Contents

Chapter 8

Going Live

At 8.38 on Monday morning, Maya signed the decision to release the retrieval feature to a small group of service advisers.

She read the last line before signing it.

The feature may assist an adviser to locate and check service-credit procedures. It may not determine a customer’s entitlement, approve a credit or send a customer message.

Leila had signed off the source collection on Friday. Daniel had tested access using an ordinary adviser’s account, checked the failure display and confirmed that he could disable the feature for the group. The team leads had worked through the incorrect approval answer. Priya had helped prepare the staff session and the route for reporting a doubtful output.

Maya authorised a limited live use. The group would start that afternoon. Leila would own its effect on customer operations. Daniel would own the configuration and access. Nadia would review what the team learned before Maya considered any wider release.

The decision did not say the feature was safe in every situation. It said who could use it, for which work, with what checks, and who could stop it.

Maya sent Leila a short message: You can begin with the agreed group.

Leila read it, then looked at the service roster. One of the advisers scheduled for the afternoon session had been moved to cover an absence. Another was training a new colleague. The launch had been approved. The working day had not arranged itself around it.

The first change was to the roster

Leila called Priya and the two team leads together.

“We have permission,” she said. “Who can actually attend?”

They moved the session by an hour and arranged cover for the advisers who would join it. That took more effort than enabling their accounts. Leila did not want half the group to receive a hurried demonstration between calls and the other half to learn by asking a colleague.

The Implementation Guide gave Daniel a sequence for access, checks, release and support. Leila used the Change and Adoption Management Plan for a different sequence: who needed to understand the change, what they would do differently, and how their ordinary work would continue if the feature was wrong or unavailable.

The feature sat inside a knowledge base the advisers already used. That made it easier to find. It also made it easier to assume that anything on the screen carried the same authority as an approved procedure.

Leila wrote the change in two sentences for the team leads.

Advisers may use the feature to find relevant procedure guidance. They must read the current source, establish the case facts in the case system and follow the existing approval route.

The team leads checked the wording. One asked whether “may” meant use was optional.

“Yes,” Leila said. “If the ordinary procedure search is clearer, use that. We want to learn when this helps.”

That answer mattered. A tool introduced with a target for use can turn every problem into an opportunity to use it. Leila wanted advisers to choose the route that helped them establish the instruction. She did not want them to create questions for the assistant so a launch report would show activity.

Could your adoption target reward use that makes the work harder?

Guidance

Consider a case where the ordinary process already works well. Describe how you would recognise useful assistance without treating every decision to work without the feature as a failure.

Your answers are saved only in this browser on this device. Nothing is sent anywhere.

Adoption is not the same as use. People can use a tool often and still do the work worse.

The team leads also asked what would happen to calls that took longer while advisers checked an answer. Leila would allow time in the initial group. She would not judge the release by the first week’s call times alone. A faster wrong decision would be a poor result.

That was the first practical cost of going live. Someone had to cover the work while people learned, and someone had to let them use what they learned.

What staff were told

At 11.06, Priya opened the draft announcement.

The first version began: We are pleased to introduce an AI assistant that will streamline service-credit decisions.

She deleted the sentence.

“It does not make the decisions,” she told Daniel.

He had not written the draft. It had come from a standard announcement outline. He agreed that it promised the wrong change.

They used the Staff Communications Pack to produce a message that an adviser could act on. It named the small group with access, the work in scope, the start time and the person to contact. It said the feature could help locate procedure material. It also said what advisers must still do themselves.

The message gave an example of a permitted question: Which procedure covers a credit after an earlier decision is reversed? It told staff not to paste names, account identifiers, full customer messages or unnecessary case history into the question box. The customer record would remain in the authorised case system.

Daniel added how access would appear in the existing knowledge base. Advisers would not be asked to verify access through an emailed link. That sentence came from the suspicious message Priya had reported during preparation. The launch email itself contained no sign-in link.

The message also explained how to flag a doubtful answer. Advisers could use the existing service support route, sending the question, output and source link without adding unnecessary customer details. If a customer decision had already been affected, they would tell a team lead first.

