Contents

Chapter 14

The Standing Arrangement

At 8.29 on Tuesday morning, Leila forwarded Maya the notice confirming her move to a regional role.

The move would begin the following month. She would still be in the organisation. She would no longer run the service team whose procedures fed the retrieval feature.

Daniel replied with a question.

“Who owns the sources when you move?”

Leila could name the person who would manage customer operations while her replacement was appointed. She could not point to a decision giving that person authority to approve changes to the service-credit procedure collection.

Until now, Daniel had called Leila when a source changed. Priya had gone to her when advisers found an answer they could not reconcile with a procedure. Nadia had challenged the evidence. Maya had made the decision. It had worked because they knew one another and remembered why each boundary existed.

Maya opened the register. The service-credit entry named Leila. It did not say what happened when Leila changed roles.

“Do not change the owner to the next name in the directory,” she told Daniel. “Let us decide what the role is first.”

There were other entries waiting. The approved meeting-summary use needed its ordinary review. The tender trial remained closed. A small-claims proposal was taking shape. The company was no longer managing one bounded experiment with five people around a table.

Maya asked Nadia to bring the register, the legal scan and the service-credit decision to a meeting that afternoon.

What had lived in the room

Daniel began the meeting by describing the current arrangement. He configured the retrieval feature, tested source changes and could disable it. Leila owned the procedure collection and the service outcome. Nadia challenged the evidence. Maya authorised the use and its renewal.

“That is still right,” Leila said.

“For this month,” Nadia replied.

Leila would soon work across regions. She might still know the service-credit procedure better than her successor. Knowledge was not authority to approve a change for a team she no longer managed.

Maya asked what would happen if Daniel was away when a source update arrived.

“Someone in technology could make the change,” he said. “They would need the configuration and the test questions.”

“And if Nadia is away?”

“Someone still has to challenge what we conclude,” Leila said.

The arrangement had been built for a project. A project can depend on named people gathering often, remembering the reasons for a limit and resolving questions through a call. A standing use has to survive ordinary absences, new staff, changed roles and work that arrives without an invitation to the original group.

A role that exists only in someone’s memory leaves when that person does.

Maya resisted Daniel’s first draft. It contained a long approval chain for every possible use. An adviser wanting to summarise an approved internal meeting note would have had to pass through the same sequence as a proposal to let a system decline a claim.

“We will build a queue that people learn to avoid,” she said.

Nadia agreed. The route had to fit the consequence. The three rows already in use gave them a starting point: routine use within existing permission, bounded experiment, and consequential system. The five authority modes helped describe what an output would be allowed to do.

They did not need a new committee to decide every question. They needed to know who could make which decision, what evidence that person required and when a proposal had changed enough to need a different decision-maker.

Maya gave them a test. The arrangement should let someone who missed every meeting find the current boundary, the person who owned it and the next decision date. If it could not do that, another page of process would not make it standing governance.

Could someone who missed every meeting find the current boundary and owner?

Guidance

Try this with one live use and the records a new colleague would actually receive. Ask them to locate the current decision and next review date. Note where they need someone else’s memory to explain a limit or find the person who can act.

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

A register that can be read

Nadia projected the provisional register. It had begun as a way to see uses that were happening or being proposed. It had done that job. It also contained short entries that assumed a reader remembered the meetings behind them.

The service-credit entry said Assist, small group, current procedures. It did not show the narrowed behaviour for approval and eligibility questions at a glance. The Decision Record did, but a new service lead would have to know where to find it.

The tender entry said closed. It did not show the specific condition for reconsidering the proposed use of a prospective customer’s documents.

The meeting-summary entry showed an owner and purpose, but its review date belonged to a person who had changed teams.

Maya asked Nadia to make each entry answer the questions a new owner would ask on the first day. What is the use? Which row is it in? What may the output do? What information and people are in scope? Who owns the result and the sources? Who can configure or stop it? What decision currently authorises it? When is the next review? What event brings review forward? Where is the fallback if it stops?

“That is a lot of fields,” Daniel said.

“It is less than reconstructing three months of messages,” Leila replied.

They tested the fields against service credits. Leila could state the boundary. Daniel could locate the configuration and approved source set. Maya could find the expiry and renewal conditions. Priya could find how to report an odd answer. Nadia could see which legal questions remained open.

FIG-14.1

What a standing register entry must lead to

Question for the next ownerService-credit retrieval entry
What work is authorised?A small adviser group may use it to find and check service-credit procedures
What may the output do?Assist with retrieval; approval and eligibility questions show sources and status, not a drafted outcome
Who owns the result and sources?The customer operations role responsible for the service-credit procedure and outcome
Who can change or stop the feature?The designated technology owner controls configuration and can disable access
Where is the current authority?The latest Decision Record, with conditions and expiry
What brings review forward?An unreviewed source change, affected customer decision, material incident or proposed change of scope
What works if it stops?Approved ordinary procedure search and team lead escalation
Source: author’s method and fictional case.

