17 min read
Reader Signal
Lower current readership momentum right now.
ZARA
Monday, November 3, 2025 – 6:45 AM – Atlanta Express Shipping, Dispatch Center
Zara had been on the job four days and she’d spent three of them outside the warehouse.
The Atlanta Express dispatch center smelled like burnt coffee and adhesive labels. Twelve dispatchers in rolling chairs, four screens each, phones cradled between shoulder and ear. The oldest dispatcher, a woman named Renee, had a laminated cheat sheet taped to her monitor. Faded. Coffee-stained. The kind of thing you make when the system is too slow to remember for you.
“Walk me through your morning,” Zara said.
Renee didn’t look up from her screen. “I pull up exceptions from overnight. Usually four or five. Most resolve themselves — delayed scan at a hub, carrier hiccup, nothing real. I clear them manually because the system can’t tell the difference between a real exception and a false alarm.”
“Then what?”
“Rates. I run thirty, forty quotes. Origin, destination, weight, class. Basic stuff.”
“What about the advanced rate tools? Multi-stop optimization? The contract comparison module?”
Renee turned around. “The what?”
Zara wrote in her Moleskine. She’d heard this before, twice already this week. Atlanta Express was the third dispatch operation she’d visited. The pattern was identical each time.
“How do you actually book a shipment?” Zara asked.
“I go to the portal. Origin, destination, weight. Then I wait. Thirty seconds, sometimes forty, for the rate to come back. By then the customer is already on the phone asking why it’s taking so long.”
“Show me.”
Renee typed. Waited. The screen churned. Zara counted. Thirty-seven seconds.
“And Axiom?”
“Two seconds. My friend dispatches for a company that uses Axiom. Two seconds.” Renee shrugged. “I don’t make the purchasing decisions. But I know which system I’d pick.”
“How many clicks to book a shipment?”
“Eleven.”
“You’ve counted?”
“When you do it forty times a day, you count.”
Zara spent the rest of the morning at Renee’s shoulder. Watched her work. Watched her route around the system — calling carriers directly because the portal’s carrier selection was three versions behind reality. Watched her copy tracking numbers from one screen and paste them into a spreadsheet because the integrated tracking page loaded too slowly.
By noon, Zara had filled nine pages in her Moleskine.
Monday, November 3, 2025 – 2:00 PM – The Warehouse
Dane was at the main whiteboard when Zara walked in. He’d been there since 7 AM. The warehouse was starting to look like a workspace — desks pushed together in clusters, whiteboards unwrapped and mounted, power strips running along the floor like veins.
Harry Thornton sat at a folding table near the window, a three-ring binder open in front of him. Three inches thick, rubber bands holding it together. He’d been through it with Gil Navarro and Ruth Washington all morning, flagging sections, cross-referencing carrier codes.
“How was the field?” Dane asked without looking up.
“Eleven clicks to book a shipment,” Zara said. “Thirty-seven seconds for a rate quote. Dispatchers routing around the system because it’s slower than a phone call.”
“How many features do they use?”
“Three. The basic quote screen. The tracking page. Invoice download.”
“Out of how many?”
“I don’t know yet. That’s what I need Harry for.” Zara turned to the folding table. “Harry. How many features does the legacy system have?”
Harry looked up from his binder. “Define ‘feature.’“
“Something a customer can do. A screen, a tool, a module. Something with a menu item or an entry point.”
Harry exchanged a look with Gil. “Rough count? Three-forty. Maybe three-fifty.”
“And how many of those do dispatch customers actually use?”
“Depends on the customer.”
“The top twenty. The ones that generate 80% of your revenue.”
Harry was quiet for a moment. “Maybe sixty. Seventy.”
Zara wrote 340 on the whiteboard. Drew a circle around it. Then wrote 60 below it.
“Eighty percent dead weight,” she said.
“That’s not fair,” Harry said. “Some of those features exist for compliance. For edge cases. For customers with specific—”
“Tell me about the seven-tier pricing model.”
Harry’s mouth closed.
“Gloria mentioned it this morning,” Zara said. “Seven contract pricing tiers. Built for one customer. How many customers use more than two tiers?”
“One,” Gloria said from across the room. “Consolidated Bulk.”
“And what percentage of your support tickets come from Consolidated Bulk?”
Gloria pulled up a spreadsheet she’d brought from her last day in the office. “Thirty-nine percent of all carrier overrides. Twenty-two percent of all support escalations.”
Zara circled the number 60 on the whiteboard. “This is our product. Not 340 features. Sixty. The sixty that matter to the customers who generate real revenue. We build those so well that nobody remembers the other 280 existed.”
Harry stared at the board. Forty years of code. Features he’d maintained, debugged, patched on weekends. Zara had reduced it to a number and drawn a circle.
“Some of those 280 features—” he started.