Priya read the draft aloud. It sounded less exciting than the first version. It also told her what to do at 3.00 that afternoon.

Nadia read it next.

“Where does it say what this is called?” she asked.

The team had used “retrieval feature”, “assistant” and “knowledge-base feature” in different meetings. The screen itself used another label. Priya suggested that the announcement use the label advisers would see on screen, followed once by a plain description: the service-credit procedure search. They made that change in the training material as well.

A small difference in names can be managed in a meeting. On a working floor, it can send someone to the wrong support page or make them think a different feature has been approved.

The announcement went to the limited group and their team leads. Leila also sent a brief note to the wider customer operations team saying that the feature was being introduced to a defined group. Others were not to borrow a colleague’s access or treat a screen capture as approved guidance.

She expected some staff to ask why they were not included. That was easier to answer than an unexplained difference in access.

Training with the wrong answer

At 1.41, Priya projected the answer from the previous week.

The adviser may approve the credit.

Several people in the room noticed the current procedure linked beneath it. One adviser said it looked convincing. Another saw the wrong approval instruction before Priya pointed it out.

Leila asked them to use the answer as they would during a call.

A team lead interrupted.

“Do the case first,” she said.

They began again. What did they know about the customer’s case? Was there a recorded reversal? Had a credit already been promised? What did the current procedure say? Which step could the adviser perform, and which needed approval?

The source link was useful. It took them to the relevant procedure. Reading the passage revealed that the answer had combined identifying a possible credit with approving one. They then checked the case facts in the case system. The exercise ended with a referral to the team lead, not a promise to the customer.

Priya did not ask the group to memorise a list of possible AI errors. She asked them to show where the answer stopped being evidence and started becoming a proposed action.

A source link helps you check an answer. It does not approve what the answer says.

The next exercise was less clear. An adviser asked which procedure applied after a reversal, but the prepared case did not say whether the reversal had been authorised or merely discussed with the customer. The feature returned a relevant passage. It could not supply the missing fact.

One participant said the tool had failed because it had not given a definite answer.

“What would a definite answer be based on?” Nadia asked.

Nobody could point to a record establishing the reversal. The group went back to the case system and the ordinary escalation route. A useful assistant would leave that gap visible. An answer that filled it would make the work easier to finish and harder to defend.

The final exercise used an unavailable feature. Daniel had arranged a test view showing the failure message he had repaired. The advisers used the approved working copy of the procedure and referred the case when they could not establish the instruction. They did not need to wait for Daniel to restore the screen before helping the customer.

The session took longer than planned. Leila did not cut the reporting exercise to recover the time.

Priya put up a doubtful answer, asked each adviser to find the source, then asked them to report the problem. The report needed enough detail for Daniel and Leila to reproduce and assess it. It did not need the adviser to diagnose the technical cause.

One adviser asked whether reporting a wrong answer would count against her.

Leila answered directly.

“No. If you notice it and show us, that is part of using this properly.”

She knew the sentence would have to hold up later. If a team lead treated reports as delays or mistakes, staff would learn a different rule.

Who signs off on what

After the session, Maya asked to see the Delegation Matrix and Accountability Map.

The two documents answered related questions. The Delegation Matrix showed which decisions the feature could support and where a person’s authority began. The Accountability Map showed who would maintain the conditions around the use and who would respond if they failed.

Daniel had marked the feature as Assist. Maya asked him to explain why it was not Recommend.

“It can suggest which procedure is relevant,” he said.

“That sounds like a recommendation.”

“It may be, in ordinary language. But the authority mode is about what we let the output do. It helps the adviser locate and interpret material. We do not treat its proposed answer as an approved case recommendation.”

Nadia thought that distinction needed care. If an adviser copied the output into a case note as the reason for granting a credit, the organisation would be using it as more than retrieval. The label in the matrix would not prevent that drift.

