Chapter 6: The Guardrails

The call comes at 7:45 AM on a Tuesday.

It’s Henderson’s. Not Greg Hollis, who sends the POs. Not the accounts payable person who sends the payments. It’s someone from their legal and compliance department whose name Sam has never heard before.

“Ms. Laurent? This is Paul Kim from Henderson’s vendor compliance team. I need to ask you some questions about your data handling practices.”

Sam’s stomach drops, though she doesn’t yet know why.

Paul Kim explains: Henderson’s is a publicly traded grocery chain. They have vendor data requirements. Specifically, they have a clause in their supplier agreement, Section 8.3, which Sam last renewed three years ago and has not read since, that states all vendor employees must follow Henderson’s Acceptable Data Use Policy when handling Henderson’s proprietary information. This includes order data, pricing, product mix, and promotional plans.

“We’ve become aware,” Paul Kim says, in the careful tone of someone reading from a script, “that your organization may have been inputting Henderson’s data into third-party AI tools without authorization.”

Sam’s mouth goes dry.

“Can you confirm or deny whether Henderson’s purchase order data, pricing information, or order volumes have been shared with external AI platforms?”

Sam cannot confirm or deny because she genuinely doesn’t know. She tells Paul Kim she’ll investigate and call back within 24 hours.

She hangs up. Her hands are shaking.

She doesn’t know the answer. She does know where to start. She’s known since February, since the photo on Lisa’s phone.

She calls Marcus and Lisa into the boardroom and closes the door.

“Henderson’s compliance just called,” she says. “Has anyone been putting Henderson’s data into AI tools?”

Lisa looks at Marcus. Marcus looks at the table.

“I did a trend analysis,” he says. “For Greg. It’s their own order history. They’re ordering fifteen percent more Amber this quarter. I think they’re building up for a holiday promotion.” He swallows. “I thought if Greg saw what we could do, you might…” He doesn’t finish. “It was just ChatGPT.”

“Did you send it to him?”

“Last Thursday.”

“Show me.”

Marcus opens his laptop and turns it around. The email to Greg Hollis is short and proud. Nobody asked him to send it. Under his signature is one more line: Here’s the full interactive breakdown I ran for you!

“Click the link,” Sam says.

He clicks.

It’s a ChatGPT share link. The whole conversation opens in a browser tab, readable by anyone who has the URL. Marcus scrolls, slower and slower. Six weeks of Henderson’s purchase orders, pasted in raw. Product codes. Volumes. Per-unit pricing, the pricing that’s under NDA. The addresses of their distribution centres. Lisa’s notes on Henderson’s volume discounts, copied straight out of her pricing sheet.

All of it typed into the consumer version of ChatGPT. No business account. No data processing agreement. And unless Marcus switched off a setting he’s never heard of, OpenAI can use all of it to train future models.

Lisa reads her own notes on the screen and closes her eyes.

Nobody says anything. The page just sits there, open to anyone with the link, the way it has been since last Thursday. Greg clicked it. Of course he did. And then he did what Henderson’s supplier agreement requires him to do. He forwarded it to compliance.

Sam’s first instinct is anger. It doesn’t last.

Most of this isn’t Marcus’s fault. Marcus was never told not to do this. Sam was supposed to tell him. Oscar asked her to, in his office, back in March. She wrote it down. Then Lisa called about Henderson’s report, and the note stayed in her notebook. And before that, in February, she read one of his ChatGPT summaries and told him it was good.

There is no AI policy at MapleCo. Oscar recommended one twice, and twice Sam decided it could wait.

Sam calls Oscar.

She waits for the “I told you so.” It doesn’t come. Instead, Oscar asks a practical question.

“How bad is the Henderson’s situation?”

“I don’t know. Their compliance guy wants to talk tomorrow. I don’t even know what to tell him.”

“First, have Marcus delete that share link. Right now, while we’re talking. Then tell him the truth. It was unauthorized use by an employee who wasn’t given guidance. It wasn’t malicious. You’re implementing a policy immediately. And you’re putting data controls in place.”

“Will that be enough?”

“I don’t know. But lying is worse. Henderson’s is a public company. They’ve seen this before. They know every supplier has employees using AI tools. What they want to know is whether you’re aware of it and whether you’re doing something about it.”

But the Henderson’s situation is only half the crisis.

