Getting an AI system into production is only the first hurdle: businesses still need evidence that customers can rely on it, says Arato co-founder and CEO Shahar Erez. His advice is refreshingly specific: test complete customer journeys repeatedly, track where failures happen and keep validating after changes.
Watch Desk analysis
What happened
In an opinion article for CTech, Erez argues that a system doing well on familiar prompts is not enough to justify giving it sensitive data or control over business processes. A bank’s voice assistant, for example, might usually work and still make a consequential mistake.
Erez says his team ran a simulation in which an enterprise AI supplied a $400 price and later admitted it had made the figure up. He argues that testing should cover whole customer journeys, including people who are confused, vague or provide information in an unexpected order, rather than a handful of standard questions.
Why it matters
A weak chatbot answer may be easy to work around. A wrong answer about a refund, claim or booking can leave a customer paying for the error, and give them little reason to return.
Erez’s practical point is that companies need to know more than an overall accuracy score. They should understand which outcomes the system gets wrong, how often, and what those failures cost. He also says changes to prompts, models or data can alter behaviour, so checks need to continue through updates.
Our read
“Works most of the time” is a thin foundation for handing an AI agent access to customer accounts. Erez offers a more useful test: can the company show that the whole task works reliably, including its awkward corners, and spot when a change breaks something that used to work?
His simulation illustrates the risk, but does not establish how frequently such errors occur in deployed systems. The durable advice is to test the work customers actually need done, then keep testing as the system changes. (Read Erez’s article in CTech)
What to watch
- Whether companies test complete customer journeys rather than a small set of familiar prompts.
- Whether they track the frequency and consequences of failures, not just an overall accuracy score.
- Whether validation is repeated after model, prompt or data changes.
Discussion spark: For AI handling customer-facing tasks, should companies have to prove reliability across complete journeys before launch, or is ongoing monitoring after release enough?
Sources and evidence
- AI that usually works is not good enough for business | Ctech (30 October 2026, 10:43 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.