Leila added a rule for the case record. The adviser would record the applicable procedure and the facts supporting the decision, with the required approval. The assistant’s wording was not the decision rationale. If the output helped locate the procedure, the adviser could still identify that source. They should not paste an unverified answer into the record and call it a reason.

The matrix placed credit approval with the existing team lead where the procedure required it. Leila owned the service outcome and the approved source material. Daniel owned configuration, access, logs and the ability to disable the feature. Priya helped staff use it and report problems, but she was not made responsible for approving every answer it produced. Nadia challenged the evidence and any proposal to expand its use. Maya owned the decision to authorise, widen, narrow or stop the release.

FIG-8.1

Authority during limited live use

WorkAssistant’s partPerson with authority
Find a service-credit procedureRetrieve and display relevant materialAdviser checks the current source
Establish what happened in the caseNo access to decide case factsAdviser checks the case record and customer information
Identify a possible creditHelp locate the applicable instructionAdviser applies the procedure and refers where required
Approve a creditNo authorityTeam lead or other authorised approver under the procedure
Change the source collectionNo authorityLeila approves source status; Daniel makes the controlled change
Pause the featureNo authorityDaniel can disable it; Leila and Maya direct the service response
Source: author’s method and fictional case.

Maya asked whether Daniel needed to call her before disabling access after a serious problem.

“No,” Leila said. “He should stop exposure first and tell us.”

Daniel agreed. The Accountability Map recorded the difference between an immediate technical pause and a decision about how the service team would handle affected work. Daniel could disable the feature. He could not decide that a customer’s credit was invalid. Leila could direct the service response. Maya would decide whether the authorised use should resume or change.

Who can pause the feature, and who takes care of the work affected by that pause?

Guidance

Walk through one plausible failure with the named owners. Keep the technical stop separate from decisions about a customer case, then identify who decides whether use may resume.

A person who can stop a tool needs to know when to stop it. A person who owns the outcome needs to know what stopping changes.

The team tested the map against the wrong approval answer. Priya would report it. A team lead would check any affected case. Daniel would preserve the relevant output, assess whether the failure could recur and pause access if needed. Leila would confirm the procedure and direct corrections. Maya would be told if the boundary or release decision was in question.

No one had to guess whether “the AI team” or “the business” owned it.

A skill nobody had asked about

At 3.18, the first advisers signed in.

Daniel stood beside Priya while they checked access. One could see the feature. Another could not. Daniel corrected the group permission, then verified that the adviser saw only the intended source collection. It was a routine access problem. It happened at the point where a launch plan met individual accounts.

The adviser who gained access asked whether she was supposed to use the assistant on every service-credit case.

“No,” Priya said. “Use it when it helps you find the procedure. Tell us when it does and when it does not.”

She stayed for the first few cases. One adviser asked a broad question and received a long answer. Priya suggested a narrower question about the procedure step the adviser needed. The shorter answer was easier to check.

Another adviser knew exactly where the procedure was and opened it directly. She did not use the feature. Priya marked that as ordinary work, not a failure of adoption.

Leila had used the Skills Self-Assessment before the session. It showed that advisers needed different kinds of help. Some were comfortable finding a source but uncertain about whether an answer overstated it. Others could assess the case well but were unfamiliar with the knowledge-base search. A few wanted to know how to phrase a question. They did not all need the same extra training.

Priya arranged short follow-ups around those needs. She also asked team leads to watch how advisers worked, not simply whether they had attended the session.

At the end of the afternoon, one adviser told Leila the feature felt like another place to check. It had helped her find a passage. It had also added a step because she now read the answer and then the source.

Leila thanked her. The complaint was useful. If the feature merely inserted a polished paragraph between the adviser and a procedure she could already find, the team needed to know. The benefit could not be assumed from the presence of an answer.

The customer is not in the room

The limited release was internal. Customers did not see the feature or its outputs. They could still feel its effects.

At 4.26, Priya listened while an adviser handled a service-credit question. The adviser used the feature to find the current procedure, then checked the case record. A previous decision had been reversed, but the record did not show who approved the reversal. The adviser told the customer she needed to confirm one detail before giving an answer.

