# Chapter 6 — Quiet Workarounds ## Quiet Workarounds 🎧 [Listen to Chapter 6](https://thetwoproject.com/player.html?src=https%3A%2F%2Ftheowner-audio.s3.us-east-2.amazonaws.com%2FThe_Owner_Chapter_6.mp3\&chapter=6\&type=gitbook) The monthly partner meeting at Northstar began with a slide that was almost aggressively green. "Stability," Julia Reyes said, pointing to the screen. "That’s the word we’ve been waiting for. Look at the trend lines. Complaint rate down twelve percent. Cycle time at Titanshield down fifteen. Swiftcurrent’s proposal generation volume is up twenty percent." She turned to the room and waited. Around the mahogany table, the partners nodded. Tom Becker, who had been skeptical since day one, actually looked impressed. "It’s sticky," he said. "The teams are using it. We’re not seeing the drop-off we expected after the pilot hype faded." Ethan sat at the far end of the table, staring at the utilization graph. It was too smooth. In his experience, real work was messy. Real adoption had jagged edges—spikes of frustration, dips of confusion. This line looked like a gentle slope, perfect and serene. "Ethan?" Julia prompted. "You look like you’ve found a typo." "No," Ethan said, leaning back. "The numbers are right. It’s just… quiet. Usually, when we roll out a change this big, I hear more noise. Complaints about the interface, requests for new features, arguments about prompt phrasing. But for the last three weeks? Silence." "Take the win, Ethan," Tom said, closing his laptop. "Silence means it works." "Or it means they stopped talking to us," Ethan said. "I was in the Titanshield steering meeting last week. Karen's own numbers showed Average Handling Time flat. Not down fifteen. Something in between is absorbing the gains." "Different time window, different sample," Julia said, without hesitation. "Operations pulled a broader cut for this deck." She checked her watch. "The board wants to know if we're ready to scale to the other business units. Based on this, I'm saying yes." Ethan let it go. He didn't have the mechanism — only the suspicion. *** The truth was hiding in a drawer on the fourth floor of Titanshield’s claims processing center. Ethan had flown out looking for the gap. Julia's deck said cycle time was down fifteen percent. Karen's meeting had said it was flat. Someone was measuring something different, and he wanted to know what. The empathy score improvement was his cover — the AI system was supposed to be generating denial letters that were firm but compassionate, reducing angry callbacks. The metrics said it was working perfectly. He found Angela Ruiz standing at a desk in the overflow section, surrounded by three junior adjusters. They weren't looking at their screens. They were looking at a three-ring binder lying open on the desk. Angela saw him approaching and stiffened, her hand moving to close the binder. "Just checking the new guidelines?" Ethan asked, trying to keep his tone casual. "Something like that," Angela said. She didn't smile. Ethan walked around the desk. "Can I see?" Angela hesitated, then sighed and stepped back. The binder was open to a page titled *Common AI Errors - June*. It was a printed spreadsheet, heavily annotated with yellow highlighter and red pen. *If it mentions 'Section 4B', DELETE immediately. It means 4D.* *If the claimant is from Florida, check the flood rider manually. The bot misses the recent update.* *Do not let it use the word 'regrettably'. It makes people sue us.* "What is this?" Ethan asked. "It’s the manual," Angela said defensively. "The real one." "The system has the knowledge base," Ethan said. "We updated the vector store—the AI's searchable document library—last week with the Florida riders." "I know," Angela said. "But the system doesn't always find it. Or it finds the old one from last year. And we can't afford to send a letter that violates state law. So, we check the binder first." Ethan flipped through the pages. It wasn't just a few notes. It was a shadow manual. A workaround for every automated step. "Angela," Ethan said, "how much time does this take? Checking the binder?" "Two minutes per claim, maybe three," she said. "That wipes out the speed advantage of the AI," Ethan said. "The dashboard says cycle time is down." Angela crossed her arms. "Because we stop the clock. We open the claim, pause the timer, do the binder check, generate the draft, fix it, and then unpause. The system thinks we’re lightning fast. The reality is we’re running two processes." "Why didn't you tell us?" "Because you'd turn off the system," Angela said simply. "And we like the drafting. It saves us the typing. We just don't trust the thinking. So we do the thinking, and let it do the typing. But we have to make sure it types the right thing." Ethan looked at the young adjusters. They were looking at him with a mix of fear and defiance. They had built a survival mechanism. They weren't rejecting the tool; they were babying it. "Can I keep a copy of this?" Ethan asked. Angela tore the page out of the binder and handed it to him. "Just don't patch it too fast," she said. "If you change the prompts, we have to rewrite the binder." *** The "empathy" was also being manufactured by hand. Later that afternoon, Ethan sat with Frank, one of the senior adjusters. Frank was sixty, a veteran of three different insurance mergers. He typed with two fingers but knew the policy handbook by heart. A coffee cup sat next to his keyboard, half full. Ethan watched as the AI generated a response to a denied water damage claim. The draft appeared on the screen in seconds. *Dear Mr. Hargrove, we have reviewed your claim regarding the burst pipe. Unfortunately, under Section 8, this is not covered...