12 min read
Reader Signal
Lower current readership momentum right now.
DANE
Monday, December 1, 2025 – 8:15 AM – The Warehouse
Dane had the narration schedule on the whiteboard. Six weeks, six veterans, color-coded by domain. Monday through Friday, three hours every morning. Each session recorded, AI-transcribed, structured into what he called “knowledge objects” — testable assertions about how the business worked.
“The routing module checks carrier availability three times,” Harry said. He was at the front of the room, whiteboard marker in hand, drawing boxes and arrows. “First check is live. Second check is from cache. Third check is a fallback to yesterday’s data.”
“Why three?” Maya asked. She had two screens open — the AI transcription on the left, a code scaffold on the right.
“Because the first check sometimes returned stale data. Someone added a second check as a workaround. Then a third when the workaround didn’t always work.”
“So three layers of duct tape.”
“Three layers of duct tape that process three million shipments a day.” Harry set the marker down. “Each layer was added by a different developer, in a different decade, for a different problem. By the time anyone noticed, customers had built their own systems around the triple-check behavior.”
Dane was watching from the side, tracking the session on his laptop. The AI had already tagged the triple-check pattern as a “legacy dependency” and cross-referenced it with Harry’s earlier narration about customer integrations. It was building a dependency graph automatically — which legacy behaviors were load-bearing and which were scar tissue.
“Knowledge object logged,” Dane said. “The AI is flagging the triple-check as a migration risk. We’ll need a compatibility shim for customers who time their requests around it. Zara — is that on your customer map?”
Zara checked her annotated wall map. Red pins for pain points, blue for workarounds, green for features customers actually valued. “Nobody mentioned the triple-check. But two dispatchers described a ‘rhythm’ to how they submit quotes — some said they wait a beat after the first submission. That might be this.”
“Or it might be nothing,” Harry said.
“We’ll validate next customer call,” Zara said. “But we don’t block the build. Maya — can the new system support both patterns? Fast path for new clients, compatibility path for legacy?”
“Already in the spec.” Maya didn’t look up from her screen.
Dane added a tick mark to his session board. Three weeks in, and the pattern was running clean: veteran narrates, AI structures, engineer builds, product validates against customer reality. The loop was tightening.
Thursday, December 11, 2025 – 10:00 AM – The Warehouse
The domain download expanded across every subsystem.
Gloria spent two mornings explaining customer exception workflows. Seventeen different exception codes that meant roughly the same thing — shipment delayed, carrier issue, resolution pending. The codes had accumulated over decades, each one added by a different support manager for a different reason.
“Can we collapse them?” Maya asked.
“Into what?”
“Three categories. Carrier issue. Customer issue. System issue. With a detail field for specifics.”
Gloria pulled out her notebook. “The customers expect the seventeen codes. Their internal systems parse them.”
“How many customers parse them?”
Gloria paused. “Maybe ten. Twelve.”
“Out of four hundred. And the twelve — do they parse all seventeen, or just a handful?”
“A handful.” Gloria’s face shifted. She was doing the math Zara would have done. “You’re going to tell me we build for three categories and give the twelve a compatibility layer.”
“Zara’s rule,” Maya said. “Build for the majority. Accommodate the exceptions.”
“That’s not a bad rule.” Gloria closed her notebook. “It’s just not how we’ve ever worked.”
Gil Navarro explained routing optimization the next day. Carrier scoring formulas that balanced cost, speed, and reliability through a point system nobody fully understood. Magic numbers embedded in the code — weighting factors set by engineers who’d left in 2001, tuned by trial and error over decades.
“Can you explain the scoring weights?” Arun asked.
“Some of them.” Gil pointed at his notes. “The cost weight is 0.35. Speed is 0.25. Reliability is 0.40.”
“Those add up to one. Makes sense. But where did the numbers come from?”
“A carrier manager named Pete Alvarez. He tuned them by hand in 2003, based on customer complaints. Pete retired in 2011.”
“And nobody’s touched them since?”
“The market’s changed. Carrier speeds have changed. The numbers haven’t.”
Dane wrote on his board: Scoring weights frozen since 2003. Candidate for ML-driven rebalancing based on current carrier performance data.
“This is the kind of thing the narrations uncover,” he said to Robert when the CEO dropped by that afternoon. “Institutional assumptions baked into constant values. The veterans can tell you why the numbers were set. They can’t tell you if the numbers are still right. That’s where the AI and current data take over.”
ZARA
Wednesday, December 17, 2025 – 7:00 AM – Atlanta Express Shipping
Zara went back to the dispatch centers.
