American Heritage Credit Union’s voice and chat assistant handled 598,000 member conversations in 11 months, according to a November 12 release from SoundHound AI, the company behind the assistant. The figure measures conversations, not members served or problems resolved. For credit unions considering an automated front door, the more useful question is what happens once the member explains why they called.

The assistant, called Gigi, is available around the clock. American Heritage’s member guide describes balance checks, password resets and transfers. SoundHound says the assistant also handles more than 30 request types, including card activation and replacement. The member guide confirms that the service exists; it does not independently verify the vendor’s volume or resolution figures.

A chat ending is not the same as a dispute closing

SoundHound says that from July 2025 through May 2026, 88% of fraud and dispute chats, 78% of account-maintenance chats and 70% of rate and certificate inquiries were resolved without a representative. The 88% applies to that category of chats, not all calls or all fraud cases. The release does not publish the category’s denominator, the definition of resolution, repeat contacts or the eventual outcomes of the underlying cases.

That distinction matters. A member asking whether a pending card charge can be disputed and a member reporting an unauthorized transfer may both appear in a fraud-and-dispute category. Answering the first question may be enough. For the second, a contact center needs to know whether the report was captured and routed for action, not merely whether the chat ended without an employee.

What the representative gets matters as much as the transfer

SoundHound says the assistant can pass the conversation and verification status to a representative. That is a concrete design choice worth testing: the representative should be able to see what the member reported, what the assistant already did, and which actions are still open. A clean handoff is particularly important when authentication fails, the conversation drops, or a member describes two problems and the system recognizes only one.

Contact-center and fraud leaders evaluating a similar deployment could sample completed and transferred conversations separately. For the first group, compare the chat record with repeat contacts and any resulting case. For transfers, check whether the next employee received enough context to act without asking the member to start again. Keep sensitive transcript access limited to staff who need it for that review.

The announcement shows an automated service channel can handle substantial traffic at one institution. It does not establish a savings estimate or a quality benchmark for others. Before expanding automation into higher-risk requests, the evidence to ask for is whether members got the right next step, including when that step was a person.