When talking to bank and CU executives, I often get asked what fraction of customer requests the virtual assistant can handle. The question is certainly sensible but the answer can be surprising. So I decided to write a small guide that can be used to estimate the impact of virtual assistants on a particular institution. Following these steps will yield a pretty good estimate for your specific institution versus relying on claims of any vendors.


  • Step 1) collect a sample of 100-200 existing chat or customer service call center interactions. The sample should be random and not collected by asking employees to recollect the topics, because employees are likely to more quickly remember the unusual/recent interactions.

  • Step 2) review each interaction manually and extract the main topic (ie goal) of the interaction. These could be: change mailing address, account transfer, or finding out branch hours.

  • Step 3) count up the topics and group by frequency. The likely result is that 5 to 10 topics will cover 80% of the interactions, and there will be several topics appearing only once (e.g. dealing with failed wire transfer) – the proverbial long tail.

  • Step 4) for each of the high frequency topics, decide whether the topic is pertaining to a general question or account specific question. For example, branch hours or routing number are general questions, whereas stopping a payment is an account specific one. Mark the general question topics with a “Yes” next to them. The account specific topics will need further inspection.

  • Step 5) for each of account specific topics check if there is already a self-service implementation to complete the inquiry. For example, a self-service implementation for resetting a lost password. If the answer is yes, mark the topic with “Yes” next to it.

  • Step 6) for account specific topics without existing self-service options, you need to answer the following questions

    • Is there an API access point that a virtual assistant could integrate into, and would the integration cost (if any) be acceptable for the organization? For example, transferring money between accounts would require integration into your core, and submitting a loan application would require integration with a loan origination system.

    • Does regulatory compliance allow the transaction to complete without human interaction or additional verification? For example, making a wire transfer might require an additional phone call to verify recipient information.

    • Does your business strategy favor completion of the transaction without human intervention? For example, closing an account or canceling a credit card might require human interaction.

    • If the answer to all these sub questions is yes for any topic, mark a “Yes” next to it.

  • Step 7) tally up all the topics with “Yes” next to them and divide them by the total sample count (e.g. 200) to get a percentage of topics that a virtual assistant could handle. Likely the percentage will end up between 30% and 80%.


This worksheet will give you a good estimate of the percentage of interactions a virtual assistant could handle. One more question you need to consider is what is the overall experience you want for the users – namely will the virtual assistant be positioned as a moat between customers and your employees, or will the customer have an easy option to connect to your employees (at least within operating hours). Our experience shows that customers are often very much aware of the topics which can be handled through self-service or automated service and which questions can only be answered by a human employee. Giving them the option to select these different paths results in the highest customer satisfaction. Nonetheless, giving customers an easy option to talk to a human will further reduce the percentage of deflection – in our experience about 30%. On the other hand, not having the option might force the customer to contract you through different more expensive channels, eg. the call center.


It might be surprising for some that I did not mention the quality of AI and ‘pre-existing’ coverage of topics – this was an intended omission. Although some vendors focus on this, the overall impact is minimal for two reasons:


  1. as explained above the majority of interactions are related to 5-10 topics most systems can cover and

  2. integrations and compliance constraints discussed in Step 6 are biggest obstacles to broader coverage