This time she brought her wall map — a printed version, scaled down to poster size. Every feature mapped against actual usage. Customer workflows drawn as paths, with red marks where dispatchers deviated from the system.
Renee looked at the poster. “That’s my job,” she said. “You drew my job.”
“Does it look right?”
Renee traced the main path with her finger. Quote → book → track → invoice. “That’s 90% of my day. But you’re missing the callbacks.”
“Callbacks?”
“When a shipment goes sideways, I have to call the customer. But I also have to call the carrier. And I have to call my manager. Three calls, in that order, every time. And I have to log each one in a different screen.”
“Why three screens?”
“Customer CRM is one system. Carrier management is another. Internal escalation is a third. They don’t connect.”
Zara drew a new line on the map. Exception → three calls → three systems → sixty minutes of overhead per incident.
“How many incidents per day?”
“Four, five. On a bad day, ten.”
“So you spend an hour a day — maybe three on bad days — making the same calls and logging them in three places.”
“Welcome to my life.”
Zara wrote in her Moleskine: Exception resolution. Single interface. Three-system merge. Top priority for customer-facing pain.
She drove to two more dispatch centers that week. Same pattern. Same pain. Dispatchers who’d adapted to the system’s failures with workarounds so ingrained they’d stopped noticing them.
Friday, December 19, 2025 – 2:00 PM – The Warehouse
“Exception resolution just moved to the top of the priority list,” Zara said.
The team was in the afternoon build session. Maya looked up from her screen.
“Above billing?”
“Above billing. I’ve been to three dispatch centers this week. They spend an hour a day — three on bad days — handling exceptions across three systems that don’t talk to each other. That’s more wasted time than slow invoicing.”
“Billing is a bigger revenue number,” Gloria said.
“Billing is a bigger revenue delay. But exception handling is where we lose customers. The dispatchers at Atlanta Express told me they’ve recommended Axiom to shippers who call during an exception. Because Axiom gives them one screen. We give them three.”
The room went quiet.
“They’re recommending our competitor,” Gloria said.
“Our dispatchers are recommending our competitor. Because our system makes their job harder during the moment the customer needs help most.”
Dane stood up. “Exception handling. New priority stack. Maya — can you and Kevin scope the unified exception interface by Monday?”
“If Harry can narrate the exception workflows tomorrow morning.”
“Harry?” Dane turned to him.
Harry was already reaching for his binder. “I’ve got the exception codes. All seventeen of them.”
“Gloria, you’re in that session too,” Zara said. “I need the customer perspective alongside the system perspective.”
“I’ll be there.”
MAYA
Friday, January 9, 2026 – 9:00 AM – The Warehouse
The first major integration test.
It had been five weeks since the first narration sessions. The team had built through a rhythm — Dane’s morning narrations feeding afternoon builds, Zara’s customer map reshaping priorities twice already, the veterans getting faster at narrating because they’d learned what the AI needed.
“Running full quote-to-booking flow,” Kevin announced. “Customer requests quote. System calculates price. Customer selects carrier. System books shipment.”
Green. Green. Green. Green.
“Four core steps passed. Moving to exception handling.”
This was the part that killed projects. Happy paths worked in demos. Edge cases killed them in production.
“Simulating carrier rejection. Carrier doesn’t serve this ZIP code.”
The system paused. Then:
Carrier rejected. Reason: Service area restriction. Alternative carriers found: 3. Presenting options to customer.
“That was Harry’s specification,” Maya said. “Carrier rejections happen in 8% of bookings. The alternative-finding logic came from his December 12 narration session.”
Kevin continued. Carrier capacity exceeded. Weather delay. Hazmat restriction. Detention time. Customer contract override.
Green. Green. Green. Green. Green.
Twenty-three edge cases. Twenty-three passes.
“That’s not possible,” Gil said. “The old system took six months to handle half these.”
“The old system was built by people who had to guess the domain. We built this from narrations.” Maya pointed at the screen. “Every green test maps to a specific knowledge object from a specific veteran. Harry’s routing rules. Gloria’s exception codes. Ruth’s hazmat logic. Gil’s carrier scoring. It’s your knowledge, codified.”
Harry was standing by the screen. He put his hand flat on the desk.
“Forty years,” he said. “Nobody ever built from what I told them.”
“The AI listens differently than a requirements analyst,” Dane said. “It doesn’t filter. It doesn’t summarize. It captures the full context — including the edge cases the analyst would have trimmed for clarity.”
Thursday, January 15, 2026 – 4:00 PM – The Warehouse
The domain knowledge model was Dane’s idea. Maya built it.
Every narration session had been recorded. Every knowledge object tagged and cross-referenced. Thousands of assertions about how Meridian’s business worked — from Harry’s routing rules to Ruth’s compliance logic to Gloria’s customer behavior patterns.
