The Empty Chair

At 9.05 on a Thursday morning, after the strategy discussion in Chapter 12, another proposal arrived with the word AI in its title.

The small-claims system was one of the priorities Tom’s strategy had selected for investigation. It could compare claims with company rules and approve eligible amounts beneath a financial limit. The proposal now also asked it to decline claims it judged ineligible. That additional authority was written into the decision requested from the board.

One chair at the table was empty.

It belonged to a member of the standing customer panel introduced during the service-credit work. Her train had been cancelled. She had helped challenge the earlier human correction process. She could not speak for every customer, and the board could meet without her.

The sponsor finished his presentation. The demonstration had shown what the system could do on selected cases. Its performance evidence sat in an appendix. The question now was which live decisions that evidence could support.

Leila, attending from her regional role, asked what an employee would see when the system declined a claim. Daniel asked whether the source rule and its effective date would travel with the decision. Finance asked how many cases would require review near the threshold of the tested decision score. Nadia asked who could suspend approval authority without waiting for the next board meeting.

The score had been checked against known outcomes in the evaluation set. It was not the model announcing how sure it felt, and its earlier test did not establish how it would perform under every future condition. Finance wanted the observed workload around the proposed threshold, not a confident phrase in a generated answer.

Then Maya turned back to the business case. One line read: “Customers accept an automated decline when the reason is shown.” Beside it was the word confirmed.

“Which claim depends on customers,” she asked, “and who tested it?”

The sponsor pointed to the complaints figures and the notes from the customer panel.

“The panel asked us to explain decisions clearly. Complaints fell after we changed the correction process.”

“That was a person explaining and correcting a decision,” Leila said. “Did we ask about receiving a decline from this system?”

“No. We treated it as the same requirement.”

Nadia opened the complaints appendix. It described customers who had complained. It did not establish how many had misunderstood a refusal, abandoned a claim or found the challenge route too difficult to use.

“The empty chair isn’t the missing test,” Maya said. “We would need this evidence if she were here. One person could not confirm that sentence for everyone.”

Daniel suggested a remote conversation with the panel member that afternoon. The sponsor offered to send her a written question set. Leila proposed reviewing recorded calls about declined claims, within the permissions governing those recordings, to find what customers had struggled to understand.

Those steps could improve the test questions. None would settle whether customers could understand and challenge this proposed automated decline. The team agreed to try the proposed explanation and challenge route with a small, deliberately varied sample of customers. They would record whom the test did not represent and what uncertainty remained.

“What changes if it fails?” Finance asked.

“Declines go to a person,” Daniel said. “We consider automated approvals separately. The cost case changes because people still examine the refusals.”

“And approvals still require their own evidence,” Nadia said. “Removing one decision from the system does not approve the other.”

The board authorised a bounded comparison using appropriately authorised, minimised or de-identified historical records in a controlled environment, and the customer test of explanation and challenge. Access, retention and disposal had to be agreed before records were used. No live claim could be changed. The sponsor would return with separate proposals for automated approval and any automated decline, with the revised workload and cost case.

One director voted against the delay. He believed a limited release with a staffed correction route would teach the company more quickly. The team had evidence gaps, he said, but customers also waited while the company sought greater certainty. His objection and the proposed alternative were recorded beside the decision.

After the meeting, the sponsor caught Maya near the lift.

“You know she might tell us customers want a faster answer.”

“She might. We should hear that too.”

“So what changed?”

“We had counted an assumption as evidence,” Maya said. “Now we know which decision depends on it.”

The lift arrived. The sponsor stayed beside the open paper.

He moved the live-authority decision to follow the customer test and the revised comparison. Beside the specific claim that customers accepted an automated decline when a reason was shown, he removed confirmed and wrote not yet heard.

FIG-E.1

Correct the claim in the paper

Claim

Customers accept an automated decline when the reason is shown.

Earlier label

Confirmed

Corrected label

Not yet heard

Decision consequence

Test comprehension and challenge. If unsupported, refer declines to a person; any automated approval still needs its own evidence.