They named the independent oversight panel later that week. Cheryl agreed to chair after securing authority to publish adverse findings, and two experienced reviewers joined her: a former hospitality worker who had organized workplace training and a retired auditor with expertise in consumer complaints. Thompson Holdings would pay the costs but could not dictate individual outcomes. Representatives from civil-rights organizations were invited to advise on the framework only if they wished to participate, without presenting their involvement as automatic approval.
That distinction mattered when the first press release was drafted. A communications consultant had written that the company was partnering with several prominent national organizations, none of which had agreed to anything. Victoria crossed out the claim. "We can ask for help," she told him. "We do not get to borrow someone else's reputation just because their name would look good beside ours."
The next challenge arrived from the technology team. Marcus had suggested that monitoring software could flag problematic phrases in customer messages and staff reports before they turned into repeated complaints. One vendor offered a comprehensive package that would record and analyze audio in dining rooms, automatically identify language it considered discriminatory, and notify managers in real time. The demonstration was impressive, until Anthony raised a question from the back of the conference room.
"Who listens to the recordings?" he asked. "And what happens when the system mistakes a customer's story about discrimination for an employee actually discriminating?" The vendor gave a complicated answer about probabilities, confidence intervals, and proprietary models. Anthony repeated the question in simpler language. He wanted to know who could hear the conversation and who could be harmed by a machine that misunderstood it.
Victoria asked the vendor to stop the demonstration. She had nearly authorized the purchase because the prospect of an automatic warning system seemed like a modern answer to a deeply human failure. But her company had already made the mistake of letting a system classify people without sufficient review. The last thing she wanted was to replace a biased manager with a biased algorithm hidden inside expensive software.
The technology group redesigned the proposal. They began with analysis of voluntarily submitted complaints, written guest feedback, and incident reports rather than continuous recording of every dinner conversation. Where audio tools might later be useful, they required explicit consent, limited retention, independent privacy review, and a clear method for correcting erroneous alerts. No automated score would decide whether an employee kept a job, and no recording would be collected simply because a customer appeared unfamiliar or uncomfortable.
"The tool can tell us where to look," Victoria said. "It cannot tell us what a person is worth. We have tried letting a score do that before, and it brought us here." The development team wrote the sentence into the design brief. It became the principle they returned to whenever the technology promised an effortless solution that would have required surrendering judgment to a machine.