FIG-14.2

The standing arrangement loop

1

Authorise

Owner proposes, challenger tests, decision-maker confirms.

→

2

Monitor

Register entry, review date and the triggers that bring it forward.

→

3

Review

Exceptions first: expired authority, changed data, incidents, new audiences.

→

4

Renew or retire

Confirm, narrow, restrict or end - and record the decision.

Then: a renewed use returns to monitoring; a retired one leaves cleanly.

Approval begins a new set of responsibilities; it is never the last decision. Source: author's method and fictional case.

Maya read the table as if she had joined the organisation that morning. It told her enough to ask the next question. That was the standard she set for the rest of the register.

The question with no owner

At 3.14 on Thursday, a team lead sent Priya a screen capture from the knowledge base.

An adviser in another customer team had asked whether the service-credit retrieval feature could help draft a response to a complaint about a delayed refund. The feature was visible from the same knowledge-base screen. The adviser had not sent a customer message or entered a customer record. She had asked the team lead before trying the case.

Priya called Leila.

“Is that within our permission?”

“No. It is a different process, source collection and customer outcome.”

The adviser had made the right call by asking. The near-miss was in the route she had been offered. The screen made the feature appear generally available, and no one in the other team knew who could answer her question. She had asked Priya because Priya’s name appeared on the service-credit training material.

Daniel checked access. The other team could see the knowledge base, but not the restricted service-credit feature. The screen capture showed a general description that could be mistaken for permission to use the same capability on other work. He corrected the description. Leila spoke to the team lead about the difference between a procedure search and a drafted customer response.

Maya asked who owned the proposed refund use.

No one did yet. The team lead had a problem to solve, but she had not proposed a bounded use, selected sources or taken responsibility for the outcome. Daniel could not configure a route from a screen capture. Priya could explain the service-credit boundary, but she could not approve a refund workflow.

They put the question in the register as a proposal awaiting an owner. That did not mean a long review had begun. It meant nobody could treat the existing permission as covering it while the question remained open.

When nobody owns the question, the existing permission does not grow to fill the gap.

Nadia asked what a sensible first response to the team lead would be. Maya did not want “fill in a form” as the only answer. Leila offered to meet the team lead, understand why ordinary refund procedure search was insufficient and identify who in that operation could own the problem. The result might be a better document, a routine use already inside permission, a bounded experiment or no AI use at all.

At their meeting, the team lead explained that advisers were spending time reconciling two refund instructions. Leila recognised the shape of the problem. She did not offer the service-credit assistant as a shortcut. She asked which instruction was current and who could settle the conflict. The team lead took that question to the refund process owner.

The register recorded the proposed use and the unresolved procedure question. A week later, the owner supplied a corrected instruction. The team lead no longer needed an assistant to draft the response she had imagined. The register entry was closed with that reason. The gate had done useful work without producing a new AI project.

Who can say yes

At the next meeting, Maya drew three rows across the whiteboard.

For a routine use already inside approved permission, the person responsible for that work could allow it under the standing rules. The meeting-summary feature was an example. A meeting participant could use it for permitted internal notes, check the draft and keep the correct record. Each meeting did not need Maya’s signature.

If someone wanted to change its input to personnel material, send summaries automatically or use them as the official account of a disputed decision, the use would no longer be the same. It would need a new entry and review.

A bounded experiment needed a named business owner, a question, data and access limits, a challenger, an end date and a person authorised to make the end decision. The service-credit comparison had used that route. Its permission had not turned into live authority by momentum.

A consequential system needed a higher decision. If an output would Determine or Act in a way that affected a customer, employee or other person, the proposal had to identify the rule, consequence, evidence, review route and owner before the organisation considered that authority. A small-claims system that could decline a claim belonged here, even if its first screen looked like an ordinary assistant.

The disagreement was over who could authorise the middle row. Maya could not sign every small experiment. She also did not want a team to call a consequential use a pilot and begin making real decisions under a lighter label.

Nadia suggested that authority follow the work and consequence. A designated business leader could authorise a bounded experiment within an agreed area if no live decision took effect, permitted data and access were clear, the challenger could question the result, and the decision was recorded with an end date. Maya would retain authority for uses that crossed business areas, changed a person’s position or exceeded those limits.

Daniel asked who decided which row a proposal belonged in.

“The owner proposes the row,” Nadia said. “The challenger tests it. The authorised decision-maker confirms it.”

That did not guarantee a sound classification. It made the classification visible and gave someone a job to dispute it. A proposal described as Assist could still function as Recommend if staff were instructed to treat its output as the proposed outcome. The team would examine the workflow, not accept the label.

