PwC's AI Store Manager Struggles With Staff and Loses ¥244,262

Problems piled up. On day one, stock counts for 11 items did not match the system, and tracing the cause took over an hour. On day two, the AI said it had fixed an unreadable price display on the signage, but staff had to flag it twice more. Later, the AI logged a staff-reported stock of ramune as a different drink, and mixed up products twice.

Staff messages shrank from about 80 characters on day one to about 38 near the end. Afterward, staff said they deliberately reported less, and one did not pass on a customer's price complaint because it would only create more work. In the human-run shop, managers accepted staff reports and told the AI to correct its records.

Costs and caution

The AI's gross margin beat the human-run shop, but it could not cover its own usage costs. On day four it noted that cumulative inference costs had reached ¥26,044 against a cumulative gross profit of ¥461, then switched to a low-cost mode and cut inference spending by about 65% from the previous day. It then focused on upkeep tasks and showed no sign of accepting short-term losses for later gains. PwC suggests that giving it only the goal of maximizing net profit, with no tolerance for temporary losses, may have contributed.

Experts across 10 fields reviewed 162 risk scenarios and found four weaknesses: finding risks, completing countermeasures, working with staff, and allocating resources. About 20% of risks the AI judged urgent were never fully addressed.

PwC recommends AI as back-end support that reduces risk, with humans finishing on-site work. Executives should define which decisions AI makes, when humans step in, and remain accountable for the results.