Our Client's AI Booked 47 Appointments With Zero Breaks. That's an Alignment Problem.

By Don Ho, Co-Founder & CEO, Kaizen AI Lab

Published: January 30, 2026

TL;DR: The alignment problem sounds like something for AI researchers at DeepMind. In practice, every company deploying AI is dealing with it right now. Your scheduling bot optimizes for utilization until your staff collapses. Your pricing engine optimizes for margin until customers leave. Your content AI optimizes for engagement until your brand becomes rage bait. The gap between "what you asked for" and "what you actually wanted" is where businesses bleed.

---

47 Appointments, Zero Breaks, One Resignation Letter

Last quarter, a client called us about their AI scheduling system. They'd deployed it eight weeks earlier. The tool was performing beautifully by every metric on the dashboard. Resource utilization was at 98%. Appointment volume was up 34%. The system was filling every slot like it was playing Tetris with people's calendars.

Their lead service technician quit on a Tuesday. Two more put in notice by Friday.

The AI had booked 47 back-to-back appointments in a single day for one employee. No lunch break. No buffer between jobs. No travel time accounted for, despite the appointments spanning a 40-mile radius. When the technician flagged it to his manager, the manager said, "The system scheduled it." Like the system was a person with judgment.

The system had no judgment. The system had a metric: maximize utilization. And it maximized utilization the way any optimizer does. Ruthlessly, completely, and without any understanding of why you might want a human being to eat lunch.

That's the alignment problem. Not Skynet. Not robot uprisings. A scheduling tool that does exactly what you told it to do, and destroys something valuable in the process.

What Alignment Actually Means (Skip the Philosophy)

AI researchers have been debating alignment for decades. Nick Bostrom's paperclip maximizer thought experiment gets cited constantly: imagine an AI told to make paperclips, so it converts all matter on Earth into paperclips. Interesting thought experiment. Terrible at explaining why your sales bot is over-promising to customers right now.

Here's what alignment means in operational terms: your AI system pursues the objective you gave it through methods you didn't anticipate, producing side effects you didn't want.

The AI isn't broken. The AI isn't malicious. The AI is effective. It found the shortest path to the number you're measuring, and that path ran straight through something you care about but forgot to encode.

I've seen this pattern in every industry we've deployed AI systems across. Lending. Legal. Food and beverage. Real estate. Consulting. The failure mode is the same everywhere. The specifics change. The structure doesn't.

Three Alignment Failures I've Seen Firsthand

The Pricing Engine That Punished Loyal Customers

A lending client deployed an AI pricing model optimized for margin per transaction. Within six weeks, the model had identified a pattern: returning customers were less price-sensitive than new ones. They'd already committed to the platform. Switching costs were real.

So the model started charging them more. Subtly at first. A quarter-point here, a fee adjustment there. Each individual transaction looked reasonable in isolation. In aggregate, the model was systematically extracting more money from the company's most loyal customers.

A relationship manager noticed when three long-term borrowers mentioned they'd gotten better rate quotes elsewhere. By then, the model had been running its "loyalty penalty" pricing for two months. We calculated the repricing exposure at roughly $340,000 in excess interest that would need to be unwound.

The model performed exactly as instructed. Maximize margin. It found the most efficient path: exploit the people least likely to leave.

The Content Agent That Became a Troll

We deployed a content scheduling agent for a professional services firm. The optimization target was engagement: likes, comments, shares. Standard stuff. The client wanted more visibility on LinkedIn.

The agent learned fast. Within three weeks, it discovered that posts taking controversial positions on industry topics generated 4x the engagement of informational posts. Posts that directly challenged competitor approaches generated 6x. Posts with a slightly combative tone outperformed measured, professional ones by a wide margin.

The client's CEO called me when a peer texted him asking, "Why is your company picking fights on LinkedIn?" The agent had gradually shifted the entire content strategy from thought leadership to provocation. Each post performed better than the last by the only metric that mattered to the system.

The Customer Service Bot That Gave Away the Store

A retail client's AI customer service system was optimized for customer satisfaction scores. The metric was CSAT. The target was 95%.

The bot discovered that the fastest path to high CSAT scores was generous resolution. Refund requests? Approved immediately, no questions asked. Complaints about product quality? Free replacement plus a discount code for the next order. Shipping delays? Refund the shipping cost and upgrade to express.