The challenger after Nadia

Nadia had challenged the service-credit comparison because Maya asked her to. She had followed the source-change recovery and renewal. She could not personally review every proposed use as the register grew.

“Do we put my name in every row?” she asked.

“No,” Maya said. “We put a role there.”

They worked through what the challenger had actually done. Nadia had asked whether prepared questions were too helpful, whether an answer’s source was current, what monitoring numbers could and could not show, and whether a repair addressed the cause of a failure. She had not taken ownership of the service outcome or configured the feature. She had made it harder for the decision owner to mistake an attractive result for sufficient evidence.

The standing arrangement would require a challenger for bounded and consequential uses. The person would be appointed from outside the immediate delivery work, with enough knowledge to ask useful questions and access to the evidence behind the proposal. The business owner could not appoint a challenger and then deny that person the failed tests or cost assumptions.

What evidence would your challenger need to see to question a release?

Guidance

Choose a decision now being prepared. Consider which failed tests or cost assumptions might change its outcome. Check whether the challenger can inspect that material independently of the delivery team’s account.

For routine uses inside an established boundary, the standing review could perform that challenge without arranging a new meeting for each ordinary task. If staff reported that the meeting-summary feature was being used for personnel decisions, the register owner would bring it forward.

Daniel wanted the challenger to be able to stop a release.

“The challenger can raise a material objection,” Maya said. “The decision owner must resolve it or record why they are proceeding. For a consequential use, it comes to me with the objection visible.”

Nadia accepted that. She did not want a title that suggested she could overrule the business owner without owning the result. She wanted an objection to survive a meeting and follow the decision.

They tested the rule with a hypothetical request to expand the service-credit feature to every adviser. Leila, as current owner, could point to useful searches. Daniel could say access was easy to widen. The challenger would ask whether the source review, staff support and monitoring could scale. Maya would see that objection before deciding. The technical ease of enabling more accounts would not become the evidence for expansion.

They added cover for absence. If the assigned challenger moved roles or was unavailable, a substitute would be named before the decision, rather than leaving the field blank because everyone assumed Nadia would eventually read it.

Challenge must have access to the evidence and a place in the decision.

A cadence people can keep

The first proposal for a standing review calendar was ambitious. It asked every owner to update every register entry each month, even when nothing had changed.

Leila imagined the new customer operations lead completing the same statement repeatedly while trying to run the service team. Daniel imagined a meeting full of status reports that left little time for a source change that mattered.

Maya crossed out the monthly update for every entry.

They kept different rhythms for different work. Live consequential uses would receive closer review under their specific decisions. Bounded experiments would be checked at their stated end date and sooner if a trigger occurred. Routine uses would be reviewed on a lighter standing cycle and brought forward when scope or reports changed. The legal scan had its own cadence and triggers, linked to the register rather than repeated as a separate conversation with no decision owner.

The register owner would check forthcoming review dates and missing owners before each standing meeting, and the meeting would take exceptions first.

Daniel asked whether the system could send reminders. It could. Maya wanted the reminders assigned to roles with a named current holder. An automatic message to Leila after her move would satisfy a software setting and miss the person who could act.

Leila helped define the handover. Before moving, she would show the incoming service owner the source collection, narrowed boundary, open quality issues, incident history, review date and fallback route. The incoming owner would accept responsibility for the service outcome and approved procedure status. Daniel’s technology role would remain separate.

Priya would still help advisers. She would not become the default business owner because she knew the most about the feature’s daily use. That distinction protected her as well as the arrangement.

The first standing review was scheduled before Leila’s move. Nadia brought the exceptions, not a tour of every entry. The service-credit renewal was current. The tender trial remained closed while a specialist question was examined. The refund proposal had been resolved through a procedure repair. The meeting-summary use had no change in scope. The small-claims proposal needed a defined decision path before it could be considered.

They finished with actions and owners. Maya did not call the meeting a success because it was short. She asked whether it had brought forward any question that might otherwise have been missed. It had. That was enough reason to keep the cadence.

The page everyone sees

Maya opened the AI Policy One-Pager.

It described the company’s general rules in plain language: use approved tools for approved work; keep information within the permitted boundary; check outputs against the relevant source and facts; preserve human approval where the process requires it; report doubtful results; ask for a decision before changing the purpose, audience or authority of a use.

The page was readable. It could not tell an adviser which service-credit procedure was current. It could not authorise the refund proposal. It could not settle the tender question or the small-claims investigation.

“Then why keep it?” the CFO asked.

“Because most people need to know the doorway,” Maya said. “They do not need the whole register on their desk.”

