Contents

Chapter 13

The Rules That Already Apply

At 10.17 on Thursday morning, Tom turned the last page of Maya’s renewal note.

The service-credit retrieval feature had been allowed to continue for a small group, under a narrower boundary. It could help advisers find current procedure material. It could not approve a credit or draft an outcome for approval and eligibility questions. Its sources had an owner, its use had an expiry, and the ordinary service route remained available.

Tom set the paper down.

“And what does the law require of us now?”

Nobody answered at once.

Daniel knew how the feature was configured. Leila knew which procedure governed the service-credit work. The CFO knew what the use cost. Maya knew what she had authorised. None of those answers settled Tom’s question.

“We need to examine it against the uses we have,” Maya said. “I’ll bring you back an answer we can support.”

After the meeting, she asked Nadia to lead the first pass. Nadia was to work with the people who owned each use, identify questions the team could answer from its records, and mark those that needed specialist advice.

She did not ask for a list of laws that contained the word AI.

Too many open tabs

At 8.36 the next morning, Nadia’s screen was filled with open pages.

She had started with broad searches. Privacy. Customer decisions. Employment. Records. Contracts. New rules about AI in different places. Every page led to another. Some discussed systems that made decisions automatically. Some discussed tools used only to draft or summarise. Some were proposals. Others described existing obligations. A page could be relevant in one place and misleading in another.

She printed a list of questions and took it to Maya.

“It is too wide,” she said. “If I write ‘check privacy, consumer, employment and AI law’ beside every register entry, we will learn nothing.”

Maya looked at the list. It was thorough in the way a list can be thorough without helping anyone act.

“What did we do before the retrieval feature?” she asked.

“We still had to handle customer records properly. We still had to make and explain service-credit decisions. We still needed an accurate case record.”

“Start there.”

Nadia closed most of the tabs. She opened the register and took the service-credit entry first. The organisation had an existing customer process. The feature changed how an adviser might find an instruction inside that process. It did not create the first need to know who could approve a credit, what information an adviser could use, or what should be recorded when a decision was reversed.

That was the useful reframe. The team did not need to begin by asking whether a new rule specifically named the technology. It needed to ask what the organisation was doing, to whom, with whose information and with what consequence. Then it could examine the rules and commitments already relevant to that work, and ask what the new feature changed.

The obligation follows the work and its effect, not the label on the software.

Nadia did not conclude that no AI-specific rule could matter. She concluded that the scan needed a starting point small enough to test.

One use at a time

On Monday, Nadia brought Leila and Daniel together with the service-credit register entry.

At the top she wrote the authorised use in one sentence: a small group of advisers could use the retrieval feature to locate and check service-credit procedures while retaining the ordinary case, referral and approval steps.

Beneath it she put four questions.

What information entered the feature? What did its output influence? Who made the decision? What evidence would show that the organisation had followed its own process?

Daniel could describe the inputs allowed by the configuration and staff guidance. Advisers could ask general procedure questions. They were not to paste customer names, account identifiers or full correspondence into the question box. Customer facts remained in the case system. Daniel could also explain access, logs and the source collection. He could not establish from configuration alone that every adviser had followed the input rule.

Leila could describe the decision. An adviser established the case facts and consulted the current procedure. A team lead approved where the procedure required it. The feature could show source material, but it could not determine entitlement. The case record should show what happened, including a reversal and its approval.

Nadia wrote down the evidence available: the source owner and version, the release and renewal decisions, access checks, training, selected output reviews, case records and the incident trail from the source-change pause. Each answered a different part of the question. None was a certificate that every case had been correct.

She then wrote what would change the scan. If advisers began entering identifiable customer histories, the information question would change. If the feature drafted customer messages, the communication question would change. If a proposed configuration gave its output effect as a credit decision, the authority question would change. If the use moved to another country or group of customers, the applicable rules and commitments might change.

Which proposed change would make you revisit the questions for your current use?

Guidance

Describe one change to the work, such as different inputs or a stronger role for the output. Trace what the change could affect and who would need to examine it before proceeding. Record the question for advice if it exceeds your evidence or authority.

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

“That last one needs someone who knows the place,” Leila said.

“Yes,” Nadia replied. “This sheet should show us when we need that person.”

The Legal and Regulatory Scan began to take shape around actual register entries. It did not answer every legal question. It made the questions visible in relation to a defined use.

FIG-13.1

First pass on the service-credit entry

Scan questionWhat the team can establish nowWhat would change the answer
What information is used?General procedure questions and an approved source collection are in scope; customer case facts stay in the case systemIdentifiable case material entered into the feature or a new data route
What can the output affect?It helps an adviser find and check guidance; ordinary approval remains with peopleOutput begins to recommend, determine or apply a credit
What record supports a decision?Case record, procedure version, approval and any reversal should be traceableDecisions rely on saved answers instead of the governing source and case facts
Who and where are affected?The current small service group and its customers under the authorised processA new audience, business unit, country or category of customer
Source: author’s method and fictional case.