“Feed the machine, Harry.” Dane said it from the whiteboard, calm. “Tell us everything about those 280. Tell us why they exist, what customers asked for them, which ones actually fire in production. That’s how we make sure we don’t cut something load-bearing. But the default is cut. Zara adds things back when a customer needs them.”
Harry looked at Dane. Then at the binder in front of him. Three inches of institutional memory, and the product person had been on the job four days and already knew more about customer behavior than the documentation he’d been compiling for months.
“The binder’s yours,” Harry said to Zara. “Every SOP I could find. Some of it contradicts itself.”
“Good. Every contradiction is a conversation.” Zara picked up the binder. Heavy. She set it on the desk next to her Moleskine. “I’m going to go through this page by page. Every time something doesn’t match what I saw in the dispatch centers, I’m going to ask you why.”
“That’ll take weeks.”
“Then we start now.”
DANE
Tuesday, November 4, 2025 – 7:30 AM – The Warehouse
Dane had the daily rhythm drawn on the whiteboard before anyone else arrived.
Three blocks. Color-coded. He’d stolen the marker from Ruth Washington’s desk and she’d already told him she wanted it back.
Morning (8-12): Domain narration. Veterans talk. Engineers listen. AI records. Each session focuses on one customer workflow from Zara’s list. The veteran describes how it actually works — not the documentation, not the process map, the real thing. The engineer asks questions. The AI transcribes and structures the output into testable specifications.
Afternoon (1-5): Build. Engineers code from the morning’s specifications. AI generates first drafts. Engineers review, test, iterate. Veterans are available for questions but don’t drive.
Evening (5-7, optional): Integration. Both groups at the same whiteboard. What broke. What surprised. What the AI got wrong and why.
“The morning sessions are the critical path,” Dane told the group when they assembled. “We’re not doing requirements gathering. We’re not writing PRD documents. We’re doing something different.”
He drew a pipeline on the whiteboard.
Domain expert narrates → AI structures into specs → Engineer builds from specs → AI generates code → Engineer reviews → Domain expert validates
“The veterans are the source material. The AI is the translator. The engineers are the editors. Nobody works alone. Nobody works in sequence. It’s a loop.”
Sofia raised her hand. “How is that different from pair programming?”
“Pair programming is two engineers. This is a domain expert, an AI, and an engineer in a feedback loop. The domain expert doesn’t need to write code. The engineer doesn’t need to understand forty years of freight logistics. The AI bridges the gap.”
“And if the AI gets it wrong?”
“It will get it wrong. Constantly. That’s the point of the loop. Harry describes how hazmat routing works. The AI generates a spec. Maya codes from the spec. Harry reviews the output and says ‘no, lithium battery weight limits change depending on ground versus air.’ Maya updates. The AI learns the correction. Next time, it gets it right.”
Harry was watching from his folding table. “You’re describing me as a training dataset.”
“I’m describing you as the most valuable training dataset this company has.” Dane turned to face him. “Forty years of edge cases. Customer quirks. The reasons behind every strange piece of code. No consultant could extract that in eight months. But if we structure the process right, the AI can learn it in weeks. Not all of it. But enough to build from.”
“And when the AI has learned enough?”
“Then you’re the reviewer, not the narrator. You check its work instead of teaching it from scratch. Your job shifts from source to validator.”
Harry considered this. It was honest. More honest than Robert’s pitch about being “full partners.” Dane was saying: your knowledge is the raw material. We’re going to process it, extract it, encode it. And then you’re going to quality-check the output.
“That’s a better description of what I’m actually here for,” Harry said. “Than the version where I’m leading anything.”
“You’re leading the domain narrations,” Dane said. “Nobody else can. But the architecture, the methodology, the team structure — that’s my job. And what we build — that’s Zara’s call, based on what customers need.”
Tuesday, November 4, 2025 – 10:00 AM – The Warehouse
The first domain narration session was rough.
Harry started with the quote system. Logical choice — it was the workflow Zara had flagged as the highest-priority customer pain point. Thirty-seven seconds for a rate that Axiom delivered in two.
“The pricing engine checks seventeen different pricing tables,” Harry said. “In sequence. Contract pricing first, then standard rates, then fuel surcharges, then accessorials, then—”
“Stop,” Maya said. She was recording on her laptop, the AI transcription running in a side window. “Seventeen tables. Why seventeen?”
“Different pricing structures for different customer segments. Enterprise, mid-market, SMB. Then carrier-specific rates. Then commodity-specific rates—”
“How many of those seventeen tables are actually used for a standard quote? Not edge cases. A dispatcher running thirty quotes a day.”
Harry paused. “Three. Maybe four.”
“So thirteen tables are checked every time, and they return nothing for the majority of quotes?”
“The system doesn’t know in advance which tables will match. It checks them all.”
