AI Case Studies

1 Pick the Right Problem

Identify the measurable outcome that matters

The company’s 1,000-person support organization was struggling with seasonal cancellation surges tied to renewal cycles. These waves created SLA misses, backlogs, temporary staffing costs, vendor auto-renewal penalties, agents spending time on non-growth accounts, and slower response times for high-value customers.

Instead of automating something “interesting,” we targeted the single workflow with the largest, clearest, most measurable operational leverage: cancellations.

This reframing turned a low-prestige, low-upside task into the biggest productivity unlock in the entire support operation.

2 Fix the Data Bottlenecks

Make sure the signals AI depends on are trustworthy

Cancellations were complex because every state had different rules, deadlines, and requirements. Before automating anything, we:

• Mapped the dataset: incoming emails, chats, voicemails, IVR requests
• Identified signal reliability: reason for cancellation, entity name, state requirements
• Aligned vendor system rules to internal CRM data
• Defined the “safe zone” where the agent could act autonomously
• Outlined the guardrails where escalation was mandatory

This ensured the AI agents were acting on clean, traceable, high-confidence signals—the foundation for any real-world AI deployment.

3 Set Up the Team to Succeed

Establish the Product-Data-Operations model

To make the AI agent stick, we designed the operational system around it:

• Human-in-the-loop for early accuracy tuning
• Frontline calibration so the tone matched a 4.7-star service standard
• Refund authorization thresholds
• Real-time escalation paths for edge cases
• Agent QA rituals to validate accuracy
• Training the support organization on what the agent would handle vs. what humans should focus on

This ensured the AI didn’t just “work”—it worked the way the team needed it to, building trust instead of resistance.

4 Prove It Works in the Real World

Deliver initial measurable results in 60-90 days, then compound through iteration

Once deployed, the impact was immediate and undeniable:

1 Pick the Right Problem

Growth was outpacing the mail team’s human capacity

The company’s mail operation is mission-critical to products like Registered Agent Service, Premium Mail Forwarding, and Virtual Office. These products promise same-day scanning and instant notifications—especially for legal mail (Service of Process). Failing an SLA isn’t just bad service—it’s regulatory risk.

The Problem Before AI

• Mail volume had grown to ~700,000 pieces per month
• The year prior, a +200,000 increase required 100+ new hires
• Growth was accelerating—not slowing
• Annual state compliance seasons created massive mail surges
• To survive peaks, the company repeatedly hired, trained, and let go of temp workers
• This churn exhausted top employees and increased errors
• Backlogs risked SLA violations and frustrated customers

The company wasn’t trying to reduce staff—they couldn’t hire fast enough to keep up. They needed a solution that protected existing employees, eliminated the burnout cycles, stabilized SLA performance, and supported exponential growth.

This made the mail operation the highest-leverage problem in the company.

2 Fix the Data Bottlenecks

Build the signals the AI needed to perform with human-level trust

We created a clear foundation:

• A full taxonomy of all mail categories across 50 states
• Recognition patterns for banking, licensing, insurance, legal notices, state documents, and regular mail
• Workflow mapping from OCR → matching → CRM routing
• Training data structured around “where to look” on each mail type
• Guardrails for legal notices and compliance-sensitive documents

Accuracy Transformed Over Two Quarters

• 80% OCR accuracy at launch
• 98% accuracy by month six
• Higher than the average human accuracy (~97%)

Mail Type Coverage Expanded

• 50% at launch
• 90% after two quarters

Fixing the data bottlenecks meant the AI wasn’t just fast—it was trustworthy.

3 Set Up the Team to Succeed

AI wasn’t built to replace people—it was built to support them

We rolled out the system in human-first phases:

Phased Human-in-the-Loop Rollout

We introduced AI by mail type, not by volume:

1. Low-volume states
2. State notices
3. Banking mail
4. Insurance & licensing
5. Legal mail
6. General business mail

Each category had matching rules, escalation rules, tone and format consistency, and clear audit trails.

How Trust Was Built

The operations team trusted the agent because:

• Side-by-side validation showed rising accuracy
• Dashboards made decisions transparent
• Human override was always available
• Error reporting was clear
• Processing sped up immediately
• SLAs stabilized within weeks

Then something powerful happened: After a few cycles, frontline employees didn’t want to be in the loop. The agent became the most reliable “employee” on the team.

Morale increased because repetitive work disappeared, people focused on judgment-driven tasks, there was no more temp-hiring chaos, and no more burnout each compliance season.

This is responsible AI in its ideal form: Humans do better work, not more work.

4 Prove It Works in the Real World

Initial results in weeks, full transformation over two quarters

1 Pick the Right Problem

Customers weren’t activating—and activation is everything in PLG

The company offered a Website + Hosting service immediately after LLC formation, registered agent signup, or domain purchase. On paper, it should have been a high-attachment, high-ARPU product.

But the data told a different story:

Before AI

• Building a website took 2-4 hours of manual work
• Customers had to fill out a long onboarding questionnaire
• Many said “I’m not ready for a website yet”
• <20% even started the free trial
• Only 20% of trials converted to paid
• Most customers never came back—”call me later” was a dead end
• The product was a PLG dead zone

The real problem wasn’t the product

It was the activation barrier—too much friction, too many decisions, too much cognitive burden.

We diagnosed that reducing friction (not adding new features) would drive the biggest revenue unlock.

2 Fix the Data Bottlenecks

Infer what customers mean, not what they explicitly tell you

We replaced the long questionnaire with a smart inference engine.

AI used:

• Domain name → infer business type (“acmesalon.com” → hair salon)
• Billing address → infer local customer expectations, pricing norms, SEO phrasing
• Vertical patterns → determine layout, language, structure
• Persona generation → speak in the tone a business owner naturally uses
• Competitive research prompts → make the site look like the best in that category
• WordPress templates → make the site fully portable and owner-controlled
• Image + copy blocks → industry-specific, conversion-optimized

The result:

A full, personalized 3-page website—home, about/services, contact—generated in 90 seconds with no questionnaire.

Customers went from: “I don’t know where to start.”
to: “You nailed it—this is exactly what I needed.”

This solved the cold start problem that blocked adoption.

3 Set Up the Team to Succeed

AI eliminated the grunt work—humans did the high-value work

At first, designers and copywriters were skeptical:

• They didn’t like the assumptions
• They worried the templates looked formulaic
• Support teams reacted to a loud minority of unhappy callers
• Internally, it felt risky

But then the data changed the narrative:

Customer Behavior Proved It Worked

• Most didn’t request changes
• Most kept the site as-is
• Most converted from trial to paid
• Most renewed month after month
• Interviews showed customers were relieved, grateful, and felt understood

Human roles shifted upward:

• No more repetitive 2-4 hour website builds
• No more chasing questionnaire follow-ups
• No more hand-holding early decisions
• Teams focused on: prompt engineering, quality governance, improving templates, customer insight work, enhancing the product

This became one of the best internal examples of responsible AI: AI didn’t eliminate jobs—it eliminated the parts of the job everyone hated.

4 Prove It Works in the Real World

Immediate activation improvements, compounding conversion over 90 days