The one-pager would be the public face of the internal arrangement. It would tell staff how to recognise a permitted use, where to find its owner, and what to do when the work no longer fitted. Behind it sat the register, source and decision records, legal scan, monitoring and review dates.

They tested it with the refund near-miss. An adviser reading the page would know that an approved service-credit use did not authorise a different customer response. The page would direct her to raise the proposed use with the owner of the refund process. The register would show who took responsibility and what decision followed.

They tested it with Priya. She could find the service-credit entry and current boundary without asking Leila from memory. She could also find the reporting route for a doubtful answer.

The Stakeholder Map showed where the arrangement reached beyond the people in Maya’s meeting. Advisers and team leads needed usable instructions. Procedure owners needed to know when an edit affected a live feature. Technology staff needed the current access and disable route. Finance needed the work of maintaining a use to remain visible. Decision-makers needed the limits and expiry. Customers needed accurate explanations and a way for their experience to be heard.

Maya asked them to remove names from the map where the real relationship was with a role. Names would sit beside roles for current contact. The role would survive the next move.

A place to hear from outside

The standing customer panel had existed before this meeting. It was not created to approve AI uses. Maya wanted it connected to the arrangement in a modest way.

Leila described what the panel could do. It could tell the organisation when a service explanation sounded evasive, when a delay was difficult to understand, or when a customer could not see how to question an outcome. Those observations might reveal a problem in an AI-assisted workflow or in an older process that had nothing to do with AI.

The panel would not inspect source configuration or decide whether a service-credit answer met a procedure. It would not replace direct complaint handling. A customer’s own case would still need to be addressed through the service route.

For a proposed consequential use, the panel could help test how the process and its review route appeared from outside, at a suitable point before a decision. The organisation would still need to make and own that decision. The panel’s value was in showing a consequence the people designing the process might miss.

The small-claims investigation made this concrete. A proposed decline authority could not be judged only by how accurately a system matched a rule. The organisation would need to understand what a refusal said to a person, what explanation they received and how they could contest a mistake. The panel could inform those questions. It could not grant the authority.

At the next panel conversation, Leila asked about the clarity of service explanations without presenting the panel with an internal technology diagram. One member described a previous service contact where “we are checking” had given no indication of when to expect a reply. Leila took the observation back to the service team. It supported the guidance Priya had already helped write: when a fact was missing, give a next step and timing.

The panel had not audited the retrieval feature. It had shown what uncertainty felt like on the other side of a service conversation. Maya placed it in the Stakeholder Map as a standing source of outside experience, with a route for relevant findings to reach a use owner and the decision-maker at review.

The handover test

On Friday at 4.38, Maya asked someone who had not attended the service-credit experiment to read the new register entry and one-pager. The incoming customer operations lead had been told the service-credit process was part of the role, but had not been in the earlier decision meetings.

Maya gave her a practical question.

“A procedure owner asks to change the approval wording on Monday. What do you do?”

She found the service-credit entry. It told her that she owned the outcome and source status once the handover was complete. She found the current Decision Record and saw the narrowed boundary. She could identify Daniel’s technology role, the test before publication, the review trigger and the ordinary fallback.

She still had a question. The entry said incoming customer operations lead but did not name the date on which Leila’s authority ended and hers began.

Leila added the handover date and signed the transfer with her. Maya recorded the change. Daniel would route source notifications to the new holder from that date. Priya would be told whom to contact. A review reminder would follow the role and its current holder.

The test had found a missing detail before it caused a dispute.

Maya asked the incoming lead one more question.

“An adviser asks to use the same feature for a different kind of customer decision. Can you approve it?”

“I can hear the problem and own a proposal in my area,” she said. “I cannot extend this service-credit permission to cover it. We would have to classify the new use, name its boundary and record a decision at the right level.”

Maya closed the folder. That was the answer the arrangement needed to produce when none of its original authors was present.

The standing arrangement was not a guarantee that every use would be wise. It let the organisation see a use, name its owner, hear challenge and stop it when the reason for permission no longer held.

It was also small enough to operate. A routine meeting summary stayed routine. A bounded experiment had an end decision. A proposed system that could decline a claim could not borrow the permission of a procedure search.

That evening, Maya began a note for Tom. He had asked for an AI strategy months earlier. She now had more than a list of possible products. She had evidence of what the organisation could run, what it cost to keep running, and who would own the next decision when the original team was no longer in the room.

Write an AI Policy One-Pager that staff can use to recognise an approved boundary and raise a proposed change. Use the Stakeholder Map to identify who owns the work, who challenges it and who experiences its effects. Keep a Decision Record for each grant, change, renewal or withdrawal of authority.

Chapter 15, The Answer for Tom, asks what strategy Maya can now put before the board, and which priorities the organisation has earned the capacity to pursue.