Nadia asked Leila whether the case record always made reversals traceable now. Leila did not claim that it did. The earlier diagnosis had found missing reversal records. The organisation had changed the recording step, but she still reviewed it through ordinary service work.

The scan preserved that gap. A tool did not become compliant because it was kept out of the case system. The customer decision and its record remained part of the work.

A quieter entry

The next register entry looked less demanding.

An approved summarising feature helped staff make short notes from internal meetings. Its boundary excluded customer records and personnel matters. A person checked the summary against their own notes before using it. It did not send the summary outside the team or make a decision for anyone.

Nadia had initially placed the same long list of possible legal topics beside this entry as she had beside service credits. Daniel asked what decision that list would change.

They ran the same questions instead.

The input was a permitted internal meeting note. The output was a draft summary. The person who attended the meeting checked it, corrected any missing or mistaken action, and decided what to keep in the ordinary meeting record. The feature had no authority to approve an action, assign responsibility or send a commitment to someone outside the room.

That did not mean the entry needed no care. A staff member might paste material outside the approved boundary. A summary might wrongly attribute a decision. If the use expanded to personnel discussions, confidential customer work or automatic distribution, the scan would need to be revisited.

For the current use, Nadia could name the owner, boundary and check without arranging an elaborate review of every note. She marked the main open question for the next ordinary register review rather than calling an emergency meeting.

Leila noticed the difference.

“We are not giving it a free pass because it sounds routine,” she said.

“No. We are asking what it can actually affect.”

The service-credit feature sat beside a customer decision and an approval route. The meeting summary was a draft that a participant checked before it became a record. Both required an owner and a boundary. The amount and kind of examination followed the use.

Scan every entry. Spend the most effort where the output can change someone’s position.

Maya read both entries that afternoon. The contrast helped her more than a long general memo. She could see why the same technology might demand different questions in different settings.

The question they could not settle

Nadia turned next to the tender tool, whose trial remained closed.

The sales director wanted to use it to help prepare responses to customer requests. He proposed starting with documents from a prospective customer operating in another country. The documents included commercial terms and information supplied for a limited purpose.

Daniel could describe where the proposed tool would process the material and what settings he could inspect. The sales director could explain why the documents were useful for a tender response. Neither could say whether the customer agreement permitted that material to be used in the proposed way, or which requirements would apply across the places involved.

Nadia did not mark the use “prohibited”. She also did not mark it “approved with care”. The trial was closed, and the unanswered questions were material to whether it could open.

She used When to Seek External Help to state the referral narrowly. The team needed appropriate advice on the customer documents, the relevant agreement, the proposed data route and the places involved before any material was placed in the tool. They would provide the actual terms and proposed workflow to the specialist, rather than ask an abstract question about whether AI could be used for tenders.

What would a specialist need to understand the question you cannot settle?

Guidance

Describe the proposed workflow and the material involved before framing the question. Identify the agreement or other evidence the specialist would need through an appropriate route. Keep any material gap visible rather than treating the referral itself as permission to proceed.

The sales director was impatient.

“Could we just use a few pages without names?”

“That might change one question,” Nadia said. “It does not tell us what the agreement permits.”

He asked whether the referral meant the idea was dead.

“No. It means the next decision needs evidence we do not have.”

Maya kept the tender trial closed while the question was examined. The Decision Record stated what had been requested, why the current boundary did not cover it, what information the adviser would need, and who could make a fresh decision after the advice.

That was the method working. The team had reached the edge of its knowledge before moving customer material.

The same approach would apply if someone proposed using an AI output for an employment decision or another use that could materially affect a person’s rights or opportunities. The team could describe the workflow, data and proposed authority. It should not improvise a legal conclusion where specialist judgement was needed and a wrong decision might be difficult to repair.

An answer that belongs in the record

Tom asked for an update later that week. Maya brought Nadia’s first scan, not the stack of open pages.

For the renewed service-credit use, they could describe the existing customer decision process, information boundary, approval route and record evidence. They had identified where the retrieval feature changed the work: it introduced a new way to find and phrase procedure material, which required source control, checking and a record of authority. They had not found a reason to treat an answer from the feature as a separate permission to make a credit decision.

For the internal summarising use, the scan showed a smaller effect and a simpler check. The meeting participant remained responsible for the note that entered the record.

For the tender proposal, the team had a question it could not settle from internal assumptions. The trial remained closed pending specialist advice on the actual material and route.

Tom looked at the pages.

“So what is the legal position?” he asked.