After the call, the adviser worried that she had sounded uncertain.

“She was uncertain,” Priya said. “That was the truthful part.”

Leila checked the case with the team lead. They traced the missing approval and corrected the record before responding to the customer. The feature had helped find the relevant procedure. It had not solved the old weakness in recording reversals.

The case gave Leila two pieces of work. One was to support the adviser in explaining a pause to a customer without blaming a system. The other was to make sure the recording change promised after the experiment was happening in ordinary cases.

She added a short line to the team guidance: when a required fact is missing, explain the next step and when the customer can expect an answer. Do not say “the AI cannot decide”. The customer had asked the organisation for a decision. The organisation still owned the response.

Nadia asked whether customers should be told that an adviser had used the feature.

Maya did not answer with a general rule for every possible AI use. In this limited release, the feature was helping staff locate internal procedure material. It was not speaking to customers, generating a customer message or deciding entitlement. The service team would continue to give accurate explanations of its decisions and respond honestly to questions about its process. If the use changed, Maya would revisit what customers needed to know.

The standing customer panel would be one place to test whether the organisation’s explanations made sense to people on the receiving end. It was not a substitute for dealing properly with the customer in front of the adviser.

The end of the first day

At 5.07, Leila stood beside Priya’s desk while the last adviser finished a case note.

Access had been enabled for the agreed group. People had attended training, though one needed a follow-up after missing the session. The source collection and fallback were in place. The ordinary approval process had held in the cases they had examined.

None of that established a time saving. It did not show how advisers would behave after the first week, when Priya was no longer nearby. It did not show whether a later source update would change answers quietly.

Leila sent Maya a short account of the first day. She included the access problem and the missing reversal record. She also described the case where the feature helped an adviser find the right passage and pause for a missing fact.

Maya replied: Keep the boundary as agreed. Bring me the pattern, not just the incidents.

Leila read that twice. Incidents mattered, especially where a customer could be affected. But a pattern might be forming without a dramatic failure. Staff might stop opening sources. They might use the feature for questions outside service credits. They might avoid it because checking took longer than ordinary search. A quiet change in any of those behaviours would matter.

The go-live decision had put the feature into real work. It had also started a new obligation: to notice what the work became.

The first doubtful answer

At 10.52 the next morning, a report arrived through the service support route.

An adviser had asked about a service credit after a missed appointment. The feature found a relevant procedure but included a sentence about a different kind of service failure. The adviser had opened the source, noticed the mismatch and used the correct passage. No credit had been approved from the wrong sentence. She submitted the question, answer and source link without copying the customer’s details.

Priya called to thank her. Then she took the report to the team lead and Daniel.

Daniel could reproduce the answer. He wanted to adjust the source collection immediately. Leila asked him to wait until they understood whether the two procedures were poorly separated or whether the answer had joined them despite clear boundaries. A quick change might improve this question and weaken another.

They classified it as a doubtful output, kept the limited group working with the existing checking rule, and reviewed related questions. Daniel prepared a controlled change. Leila checked the source labels. Nadia asked them to keep the original result for comparison.

The report consumed time. It also showed that the route for reporting worked in the first live use. The adviser had noticed an error, avoided an affected decision and shown enough evidence for the owners to investigate.

Maya did not announce that the launch was a success. One correctly handled report was evidence of one control working once. She asked how the team would know whether other advisers were encountering similar answers and saying nothing.

Leila could ask team leads what they were seeing. Daniel could examine permitted use records for patterns without treating every question as a staff performance measure. Priya could collect examples during follow-ups. Those views would still miss some problems.

The team needed a monitoring plan. They also needed to decide how much review their small release could sustain.

Use the Implementation Guide, the Change and Adoption Management Plan, the Staff Communications Pack and the Skills Self-Assessment to prepare staff. Record decision rights and owners in the Delegation Matrix and the Accountability Map.

Chapter 9, Watching It Work, asks what the team should measure now that the feature is in use, and which changes should bring Maya back to a decision.