Hypothesis 3

Meet the customer where they are.

Every organization values T2C data differently. What if the system adapted to each one?

The Problem

Every org has different rules.

PB
Pitney Bowes (Tiara)
BPN State Reg Address Website Reviews Social Lawsuits Revenue
F50
Fortune 50 Client
Is Business Active? Revenue Done
RC
Regional Credit Union
BBB Rating Credit Score Sec. of State Lawsuits Local Reviews
PB
Custom build #1
F50
Custom build #2
RC
Custom build #3
?
Custom build #4
?
Custom build #5
?
Custom build #6
...
Custom build #30
...
Custom build #31
...
Custom build #32
30 customers = 30 custom engineering projects. Doesn't scale.
The reality

Tiara at Pitney Bowes checks 8 sources in a specific order. A Fortune 50 client only cares if the business is active. A credit union wants BBB and credit scores. They're all right - for their context.

"Everybody's just doing their own damn thing. It is in their head. It's just a decision that occurs in their mind." Bill Phelan
The old constraint

Building a custom model for each customer meant custom engineering. 30 customers, 30 builds. That doesn't scale. You'd need to reconfigure the product for every different way an analyst thinks.

"Does that mean we got to reconfigure product 30 different ways to satisfy 30 different crazy ways analysts think?" Bill Phelan
The Unlock

The AI follows rules written in English.

We write a search spec - what to check, what to prioritize, how to weigh results - in plain English. The AI agent follows it exactly, using T2C's data as its tools. No custom code. The spec is the customization.

Every customer needs something different
PB Pitney Bowes
8 checks in specific order. BPN and state reg are non-negotiable.
F50 Fortune 50
One question: is the business active? If yes, approve.
RC Credit Union
BBB rating and credit score first. Lawsuits are a dealbreaker.
A skill bridges the gap. Written in English. No code.
pb-wheeler.skill
Check BPN first.
Verify state reg.
Check reviews + social.
Flag if no web presence.
fortune50.skill
Is business active?
If yes, approve.
Done.
credit-union.skill
Check BBB rating.
Check credit score.
Search lawsuits.
Check Sec. of State.
T2C's data. Same for every customer.
◆
BPN
◆
Sec. of State
◆
Address
◆
Reviews
◆
Social
◆
Lawsuits
◆
BBB
◆
Credit
◆
Revenue
Same engine. Same data. Every customer's own model.
"Let the customer program it using our engine." Bill Phelan
The Simulation

Seed, run, learn, repeat.

The skill is directionally accurate. Every correction makes it sharper.

S
pitney-bowes-wheeler.skill
Empty
Always Check
Verify BPN via Salesforce
Check state registration on OpenCorporates
Verify address on Google Maps
Check Google Reviews + Yelp
Search Facebook + Instagram
Search for lawsuits and fraud claims
Priority
BPN and state registration are non-negotiable
Flag
If no web presence, escalate for manual review
Past DecisionCheckedResult
Acme Welding LLCBPN, State, ReviewsApproved
Glacier TransportBPN, State, BBB, RevApproved
Tundra EquipmentBPN, State, RevenueFlagged
Northern Lights HVACBPN, State, ReviewsApproved
Fairbanks FuelBPN, State, Rev, BBBRejected
Harbor View BoatsAll 8 checksApproved
Skill generated from 100 past decisions
Patterns found: BPN checked 100%. State reg 100%. Social 73%. Revenue flagged in 4 of 5 rejections.

Rocky's Marine, Inc.

Petersburg, Alaska - First run
Skill match
60%
BPN #4821 - Active
State Reg Good Standing
Address Verified
Reviews 4.8 stars (12 reviews)
Social Facebook - Active
Lawsuits None found
Revenue ~$60k (estimate) "I need D&B number, not an estimate"
BBB Not checked "Always check BBB for Alaska businesses"
Week 1 - Seeded from history
Week 4 - After analyst feedback
Week 12 - Fully adapted
Auto-reject Analyst reviews Auto-approve
60%
Auto-decidable at launch
90%
Auto-decidable after 12 weeks
782
Auto-approved
< 620
Reject
621-749
Review
750+
Approve
Weighted by Tiara's priorities:
BPN (25%) + State Reg (25%) + Reviews (15%) + Social (10%) + Revenue (15%) + BBB (10%)
Step 1: Seed with history

Take 100 of a customer's past decisions - what they checked, what they skipped, what they approved. The AI reviews every one and writes a base skill in plain English.

Step 2: T2C runs first

The generated skill gets the first experience to 60%. Good enough that T2C is now the starting point, not the last resort. The results look familiar because they follow the analyst's own patterns.

"Once they see it, they'll be like, 'Aha, that's it. I got it. Done. Let's go. Move on. Do it again.'" Bill Phelan
Step 3: Analyst gives feedback

She reviews the results, corrects what's wrong, tells T2C what's missing - in her own words. The skill updates in English. T2C reruns with the new instructions.

Step 4: The skill compounds

Every correction makes the next search better. The gray area shrinks. Each customer ends up with their own T2C score, weighted by what they actually care about.

"If it's 750 or more, I approve it automatically. If it's 620 or under, I reject it. But 621 to 749, I got to look at those." Bill Phelan
The result

Each customer ends up with their own T2C score - weighted by what they actually care about. Same engine. Different skill per org. No custom engineering.

"We're expanding the delight and expanding the dependence." Kyle Williams
The Result

Every customer's own model.

No custom engineering Skills are English, not code
Analyst feels smart They see their process, not a generic one
Trust builds naturally They correct it
Scales to every customer Same engine, different skill per org