“We have a method for finding the position for each use,” Maya said. “We have evidence for some parts. We have referred the part we cannot settle. A single answer for every use would mislead you.”

Tom did not look pleased. He also did not ask her to replace the pages with a traffic-light chart.

Nadia explained why she had resisted one broad statement. The technology appeared in work that already had rules, agreements and responsibilities. The result depended on the information used, the people affected, the place, and whether the output merely helped someone or took effect as a decision. Change any of those and the answer could change.

Maya added that the Decision Record would capture what the organisation decided after each scan. The scan identified obligations and open questions. The decision record stated the authorised use, conditions, evidence and owner. An unresolved material question did not become resolved because the team filed it.

Tom asked when they would finish scanning the register.

“The live entries first,” Maya said. “Proposed uses before they cross their boundary. Then on a schedule and whenever the use changes.”

That gave him a programme of work rather than a promise of one definitive memo.

The horizon that moves

Daniel joined the next scan meeting with a concern.

Even if the team examined every current register entry, rules and guidance could change. So could contracts, product behaviour and the countries in which the organisation worked. A scan completed once might become another old procedure with a convincing date on its cover.

Nadia agreed. She added an owner and cadence to the Legal and Regulatory Scan, and owners of individual uses would report changes in audience, data, location, purpose or authority mode when they proposed them.

They also identified events that would trigger an earlier look: a proposed use of a new kind of data; a move from Assist to Recommend, Determine or Act; a customer or employee group newly affected; work in another country; a changed contract term; a material complaint or incident; or a credible change in rules relevant to an existing use.

Daniel asked who would watch the moving horizon. Nadia would coordinate it, but she would not pretend to monitor every possible rule in every place alone. The organisation would use the appropriate sources and advisers for the places and activities actually in scope. If a proposed expansion introduced a new place, the expansion decision would include that question before the use began.

Maya wanted the output of each review to be short enough to act on. Nadia proposed three lines for each entry: what had changed, what it might affect, and whether the current authorisation could continue while the question was resolved. A specialist’s advice could be attached where needed. The Decision Record would show any resulting change in scope or conditions.

A scan is a recurring way to notice a changed answer, not a certificate that nothing can change.

The CFO asked whether this was another continuing cost. Maya said yes. It belonged in the whole-cost view for uses that required it. The work would be shared where several uses raised the same question, but it would not be treated as free because it happened in a meeting rather than on an invoice.

The small-claims question

The small-claims proposal came up during the register review. It was still an investigation, not an authorised system. Someone had suggested that a future tool might decline a claim when it appeared not to meet the rules.

Nadia circled the word decline.

A system that helped staff find a claims procedure would pose one set of questions. A system whose output took effect as a refusal would pose another. The difference was not whether both systems used AI. It was the authority and consequence of the output.

They did not decide the proposal that day. Maya asked the proposal owner to describe the actual decision path: whose claim, what information, what rule, what record, which person could review it, and what would happen after a decline. The legal scan would then identify the existing obligations and the questions needing specialist advice before anyone asked for that authority.

Leila pointed to the service-credit work.

“We learned this with a credit,” she said. “Finding a rule and deciding a case are different jobs.”

Maya agreed. She would not let the service-credit feature’s renewal become evidence that a different system could decline a claim. The register would keep those uses separate.

The standing customer panel could help the organisation understand how a process appeared to people affected by it. It could not decide the legal position or approve the proposal. That distinction, too, belonged in the standing arrangement the team had yet to establish.

A question with an owner

At 4.28, Nadia closed the scan for the day. It was not complete in the sense that every possible future rule had been answered. The live entries had named owners, evidence and open questions. The tender proposal had a specialist referral. The small-claims investigation had a clear request for a decision path before anyone could scan it properly.

Maya returned to Tom’s original question.

“What does the law require of us now?” was a reasonable thing for a chair to ask. It became less useful when detached from an action. The service-credit adviser finding a procedure, the meeting participant correcting a summary, and a future system declining a claim did not present the same question merely because each might use an AI feature.

Nadia filed the first Legal and Regulatory Scan with the register. She scheduled its next review and recorded the events that could bring it forward. Maya attached the decisions and referrals to their respective uses.

Tom would not receive a declaration that the organisation was compliant with everything. He would receive a view of what had been examined, what evidence supported it, what remained open, and which uses could proceed under their present authority.

That was an answer the team could maintain.

Run a Legal and Regulatory Scan against each real use in your register. Record the information, audience, decision effect, existing obligations, evidence and conditions that would change the answer. Keep the resulting authority and open questions in the Decision Record. Use When to Seek External Help when a contract, place, data route or consequential decision outgrows the team’s expertise.

Chapter 14, The Standing Arrangement, asks how these entry-by-entry decisions become an enduring way to authorise, review and stop uses across the organisation.