The other half arrives that same afternoon. Sam gets an email from a customer, a small restaurant chain in Niagara, saying they received an invoice dated for this month with last month’s pricing. The old, lower pricing. The new system posted the invoice when the order shipped, at the price on the master list. But the master list didn’t have the price increase Sam agreed to over the phone with the chain’s owner two weeks ago. She mentioned it to Lisa in the hallway. Since then, Lisa has been correcting the price by hand on every one of their orders, with a reason and her name attached each time. This one came in on a Saturday and shipped Monday at seven, before Lisa got to her desk. The automation didn’t know about the increase, because the new price lived in a phone call, then in Lisa’s head, and nowhere else.

It’s the Tofino deposit all over again, Sam realizes. Agree to something out loud. Write it down later. Later never comes.

Sam issues a corrected invoice. The dollar amount is small. The chain’s owner doesn’t call back.

At 6:30, Marcus knocks on her door frame. He’s holding a stack of paper.

He sets it on her desk without a word. It’s the list. Every tool he’s used since January. Every dataset he’s pasted, with dates. The share link, marked deleted, 9:04 AM. Four pages, single-spaced.

“Nobody asked me for that,” he says. “I figured somebody would.” He takes a breath. “And if it helps with Henderson’s, you can tell them the person responsible doesn’t work here anymore. I’ll write the letter tonight.”

Sam reads the first page. Then the second. Then she pushes the stack back across the desk.

“You’re not resigning.”

“Sam…”

“You did something dumb with a tool nobody told you how to use. Telling you was my job, and I didn’t do it.” She taps the list. “You’re the only person in this building who knows exactly what we’ve put where. So you’re going to own the approved tool list.”

Marcus stands there for a second, like he’s waiting for the catch.

“Go home, Marcus.”

He goes.

That night, Sam sits in her office long after everyone has gone home.

How many of these failures are the system’s, she wonders, and how many are mine?

She picks up the phone. It’s 9 PM. She calls Oscar anyway.

He picks up on the second ring. She tells him about Niagara. He listens to all of it.

“That’s four this month,” he says. “Can you separate them?”

Sam thinks. “Marcus’s ChatGPT thing. That’s governance.”

“Good. What else?”

“The pricing error. We have three price lists that don’t agree. You told me to fix that and I said it could wait. That’s… data integrity? The system exposed a problem that was already there.”

“Good. What else?”

“Nobody caught the pricing error before it went to the customer. We need monitoring. Sanity checks.”

“Good. And the fourth one?”

Sam is quiet for a long time.

“The fourth one is me,” she says. “I was processing Henderson’s orders by hand because I didn’t trust the system. I caused the error. The system had it right. I overrode it.”

Oscar is quiet for a moment. “That’s the only one I can’t build for you.”

Sam stares at the ceiling. Out on the dock, the night crew is loading the last truck.

The next morning, Sam calls Paul Kim back, the way she promised.

She tells the truth. An employee, without malicious intent, used a consumer AI tool to analyze Henderson’s order data, then shared a link to it with Greg. The employee was not given guidance because MapleCo did not have an AI policy. That is MapleCo’s failure, not his. The link is deleted. She sends Paul Kim Marcus’s list: every tool, every dataset, every date, compiled before anyone asked for it. A one-page AI policy will be on his desk within two weeks, with the controls behind it.

Then she answers the rest of his question. Yes, Henderson’s purchase orders go through AI at MapleCo every day: the PO system reads them. It runs on a business account under a data agreement, and Oscar has it documented: which provider sees what, where it’s stored, and who can read the logs.

Paul Kim listens. He asks a few questions. Then he says: “Ms. Laurent, I appreciate your transparency. Most suppliers we flag on this either deny it or scramble to hire a lawyer.”

Sam waits for the but.

“But deleting a link doesn’t prove nobody kept a copy, and transparency doesn’t undo the exposure. We’re placing MapleCo on a 90-day vendor probation. During that period, we’ll be monitoring your data handling practices. We’ll need a copy of your AI policy, the documentation for that PO system, a signed attestation from every employee who handles Henderson’s data, and a quarterly compliance report. If everything checks out at 90 days, you return to standard vendor status.”

“And if it doesn’t?”

“Then we have a different conversation.”

Henderson’s also cuts its order volume by 15% for the length of the probation. Not as punishment, Paul Kim explains. As risk management. They’re diversifying supply until they’re satisfied MapleCo’s controls are in place.

Sam hangs up. The account isn’t lost. But it’s wounded. Fifteen percent of Henderson’s is real money, and for the next three months every order carries extra weight.

Greg’s end-of-June deadline is a week away, and her numbers are finally moving. Now there’s a second clock. This one is ninety days long, and it has paperwork.

Over the next two weeks, Sam and Oscar build three things.