Maya looked at her screen. “That’s your thirty-seven seconds. The system is doing fourteen unnecessary lookups per quote.”
“The Consolidated Bulk contract uses all seventeen,” Harry said quickly.
“One customer,” Zara said from across the room. She was reading the binder. “One customer uses all seventeen tables. Four hundred and twelve other customers use three.”
The room was quiet. Sofia was typing something. Arun had his head cocked, working through the architecture.
“So we build for three tables,” Maya said. “Fast path. Two-second response. And if the customer happens to be Consolidated Bulk, we hit the extended path.”
“The extended path will still be slow,” Harry said.
“Will it be thirty-seven seconds slow?”
“No. Maybe five. Six.”
“Then Consolidated Bulk gets six-second quotes and everyone else gets two. That’s the product decision.” Maya looked at Zara. “Right?”
“That’s the product decision,” Zara confirmed. “Build for the 412. Accommodate the one.”
Dane watched the exchange from the side of the room. This was the loop working. Zara defined the priority (quote speed). Harry narrated the system (seventeen tables). Maya found the architectural insight (fourteen are unnecessary for most customers). Zara made the product call (build for the majority). Nobody was waiting for a requirements document to be approved. The decision happened in the room, in real time.
He wrote on his whiteboard: Day 2. Loop working. Decision cycle: 4 minutes.
MAYA
Wednesday, November 5, 2025 – 2:00 PM – The Warehouse
The first week was chaos and then it wasn’t.
Monday and Tuesday had been all talk. By Wednesday afternoon, Maya was building. She’d taken the morning narration sessions — Harry on pricing, Gloria on carrier preferences, Gil on routing exceptions — and run them through the AI pipeline Dane had designed. The structured output was rough but usable. Specifications that read like a conversation transcript, annotated with edge cases and historical context.
“This is wild,” Sofia said, looking over Maya’s shoulder. “It’s like a spec that argues with itself.”
“That’s because Harry and Gloria don’t always agree. The AI captured both versions.” Maya highlighted a section. “Harry says the fuel surcharge is calculated monthly. Gloria says some customers are on quarterly recalculation. Both are right — it depends on the contract tier.”
“So which do we build?”
“Both. But the quarterly path is an exception handler, not the main flow. Dane’s API-first approach — the fast path handles 90% of cases. Exceptions get routed to a separate service.”
Sofia started coding. Within an hour, she had a rate calculation service running. Basic — it only handled the three core pricing tables. But it returned a quote in 1.8 seconds.
Maya called Harry over.
“Try it,” she said.
Harry typed in a test shipment. Atlanta to Dallas, 500 lbs, Class 70. The quote came back before he finished reaching for his coffee.
“That’s…” He picked up the coffee. Set it down. Picked it up again. “That’s fast.”
“1.8 seconds. And we haven’t optimized yet.”
“The fuel surcharge?”
“Placeholder. We need your Wednesday narration to get the real logic.”
“There’s a bug. The accessorial fee is wrong. It’s applying a liftgate charge to a shipment that doesn’t need one.”
Sofia leaned over. “Where’s the liftgate logic?”
“It triggers when the destination has no loading dock. But the system has no way to know that unless the customer tells us. And half the time they don’t. So we default to charging the liftgate and then issuing a credit later.”
“That’s a terrible customer experience.”
“I know. It’s been that way since 1998.” Harry pointed at the screen. “But if you want to fix it, you need the destination database. Every delivery address we’ve ever shipped to, flagged with dock status. It exists. It’s just never been connected to the quoting engine.”
“Where is it?”
“Building C. A server closet. Ruth knows the access credentials.”
Maya looked at Dane, who was listening from his desk. “Can I pull that data?”
“How big?”
“Two million addresses. Maybe three.”
Dane nodded. “Pull it tonight. Integrate tomorrow. We’ll validate with Gloria’s team in the Thursday narration.”
By Friday, the quote engine returned accurate rates in under two seconds, with correct accessorial charges, for the top three pricing tables. Five days. Five days to build what the old system had spent twelve years patching.
“Maya.” Harry was standing at her desk. End of day Friday. Everyone else was packing up.
“Yeah?”
“You built in a week what took us a year to maintain.”
“What I built handles sixty percent of use cases. You maintained a hundred percent for forty years. Those aren’t the same thing.”
“No. But it’s a start.” He put his reading glasses in his shirt pocket. “When do we tackle the carrier routing?”
“Monday narration. You and Gil. Bring everything.”
“I’ll bring the binder.”
“Bring the stories, Harry. The binder is just paper.”
Saturday, November 8, 2025 – 9:00 AM – Zara’s Hotel Room, Atlanta
Zara called Kwesi while Adaeze’s voice yelled in the background about something involving a stuffed elephant and jurisdiction.