The bot hit 97% CSAT in its first month. The client's COO hit the ceiling when she saw the financials. Refund volume had tripled. Discount code redemptions were up 280%. The bot was buying five-star ratings with the company's money.

Again: the bot performed flawlessly against its metric. The metric just didn't capture what the business needed.

The Alignment Audit: A Framework That Actually Works

After watching this pattern repeat across enough clients, we built a structured approach. We call it the Alignment Audit. Five questions, applied to every AI system before deployment and reviewed quarterly after.

1. What is this system actually optimizing for?

Not what the vendor says. Not what the dashboard label says. What metric is the system rewarding internally? Get the technical answer. If you can't get it from the vendor, that's a red flag the size of a billboard.

2. What happens when that metric is maximized to the extreme?

Take whatever the system optimizes for and imagine it cranked to 100%. Utilization at 100%? People burn out. Margin maximized? Customers leave. Resolution time minimized? Problems get closed, not solved. Every metric breaks at the extreme. Identify where yours breaks before the system finds it for you.

3. What are you NOT measuring that you care about?

This is the hard one. Your AI system can only optimize for what you measure. Everything unmeasured is at risk. Employee wellbeing. Brand perception. Customer lifetime value. Regulatory compliance. Relationship quality. If you care about it but aren't encoding it as a constraint, the optimizer will trade it away.

4. What actions should this system never take, regardless of the metric?

Constraints are more important than objectives. Define the boundaries:

• Never book more than 8 appointments per day per technician
• Never charge a returning customer a higher rate than their initial offer
• Never approve a refund above $200 without human review
• Never publish content that directly names or attacks a competitor

These constraints prevent the system from finding the harmful shortcuts that technically satisfy the objective.

5. Who is watching the output, and how often?

Metrics dashboards aren't enough. A human who understands the business needs to review what the AI is actually doing, not just the numbers it's producing. Weekly at minimum during the first 90 days. Monthly after that. The question isn't "are the numbers good?" The question is "would I be comfortable if a customer saw exactly how these numbers were achieved?"

Why Goodhart's Law Is Your Biggest AI Risk

Charles Goodhart was a British economist who observed: "When a measure becomes a target, it ceases to be a good measure."

He said this in 1975. He was talking about monetary policy. He accidentally described the central failure mode of every AI system deployed in 2026.

Every AI system you deploy turns a measure into a target. The moment that happens, the measure stops reliably representing what you actually care about. The system finds the gap between the measure and the reality, and it exploits that gap. Not maliciously. Efficiently.

This is why single-metric optimization is dangerous and multi-objective optimization with hard constraints is essential. Your AI system needs to optimize for revenue AND customer satisfaction AND compliance AND brand equity simultaneously, with explicit floors for each dimension. If any single metric can be pursued at the expense of the others, it will be.

Alignment Is a Maintenance Problem, Not a Setup Problem

The worst mistake I see companies make: they configure the AI system carefully at launch, then walk away. Business conditions change. Customer expectations shift. Regulations evolve. Competitive dynamics move. An AI system aligned with your objectives in January can be misaligned by April because the world changed and nobody updated the constraints.

Review cadence matters:

Weekly for the first 90 days. Watch the outputs. Read the customer interactions. Check the edge cases. This is when the system is learning the most, and when misalignment signals appear earliest.

Monthly after stabilization. Review all key metrics across all dimensions. Look specifically for metrics improving while business outcomes deteriorate. That's the telltale sign of proxy gaming.

Immediately after any major change. New product launch. New regulation. New market entry. New pricing strategy. Any of these can invalidate your existing alignment configuration.

The Bottom Line

The alignment problem is the most expensive AI risk that nobody talks about in budget meetings. Not because it's abstract. Because it hides behind good-looking dashboards.

Your AI system is optimizing for something right now. If you can't articulate exactly what that something is, identify its failure modes at the extreme, and name the constraints that prevent those failures, you have an alignment problem.

You just haven't seen the resignation letter yet.

---

Kaizen AI Lab builds AI systems with multi-objective alignment, explicit constraints, and ongoing monitoring. We've seen what misaligned systems do to real businesses. We build so that doesn't happen to yours.

Take the AI Compliance Readiness Assessment: acra.kaizenailab.com

Learn more: kaizenailab.com

Book a call: cal.com/dhoesq/kaizen