“What if we use the knowledge base as a reviewer?” Dane had said. “Not just as source material for building. As an automated validator.”
Maya pulled up a routing module Kevin had written the day before. Clean code, well-structured.
Review this code for domain correctness, she typed.
Thirty seconds.
Issues found: 3
1. Carrier scoring does not account for seasonal capacity variations. Peak season (November-December) requires 15% capacity buffer per Harry Thornton narration, December 15.
2. Hazmat routing restriction missing for lithium batteries via air freight. Reference: Ruth Washington narration, January 2.
3. Customer tier override not implemented for preferred carriers. See Gloria Reyes narration, December 28.
Kevin stared at the screen. “It caught three things I didn’t know about.”
“They’re domain requirements you’d never have found in documentation,” Dane said. “Because the documentation doesn’t exist. The knowledge was in people’s heads. Now it’s in the model.”
“So we can generate code and validate it against the domain model automatically,” Maya said. “The veterans become part of the build pipeline permanently, even when they’re not in the room.”
Harry read the output. Three bugs from three different veterans’ knowledge. All real. All things an engineer without domain context would have missed.
“This is what you meant,” he said to Dane. “When you said I’d shift from narrator to validator.”
“How does it feel?”
“Like being replaced and preserved at the same time.”
“That’s accurate.”
ROBERT
Monday, January 19, 2026 – 10:00 AM – Robert’s Office, Meridian Headquarters
Robert ran the monthly numbers. The board book was due Friday. He had two stories to tell, and only one of them was true.
The public story: Meridian’s revenue was flat. Axiom’s market share was growing. The stock had dropped 5% in two weeks. Analysts were writing their third round of obituaries. Three board members had taken calls from private equity firms.
The real story: twelve people in a warehouse had built 65% feature parity with the legacy system in ten weeks.
Robert couldn’t tell the real story. Not yet.
He drove to the warehouse that afternoon. Maya walked him through the dashboard.
“Quote system: complete. Booking: 80%. Tracking: 60%. Billing: 40%. Exception handling: 70%.”
“The board is getting nervous,” Robert said. “Axiom’s started S-1 prep work. They’re targeting Q2 2026. If they file early enough, the quiet period starts before we launch.”
“What does that mean for us?” Dane asked.
“Their customers get locked in with investor confidence before anyone sees what we’ve built. Every month they move closer to filing, our surprise gets less surprising.”
Zara looked up from her customer map. “Then we launch sooner.”
“Can we?”
Dane and Maya exchanged a look.
“We can accelerate the 60 features Zara prioritized,” Maya said. “But we’d have to defer some of the compatibility paths for edge-case customers.”
“Which customers?” Robert asked.
“The ones who rely on legacy integrations,” Dane said. “The Consolidated Bulks of the world. The ones who use all seventeen pricing tables.”
“How many customers is that?”
Zara checked her spreadsheet. “Twelve. Representing 8% of revenue.”
“And the other 92%?”
“The other 92% would have a better product than they’ve ever seen. Faster quotes, unified exception handling, real-time tracking. The stuff dispatchers have been asking for.”
Robert nodded. “Build for the 92%. We’ll negotiate with the 8% individually.”
He left. At the door, he turned back. “Nora Vasquez asked how you’re capturing what the veterans know. Scaling question.”
Dane pulled up the knowledge model dashboard on his laptop. “We’re building a domain knowledge base from the narration sessions. Structured knowledge objects that any new engineer can query. The AI validates code against it automatically.”
“Write that up. One page. I want to send it to her.”
Friday, January 30, 2026 – 5:00 PM – The Warehouse
End of month four.
Zara’s customer map covered an entire wall. Blue for features that mattered. Red for features that didn’t. Green for new capabilities customers had asked for and never received. The blue and green together formed the product roadmap. The red was everything they were choosing not to build.
Dane’s methodology board had evolved from a simple three-block schedule into a full pipeline diagram. Recording narrations → AI extraction → knowledge objects → build specs → code generation → domain validation → customer validation. Each step measured. Each loop tightening.
Maya’s dashboard showed 65% feature parity and accelerating. The narration sessions were getting shorter — not because the veterans had less to say, but because the AI was learning to ask sharper questions.
Harry packed up his laptop at 5:30. The warehouse was quiet. Sofia was still coding. Arun was reviewing test results. Gloria had left an hour ago.
“Goodnight, machine,” Harry said to nobody in particular.
“Goodnight, training data,” Sofia said without looking up.
Harry laughed. It was the first time an engineer had ribbed him in forty years. It meant something.
End of Chapter 5