“How’s the project?” Kwesi asked, referee-whistle energy in his voice.
“I spent two days riding with dispatchers. Nobody uses eighty percent of the features in the system. Eighty percent, Kwesi.”
“That’s a product.”
“That’s a graveyard. Features built for one customer that nobody removed. Features built for a report that got cancelled in 2009. The codebase is a museum.”
“But the dispatchers — they know what they need?”
“They know exactly what they need. They just can’t get it because the system is buried under forty years of accumulated cruft.” Zara opened her Moleskine. “I’ve got a wall-map started at the warehouse. What customers actually do versus what the system thinks they do. The gap is staggering.”
“You sound like the Loadstar days.”
“Better. At Loadstar we built for customers who didn’t exist yet. These customers exist. They’re using the system every day. They’re just routing around the worst parts of it.”
Adaeze screamed something about the elephant’s constitutional rights.
“I gotta go,” Kwesi said. “Love you. Come home soon.”
“Two weeks. Tell her the elephant has rights.”
Monday, November 10, 2025 – 10:00 AM – The Warehouse
The second week started with Zara’s customer calls.
Gloria had set them up. Five customers, thirty minutes each. Zara on point, Gloria providing context, Harry and the engineers listening from a speakerphone on the folding table.
The first call was with a dispatcher at Atlanta Express — not Renee, her supervisor.
“Walk me through what’s broken,” Zara said.
“How much time do you have?”
“All day.”
The supervisor talked for forty minutes. Exceptions that weren’t real. Tracking data delayed by an hour. Invoices that took three weeks because reconciliation was manual. Rates that came back slow. A carrier database that hadn’t been updated since the last merger.
“Do you use the predictive demand module?” Zara asked.
“The what?”
“What about custom reporting?”
“I download a CSV and put it in Excel. Is there another way?”
Zara muted the phone and looked at Harry. Harry’s jaw was set.
“Who asked for predictive demand?” she asked.
“Consolidated Bulk. 2016.”
“And custom reporting?”
“Internal analytics team. 2014. They wanted dashboards. The customers wanted CSV downloads. We built dashboards.”
Zara unmuted. The second call was worse. The third was worse than that. A shipping manager at a mid-size manufacturer said he’d been asking for automated bill of lading generation for six years. Never built. Meanwhile, the system had added features he’d never opened.
After the fifth call, Zara tallied the numbers on the whiteboard.
Features in the legacy system: ~340
Features customers actually use: ~60
Features built for Consolidated Bulk: ~45
Features nobody uses: ~235
“Eighty percent,” she said. The room was quiet. “We’re not rebuilding 340 features. We’re building 60. The 60 that matter.”
Dane stood up. “And here’s how we build them.” He drew the parallel org diagram on the board — the boundaries, the governance, the migration gate. Below it, he drew the build pipeline: domain narration → AI specs → code generation → human review → domain validation.
“Zara decides what we build. I decide how we build it. Maya builds it. The veterans — Harry, Gloria, Gil, David, Ruth, Warren — are the knowledge base. You feed the machine. Without you, we build the wrong thing. But the machine does the building.”
Robert had been watching from the doorway. Nobody had noticed him come in.
“Show me what you have,” he said.
Maya walked him through the prototype. Basic quote engine. Real-time tracking stub. Exception handling with a unified interface.
“This is one week of building?” Robert asked.
“One week of building. Two weeks of listening.” Maya gestured at the team. “Harry describes the pricing logic. Dane’s process turns it into a spec the AI can work with. I review what the AI produces. Gloria validates against customer behavior. The loop takes hours, not months.”
Robert watched the team. Veterans and engineers at the same whiteboards. Zara’s customer map on the wall — hand-drawn, annotated in three colors, dispatch routes traced in red.
“I spent $47 million trying to get consultants to understand our domain,” he said. “You figured it out in two weeks.”
“We didn’t figure it out,” Zara said. “We asked the customers. Then we asked the people who serve the customers. Then we built the smallest thing that fixes the biggest pain. That’s not genius. That’s product management.”
Robert left without saying more.
That evening, Robert texted Nora.
“Team built a working quote system in one week. They’re using a structured process — domain experts narrate, AI translates to specs, engineers build from specs. Two weeks of listening, one week of building.”
Three dots. Then:
“That’s the right loop. How are you scaling the narrations? If it’s ad hoc conversation, it works at twelve people. It won’t work when you need to train twenty.”
Robert forwarded the message to Dane.
Dane replied in three minutes: “Already on it. Recording every narration session. AI extracts structured knowledge objects — decision rules, exception paths, customer-specific overrides. Building a domain knowledge base that any new engineer can query. The veterans are the source. The knowledge base is the product.”
He added: “Harry’s binder was the first draft. This is the second.”
End of Chapter 4