The Scenario — TeleServ Call Center AI
Your team built an AI assistant for TeleServ agents. It suggests responses, surfaces policy, and summarizes call context in real time. You defined its scope, its behavior, its principals. You wrote the prompt carefully.
Then a customer called. And said something the prompt didn't anticipate. The AI has to make a decision. It has nowhere to go. Your job is to figure out where it should have been able to go — and what it would have needed when it got there.
What is a Context Engineering Map?
It's the design artifact that sits between your prompt and your engineering pipeline. It tells engineers what knowledge the AI needs, where that knowledge lives, who owns it, and how urgently it's needed. Without it, engineers guess. With it, they build. This document is your output — something you could hand to an engineer tomorrow.
Each table has been assigned one of the three moments below. Find yours and read it carefully before Round 1.
A
Billing Dispute
The moment
"I've been charged $47 for an international data plan I never signed up for. I want it removed, I want a credit, and I want to know how this happened. I've been a customer for eleven years."
The agent's prompt covers billing support — but not disputed charges, account history lookup, or unauthorized plan additions.
B
Outage Complaint
The moment
"My service has been out for six hours. I work from home. I've already lost a client call. I need to know exactly when it will be fixed — not 'we're working on it.' A time."
The agent's prompt covers outage acknowledgment — but not real-time network status, estimated restore times, or compensation eligibility.
C
Cancellation Threat
The moment
"I've been trying to get this resolved for three weeks. I'm done. Cancel my service. All four lines. Today."
The agent's prompt covers general retention — but not customer lifetime value, churn risk score, authorized discount tiers, or escalation pathways.
Read your table's moment. List every decision the AI has to make — not how it makes them, just what they are. One decision per line. Don't solve yet.
Compare your individual gap lists. Agree on the top gaps, then complete one entry per gap below. Be specific — "the database" is not a source. "TeleServ billing transaction database" is.
Entry 1 #001
Decision PointWhat must the AI decide?
Knowledge RequiredWhat must it know?
SourceWhere does it live?
AvailabilityWhen can it be accessed?
Real-time
Fetchable on demand
Static
Unavailable
PriorityHow critical is this?
Must have
Should have
Nice to have
Entry 2 #002
Decision PointWhat must the AI decide?
Knowledge RequiredWhat must it know?
SourceWhere does it live?
AvailabilityWhen can it be accessed?
Real-time
Fetchable on demand
Static
Unavailable
PriorityHow critical is this?
Must have
Should have
Nice to have
Entry 3 #003
Decision PointWhat must the AI decide?
Knowledge RequiredWhat must it know?
SourceWhere does it live?
AvailabilityWhen can it be accessed?
Real-time
Fetchable on demand
Static
Unavailable
PriorityHow critical is this?
Must have
Should have
Nice to have
Before share-out — answer these as a group
What surprised your group most about where the knowledge lives?
Which entry had the most disagreement at your table — and why?
What did your AI need most that it had no path to reach?
Each table reads their top two entries aloud. As you listen to other tables, note what's different across moments — especially in Source, Owner, and Availability.
Moment A · Billing
Anything marked Unavailable?
Moment B · Outage
Anything marked Unavailable?
Moment C · Cancellation
Anything marked Unavailable?
Did different tables name different sources for the same kind of knowledge?
That disagreement is a design finding. It means the knowledge architecture is ambiguous — and an engineer building the pipeline would have to guess.
How many entries were marked Unavailable? What does that mean for the AI?
Unavailable isn't a failure — it's a decision. The AI needs a behavior for when the knowledge it needs doesn't exist yet. That behavior has to be designed too.
Who owns the most critical sources? Are they in the room?
Context engineering is a cross-functional conversation. Designers map it. Engineers build it. But the owners of the data sources have to be at the table too.
What's the difference between this map and a prompt?
A prompt tells the AI what to do. A Context Engineering Map tells engineers what to build so the AI has somewhere to go when the prompt runs out.
Key Takeaway
Context engineering is the design layer between your prompt and your data pipeline. Without it, engineers are guessing at what knowledge matters. Without it, your AI will fail at exactly the moments that matter most — not because the prompt was wrong, but because nobody mapped where it needed to go next.