repeat contact rate, not just containment |
| Proactive outreach | Flagging an issue before the customer notices it | Frequency fatigue and relevance |
| Agent assist | Surfacing context and drafting replies for human agents | Agents pasting drafts without reading them |
| Triage and routing | Getting the case to the right team first time | Quality of your case taxonomy |
| Post contact summarisation | Writing up cases accurately and consistently | Summaries entering a record unverified |
Agent assist is consistently underrated. It carries almost none of the customer facing risk, it improves consistency across shifts, and it shortens handling time on exactly the complex cases that deflection cannot touch. It is usually the better first project.
A Real Example: Bank of America and Erica
The challenge. Bank of America serves tens of millions of clients, most of whom need routine information rather than advice: a balance, a transaction, a card control, a payment date. Handling that volume through branches and call centres is expensive, and customers dislike waiting for something simple.
The solution and implementation. The bank launched Erica in 2018 as a virtual assistant inside its mobile app, then expanded it steadily rather than all at once. Capability grew from navigation and basic queries to proactive personalised insights, and later into Merrill, Benefits OnLine and the CashPro platform for corporate clients. The bank also deployed an internal version, Erica for Employees, which around 97 percent of employees have adopted.
The outcome. Erica has surpassed 3.6 billion client interactions since launch, helping more than 24.6 million clients, including 1.9 billion proactive personalised insights. Clients now interact with it more than 200 million times per quarter. The bank has reported that around 98 percent of clients find the information they need, with an average interaction lasting roughly 48 seconds.
The business impact. Routine volume moved away from expensive channels while engagement went up rather than down. The detail worth copying is the proactive shift: somewhere between half and sixty percent of Erica interactions now begin with the assistant making a suggestion rather than the customer asking a question. That is a different model from deflection, and it is why the engagement numbers keep rising.
Common Mistakes in AI Driven Customer Support
- Reporting containment without reporting repeat contact rate alongside it.
- Deploying on top of documentation that has not been audited.
- Hiding the route to a human, which reliably produces complaints and bad press.
- Sizing the team on deflection rates while ignoring that remaining cases are harder.
- Letting AI written case summaries enter the record without agent verification.
- Launching customer facing before trying agent assist, which carries far less risk.
Best Practices Checklist
- Audit and repair the knowledge base before selecting any platform.
- Report containment and repeat contact together, always.
- Make escalation to a person obvious and fast.
- Start with agent assist, then move to customer facing.
- Track satisfaction by resolution type, not as one blended score.
- Review the questions the system failed on monthly and feed them back into documentation.
How to Get Started
- Pull your top fifty contact reasons by volume and check the documentation for each.
- Fix the gaps. This is the project, not the preamble.
- Deploy agent assist first and measure handling time and consistency.
- Move the highest volume, lowest risk contact reasons to customer facing handling.
- Expand only when repeat contact rate stays flat or falls.
Future Trends in AI Customer Experience
Three shifts are underway. Service is moving from reactive to proactive, as the Erica numbers show, which changes the economics because a prevented contact costs nothing. Agents are gaining the ability to take action rather than just answer, and Microsoft’s work on autonomous agents for business process illustrates where the tooling is heading. And voice is becoming credible again, which matters because phone remains the channel customers choose when something has gone genuinely wrong.
Key Takeaways
- Your knowledge base sets the ceiling on everything else. Audit it first.
- Containment without repeat contact rate is a misleading number.
- Agent assist carries less risk than deflection and often delivers more.
- Bank of America’s results came from steady expansion since 2018, not a single launch.
Frequently Asked Questions
What should we measure in AI customer service?
Containment and repeat contact rate together, plus satisfaction split by resolution type. Containment alone tells you the conversation ended, not the problem.
Will customers accept an AI assistant?
Generally yes for routine requests, provided answers are accurate and escalation to a person is obvious. Hiding the human route reliably produces complaints.
Where should we start?
With a knowledge base audit, then agent assist. Customer facing deflection should come after both, not before.
Does AI support automation reduce headcount?
Sometimes, though remaining cases become harder on average. Many organisations absorb volume growth instead of cutting, which is often the better outcome.
How long until we see results?
Agent assist typically shows measurable change within a quarter. Customer facing deployment takes longer because documentation repair usually dominates the timeline.
Where to Take This Next
The service organisations doing this well are not the ones with the cleverest assistant. They fixed their documentation first, kept the route to a person obvious, measured whether problems ended rather than whether conversations did, and built toward proactive contact rather than pure deflection. That sequence protects the customer relationship while the costs come down.
If you want an independent view of whether your documentation and contact data would support this, iSpark runs fixed scope service assessments that end with a written recommendation either way. Start with your top fifty contact reasons.
Published by iSpark.