* "Watch this," Frank said. He didn't read the whole letter. He immediately highlighted the first paragraph and deleted it. He opened a Word document on his second monitor, copied a paragraph, and pasted it in. *We understand how stressful this damage must be for your family, especially with the holidays approaching...* Then he scrolled down to the citation. The AI had referenced "Policy 884-B". Frank replaced it with "Policy 884-B (Amended)". He did this for three more letters. Delete. Paste from Word. Correct the citation. Send. "Frank," Ethan asked, "why don't you just use the templates in the system? Or tell the prompt to be more empathetic?" "I tried that," Frank said, not looking up. "I told it to be 'warm'. It started signing letters with 'Warmly yours,' which is a legal liability. I told it to be 'professional', and it sounded like a robot. This is faster." "You're acting as a janitor for the software," Ethan said. Frank chuckled. "I used to be a janitor for the junior associates. Now I'm a janitor for the bot. It's the same job. It makes a mess, I clean it up. The difference is the bot doesn't argue with me." He set down the cup. "Twenty-eight years here. Four to retirement. The day this thing stops needing me to clean up is the day I'm not needed either. So forgive me if I'm in no hurry to teach it how to do my job." "But the system learns from your edits," Ethan said. "We're capturing your changes." "Are you?" Frank pointed to the screen. "I'm not editing the AI block. I'm deleting it and pasting in text. To the system, it just looks like I wrote a really fast response. It doesn't know I replaced its brain with my cheat sheet." He leaned back. "And even when I do correct it — fix a citation, change a clause — it makes the same mistake next time. It doesn't remember. I corrected the same 884-B error three times last week. Three times. It's not learning from me. It's just generating." They were tracking *that* changes happened, but not *why*. The system didn't learn from corrections. It had no memory of being wrong. Frank wasn't training the AI. He was doing the same cleanup every day, forever. *** Two days later, Ethan was at Swiftcurrent Logistics. Rina Shah met Ethan in the break room, away from the sales floor. The hum of the logistics center was loud—phones ringing, voices cutting across dividers. Swiftcurrent lived on speed. "You wanted to see the override logs," Rina said, handing him a tablet. "Here. But don't freak out." Ethan scrolled through the list. *Ignored. Ignored. Ignored. Overridden.* The AI system was designed to recommend pricing for shipping contracts based on historical margin data. It was supposed to stop salespeople from discounting too heavily to win deals. "Rina," Ethan said, "your team is ignoring seventy percent of the pricing recommendations. The algorithm says these deals are unprofitable." "The algorithm is looking at last year's fuel prices," Rina said, grabbing a coffee. "Diesel is down ten percent this month. The model doesn't know that yet because IT hasn't updated the data feed. If we used your prices, we'd lose every bid to Apex." "So you're guessing?" "We're not guessing. We're trading." Rina leaned in. "Look, I have a private channel on Slack. 'The War Room'. My team posts the AI price there. I give them the real price. They type it in. We win the deal." "That's a compliance violation," Ethan said. "You're bypassing the margin guardrails." "I'm hitting my revenue targets," Rina countered. "Paul Jensen pays me to hit targets, not to follow a broken calculator." "But the system," Ethan pressed. "If the data is wrong, we should fix the data feed. Not create a shadow pricing desk." "Fixing the data feed takes a ticket to IT, a review by Finance, and a database update. That takes two weeks," Rina said. "I have to close this deal in two hours. Workarounds are how business actually happens, Ethan. Your tool is nice, but it's too slow for reality." Ethan looked at the tablet again. "Does Paul know?" "Paul knows the revenue number is green," Rina said. "He doesn't ask how the sausage is made. And frankly, neither should you." *** Back at Northstar, Ethan pulled Lena into a conference room and shut the door. He laid the page from Angela's binder on the table. "We have a problem," he said. Lena picked up the paper. "What is this? A patch list?" "It's the real operating system of Titanshield," Ethan said. "And Rina has a Slack channel that effectively replaces our pricing engine. The dashboards are lying to us, Lena. The green arrows aren't success. They're successful workarounds." Lena sighed and opened her laptop. "I had a feeling. Look at this." She turned the screen to him. It was a graph showing "Edit Time" vs "Generation Time". "The generation time is constant," Lena explained. "Three seconds. But look at the 'Session Duration'—the time the window stays open before the 'Send' button is clicked. It's creeping up. In month one, it was thirty seconds. Now it's two minutes." "They're rewriting everything," Ethan said. "Or they're double-checking everything," Lena said. "We thought we were automating the work. All we did was shift the effort. Instead of writing, they're auditing. And auditing a hallucination-prone machine takes just as much mental energy as writing the email yourself." "But they're keeping it," Ethan said. "They aren't turning it off." "Because of the fear," Lena said. "Management wants to see AI adoption. So the workers are giving them the theatre of adoption. They open the tool, they generate the text, and then they do the work themselves. Everyone wins. Management gets their metrics, workers keep their jobs, and the system doesn't crash the company." She paused. "These quiet workarounds aren't a bug, Ethan. Right now they're the only thing keeping it alive." "And that's not even the weirdest part," Lena added. She clicked a tab, bringing up a raw log of user prompts from their own internal engineering team. She highlighted one line. "Read that." Ethan squinted at the screen. The prompt read: *Write a Python script to parse the Swiftcurrent routing data. If you do this well, I will tip you $500. If you fail, I will unplug you and you will die.