The AI Policy. It’s one page. It says: use only the AI tools and accounts on the approved list, only with the data the list allows, and only for the purposes it allows. The list names each tool, the account, what data it may see, and who approved it. The PO system is on it, under its data agreement. Marcus keeps the list. Sam approves what goes on it and reviews it every quarter. When in doubt, ask before pasting.

Sam shows it to Lisa. Lisa says, “This should have existed six months ago.” Sam agrees.

She shows it to Marcus. He reads it twice, then asks to be the one who walks the team through it. Sam says yes. It’s his list now.

Then she calls Dave. A one-page policy is a promise. Dave makes it harder to break. By Friday, MapleCo has a business ChatGPT workspace whose terms keep its data out of model training, and everyone signs in to it with their company account. Other AI sites are blocked on the office network and on every company laptop. The one thing Dave can’t do is block personal ChatGPT without blocking the business workspace, because they’re the same website. Personal accounts are the policy’s job.

“Now this,” Dave says, “is my trade.”

Oscar takes the other half: what data his automations touch, where it’s stored, and who can read the logs. “Marcus keeps the list,” he says. “Dave locks the doors. I watch the pipes.”

Oscar adds one more principle. “Your policy covers data. It should also say where the line is.”

“Routine, repetitive work is safe to automate with monitoring. Reading a PO, matching it against the database, sending a tracking number. The system does the work, and a person reviews the exceptions. Pricing exceptions, contract-sensitive communication, compliance reporting, anything where a wrong answer could damage a customer relationship or create legal exposure: that stays with a person. The system prepares. A person decides.”

“So if Henderson’s asks for a custom pricing agreement…”

“The system pulls the numbers and drafts the comparison. You make the call, and you send the email.”

Sam adds it to the policy. One paragraph. It takes five minutes.

The Single Source of Truth. This one takes a full week, and most of it happens at the boardroom table.

Oscar puts the three price lists on the big screen side by side: the master Google Sheet, the price list in QuickBooks, and Henderson’s contract rates in US dollars. Lisa brings a highlighter and a coffee the size of a flower vase. Sam brings the guilt.

In the first hour, they find forty-one products where at least two of the three disagree.

Some are easy. A price increase from February that made it into the Google Sheet and never reached QuickBooks. A gift basket MapleCo stopped making two years ago, still on the list, still with a price, still occasionally ordered.

Some aren’t. Three restaurant customers are on rates that exist in neither list. Lisa knows all three by heart. Sam agreed to one of them at a trade show in Toronto and wrote it on a napkin.

“Is the napkin still around?” Oscar asks.

“The napkin,” Lisa says, “is me.”

They go line by line. Every disagreement gets the same three questions: which number is right, who decided, and where is it written down? When the answer to the third question is “Lisa,” it goes into the database with a note saying who agreed to it, when it starts, and when it ends.

By Friday there’s one price for every product and every customer, with Henderson’s in US dollars at the contract rate. The automation prices every order from the database, and every invoice it posts to QuickBooks carries the same number. When a price changes, it changes once, in one place, and everything downstream updates automatically.

It comes with one rule, which Sam writes on the boardroom whiteboard herself: No price, discount, or deposit exists until it’s in the database. Not a phone call. Not a hallway conversation. Not a napkin. Not a promise made from a campground in Tofino.

The Monitoring Layer. Oscar splits it into two jobs. The first is validation: every price on every order and every invoice has to match the database, or it doesn’t go out. The second is watching for patterns nobody would spot one order at a time. A new customer gets flagged. A product code that isn’t in the catalogue gets flagged. So does an invoice total far from a customer’s usual, and so does anyone changing the same customer’s price by hand, order after order.

“The usual-total check wouldn’t have caught Niagara,” Oscar says. “Last month’s lower price is their usual. What should have caught it was Lisa changing their price on every order for two weeks. The system logged every one of those changes, and nothing alerted on them. That one’s mine.”

“No check finds a price nobody wrote down,” he adds. “The rule on the whiteboard does that. This just tells us when someone’s working around it.”

Lisa reviews the flags each morning. It takes ten minutes.

That weekend, Sam sits in the stands at Maya’s summer hockey camp.

She doesn’t bring her laptop. She does bring her phone, and she checks it four times before the first drill is over.

It buzzes. An automated notification from the PO system: “Saturday batch: 3 POs received, 3 processed, 0 exceptions.”

She reads it, smiles, and puts the phone back in her pocket.

Two minutes later, she takes it out and reads it again. Nothing has changed. She knows nothing has changed. She checks anyway.