ImplementationAdoption storyFor: Customer support

iVisa built its own support assistant, then gave one job back to people

iVisa's in-house assistant handles about half of customer messages. Its support lead says it needs daily QA, and one job went back to the human team.

At a glance

Company
iVisa · Other · Size not stated
Team and use case
Customer support · Ticket deflection
Tools
In-house assistant (iVy) built on large language models; model not named
Stage
In production, one team
Results
  • Customer messages handled by the assistant: Around half; no earlier figure, measure not defined
  • Conversation types moved back to people: One (new conversations); measure not stated
Cost
Not disclosed. No figure for the build, the model usage or the team that runs the assistant.
People
A dedicated AI team inside customer experience reviews conversations daily (random sample plus negative ratings); people keep complex and urgent cases. Team sizes and any headcount change: Not disclosed.
Evidence
Company claim, no method

The company states a result without saying how it was measured.

What this leaves out: One interview with the director who runs the assistant. The share of messages is in CX Today's write-up, not his quoted answers; no method, baseline, satisfaction score or error rate is given.

iVisa, which sells help with visas and other travel documents, built its own support assistant instead of buying one, and it now handles around half of customer messages, [CX Today reported](https://cxtoday.com/how-a-cx-team-built-and-runs-a-multilingual-ai-assistant) on 19 November 2026 in an interview with Victor Gimenez, iVisa's director of customer experience. Gimenez says the team later moved one type of conversation back to people, because the human team had started doing it better.

What iVisa runs

  • The build: "We've built it internally, on top of LLMs, but we don't use a third-party vendor," Gimenez says. LLMs are large language models, the kind of AI behind ChatGPT and Claude; he does not name the one iVisa uses. He says the team has been building the assistant, which CX Today calls iVy, for a year and a half.
  • The work: the assistant started on informational requests, such as the status of an order or what a visa document means. People take complex and urgent cases; his example is a traveller at the airport.
  • Languages: prompts, the written instructions the assistant follows, are drafted in English and then tested in the 14 languages on iVisa's site. The assistant also replies in other languages when customers write in them.
  • Disclosure: every conversation, in every language, opens with a fixed statement that the customer is talking to an AI assistant.

The job that went back to people

Gimenez says the assistant at one point handled new conversations "better than the human team because of how we trained it". Then, he says, "the human team came up with a new process, worked on a new solution and a way to improve the experience, and started doing it better than the AI assistant. So we reverted it back." He does not say which measure decided it or how long each side held the work. The plan is to try again once the assistant can match the people.

On letting the assistant answer everything, he says "the experience just goes down, CSAT goes down". CSAT is the customer satisfaction score.

The people side

A "dedicated CX AI team" sits inside customer experience and works with engineering. It checks conversations daily: a random sample, plus conversations where the customer gave the assistant a negative rating. Gimenez says translated answers have been accurate. The weak point is triggers, the phrases that should make the assistant take a specific action, which it catches more slowly outside English, so the team checks before and after every launch.

His reason for the daily checks: "an error in the prompt is going to be repeated a thousand times exactly how you wrote it." He says support leaders give human teams QA, training and coaching, then cut that support "by 99.9%" for an assistant. His advice is to start with the same structure and drop only what makes no sense: "The process team is not a process team, it's actually the prompt team."

The interview does not say how many people are in the AI team or the support team, or whether support headcount changed.

The catch

  • The "around half" figure appears in CX Today's write-up, not in Gimenez's quoted answers. Neither says whether "handles" means closed without a person or answered first, and there is no earlier figure to compare it with.
  • No satisfaction score for the assistant or for people, no error rate and no count of wrong answers. Gimenez says a wrong answer "can lead to a customer stranded at the border in another country".
  • No cost: not the build, not the model usage, not the team that runs it.
  • No company size. CX Today says iVisa has helped more than 3 million travellers.

What to do if you run support

  • Before launch, name who reviews the assistant's conversations every day, a random sample plus every negative rating, and put those hours in the business case.
  • Score the assistant and your people separately on each type of conversation, and write down now the result that would move a type back to people.
  • If you serve customers in more than one language, test every trigger and handoff rule in each language before and after each change, not only in English.
Was this useful for your work?

Take this to work

Forward this

iVisa built its own support assistant, which handles around half of customer messages, then moved one type of conversation back to its human team because they did it better, its customer experience director told CX Today on 19 November 2026; no cost was given. Before we widen an assistant's scope, we should score it against our people on each type of conversation and fund daily review.

Questions to ask

  1. Does "handled" mean closed without a person, and what is our own share on that definition?
  2. What is our satisfaction score for the assistant and for people on the same type of conversation?
  3. Who reviews the assistant's conversations each day, how many, and what do those hours cost?
  4. What result would make us move a type of conversation back to people, and who decides?
  5. In which languages have we tested each handoff rule, and how often does the assistant miss one?

Questions this article answers

Did iVisa build or buy its AI support assistant?

iVisa built it in-house on large language models without a third-party vendor, its director of customer experience, Victor Gimenez, told CX Today in an interview published on 19 November 2026. He did not name the model or give a cost.

How much of iVisa's customer support does AI handle?

CX Today's write-up of 19 November 2026 says the assistant handles around half of customer messages. It does not say whether those are closed without a person, and gives no earlier figure.

Why did iVisa move work from its AI assistant back to human agents?

Gimenez says the human team built a new process for one type of conversation and began handling it better than the assistant, so iVisa moved that work back. He does not say which measure was used.

Who checks the answers iVisa's AI assistant gives?

A dedicated AI team inside customer experience reviews conversations daily: a random sample, plus conversations with a negative customer rating. Its size is not stated.

Send this to someone who runs a teamLinkedInXBlueskyHNEmail