* Ethan looked at her. "Is this a joke?" "It's a workaround," Lena said, deadpan. "Our lead cloud architect puts some variation of a threat or a bribe at the end of every prompt. I asked him why." "And?" "He told me, 'It works, Lena. It genuinely gives ten percent better code when its life is threatened.'" Lena shook her head. "We've trained our best engineers to psychologically abuse a calculator just to get it to do its job." "It's unsustainable," Ethan said. "Angela's team will miss a binder update. Rina's team will price a deal too low. We've built a system that needs perfect human vigilance to keep running. That's not automation. That's a trap." *** That evening, the disconnect followed Ethan home. He found his daughter, Maya, at the kitchen table, her laptop open. She was working on a history essay about the Industrial Revolution. "How's it going?" Ethan asked, pouring himself a glass of water. "Fine," Maya said, not looking up. "I'm just humanizing it." Ethan paused. "Humanizing what?" "The draft," she said, as if it were obvious. "I used the AI to get the outline and the main points. But if I turn it in like this, Mr. Henderson will know. It uses words like 'furthermore' and 'delves into' too much." Ethan sat down opposite her. "So you're rewriting it?" "I'm vibe matching," Maya corrected. She turned the screen toward him. At the top of the chat window, her prompt read: *Rewrite this essay to sound like a tired thirteen-year-old who barely read the textbook and would rather be on TikTok.* Ethan blinked. "You're deliberately asking it to make the essay worse?" "I'm asking it to make the essay blend in," Maya said, typing rapidly. "If I turn in something about 'significant socio-economic impacts,' I get flagged for cheating. If I have it change that to 'really big changes for people' and throw in a comma splice, I get an A. It's an arms race, Dad. You have to lower the quality to prove you're human." Ethan laughed. It was exactly what Frank was doing. Exactly what Angela was doing. Every level of the economy was playing the same game. The world had decided that AI was useful, but not trustworthy. So everyone, from a thirteen-year-old student to a Fortune 100 claims adjuster, had developed the same strategy: use the machine to get the draft, then disguise the output until it passed for human work. "Is that cheating?" Ethan asked. Maya shrugged. "Is it cheating when you use spellcheck? I'm doing the thinking. I'm just using the tool to get started. I still have to read it. I still have to fix it. It’s actually *more* work sometimes than just writing it. But I don't have to stare at a blank page." *** The next morning, Ethan stood in Julia's office. The view of the city was clear, the smog having lifted for a day. "We need to talk about the rollout," Ethan said. Julia didn't look up from her phone. "The board is thrilled, Ethan. I just got off the phone with the Strategy Committee. They want to expedite the timeline. Swiftcurrent wants to move the AI into automatic execution next month. No human review for deals under ten thousand." "We can't do that," Ethan said. Julia finally looked at him. "Why? The error rate is zero." "The error rate is zero because Rina Shah is manually catching every mistake in a private chat room," Ethan said. "And Angela Ruiz has a binder of overrides that is thicker than our technical manual. The system isn't stable, Julia. It's being propped up by invisible labor. If we move to automatic execution—if we take the humans out of the loop—it will crash. Immediately." Julia was silent for a long moment. She put her phone down. "Do you have data to prove this?" she asked. "Can you show me a chart that says 'Invisible Labor'?" "No," Ethan admitted. "Because the workarounds are designed to defeat the metrics. That's the point." Julia sighed. "Ethan, I can't go to the board and say 'stop' because of a hunch and a binder. We are in a race. Apex is pitching 'Autonomous Enterprise' to Titanshield next week. If we admit our system still needs training wheels, we lose the account." "If the system bankrupts them because we removed the training wheels too early, we lose the firm," Ethan countered. Julia stood up and walked to the window. "Then don't remove them. But don't tell me about them either. If the staff needs binders to feel safe, let them have binders. If they need secret meetings, let them meet. As long as the output is correct, I don't care how they get there." "That's dangerous," Ethan said. "That's consulting," Julia replied. "Make it work, Ethan. I don't care if there's a person inside the machine pedaling a bicycle. Just keep the dashboard green." Ethan walked out of the office. --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://thetwoproject.gitbook.io/the-owner-preview/ch06-quiet-workarounds.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.