AI Agents need an onboarding plan too

The world now agrees: agents will play a significant role in the workplace. The challenge is getting them to be effective.

Like humans, agents need business context. When you onboard a new employee, you give them a plan, materials to review, people to meet, and a few simple tasks to generate momentum. As humans, we build context in our roles over time to be effective.

Unlike humans, agents don’t need months or years to learn. They can absorb context instantly. 

But, the context needs to be generated, organized, and maintained. If done well, a context graph can be an effective, repeatable onboarding plan for an agent.

What is Context?

Context is the structural knowledge about your business, the decisions your team has accumulated over time, and the procedures for how you operate your business.

Your team, customers, internal process, product catalog, and pricing are all part of your business context. For Customer Success, it also includes accounts, renewals, champions, support tickets, success plans, etc.

Context represents all the tribal knowledge that an employee (or an AI) needs to do a good job.

How Context Graphs Support Post-Sale Teams

Well defined context helps AI move from answering basic questions to proactively surfacing risks and opportunities.

For customer success, account management, and revenue teams, this means multiplying the eyes and ears you have on your accounts so a $200k expansion opportunity doesn’t go unnoticed.

With context, an AI System knows things like:

• Which products are compatible to be sold together?

• Is a 20% usage dip in December a seasonal pattern or a revenue red flag?

• Why are we building product X,Y,Z on our roadmap?

Agents With Context (The Proactive Teammate):

• Operational Mode: The system is PROACTIVE, reasoning over your business logic to surface risks and opportunities without being asked.

• Answering "Which accounts are at risk this quarter?": Faster. The agent already knows your definitions for a "churned account," renewal dates, and a healthy engagement score, skipping discovery and going straight to the insight.

• Defining "churned account": More accurate and deep. The agent uses your team's specific definition (e.g., "didn't renew within 30 days of contract end") rather than a generic textbook one.

• Checking "health score for Account X": More repeatable. Definitions and scoring logic are encoded once and applied consistently, ensuring all your CSMs get the same, customized, and trustworthy answer.

Agents Without Context (The Reactive Tool):

• Operational Mode: The system is REACTIVE, able to answer only direct, explicit questions about known data.

• Answering "Which accounts are at risk this quarter?": Slower. Every query starts with discovery ("What does 'churned' mean here?"), which is the equivalent of the agent having its 'first day on the job' for every single question.

• Defining "churned account": Less accurate and shallow. The agent defaults to a generic definition, often leading to bad analysis, false churn flags, or incompatible upsell recommendations.

• Checking "health score for Account X": Inconsistent. The agent is vulnerable to stale definitions and conflicting logic, leading to different answers across your team.

What does a Context Graph look like?

A context graph is a structured map of your business. It’s a way to teach an AI the nouns, verbs, and numbers of your business. It includes things like:

• Entities that matter most and the attributes that define them

• Relationships between entities

• Metrics you use to track your business

Taken together, this creates a full picture of your business. AI Agents can query this stable structure when they answer a question, rather than having to reconstruct a mental image every time.

Entities

Entities are the basic objects and descriptions that make up your business.

For Customer Success, they includes things like:

• Products: The specific goods or services that the customer is using or has purchased

• Renewal Event: The process and timing of contract or subscription extensions.

• Champions: The key contacts or advocates within the customer's organization.

• Support Tickets: Records of issues or questions the customer has raised

• Success Plans: Documents outlining customer goals, milestones, and strategy

Attributes

All objects have attributes. For example, an Account object has:

• Tier

• Renewal Event

• Health Score

A Renewal Event has: 

• Risk Category

• Renewal Date

Relationships

Every object has relationships to other objects. The relationship definitions help AI agents make connections between them. 

Relationships help an agent understand things like:

• An ID in Hubspot = UUID in the warehouse = account_ID in Pendo.

• An Account subscribes to Products. 

• A Renewal Event belongs to an Account. 

• A Champion works at an Account and influences a Renewal Event. 

• A Support Ticket relates to a Product.

Relationships is where context graphs really outcompete basic “connect your data” setups (i.e. connecting your apps to ChatGPT and Claude). 

With relationships defined, AI systems start generating deep, repeatable insights.

Example of relationship definitions:

relationships:
  - from_entity: Account
    to_entity: Product Family
    type: subscribes_to

  - from_entity: Renewal Event
    to_entity: Account
    type: belongs_to

  - from_entity: Stakeholder
    to_entity: Account
    type: works_at

  - from_entity: Stakeholder
    to_entity: Renewal Event
    type: influences

  - from_entity: Support Ticket
    to_entity: Product Family
    type: relates_to

Example of relationship definitions:

relationships:
  - from_entity: Account
    to_entity: Product Family
    type: subscribes_to

  - from_entity: Renewal Event
    to_entity: Account
    type: belongs_to

  - from_entity: Stakeholder
    to_entity: Account
    type: works_at

  - from_entity: Stakeholder
    to_entity: Renewal Event
    type: influences

  - from_entity: Support Ticket
    to_entity: Product Family
    type: relates_to

Example of relationship definitions:

relationships:
  - from_entity: Account
    to_entity: Product Family
    type: subscribes_to

  - from_entity: Renewal Event
    to_entity: Account
    type: belongs_to

  - from_entity: Stakeholder
    to_entity: Account
    type: works_at

  - from_entity: Stakeholder
    to_entity: Renewal Event
    type: influences

  - from_entity: Support Ticket
    to_entity: Product Family
    type: relates_to

Metrics

Context graphs store your business metrics too.

For example, a health score is a composite metric that consists of usage, sentiment, activity, engagement, ARR, etc. But Company A and Company B likely define it differently.

An AI agent without context doesn’t know your definition; it will guess. And it may guess differently every day.

Metrics should be defined with enough specificity that an agent can compute them instantly.

There are other aspects of context, such as team process definitions and memories about how and why your team made certain decisions. We’ll explore those in another article

The Granularity Tradeoff

How detailed should the graph be? This is a real design decision.

At one extreme, you could add everything. Every single support ticket could have its own spot on the graph. In this case, the agent will have to do less “discovery” when answering questions.

On the other end, you could graph only the entity types and their relationships, letting the agent search for the specifics at runtime.

More granularity means less discovery and faster responses, but much more maintenance, and often less detailed answers because you’re prescribing how the agent must think, and your instructions might be too restrictive.

A graph with 50,000 nodes is harder to keep current and harder to search efficiently than one with 50 well-defined entity types.

There is a balance.

We've found that modeling at the entity-type level, with rich metadata and clear source-of-truth mappings, is the right for now. 

The agent knows what an Account is, where Account data lives, and how to compute Account-level metrics. It queries for specific accounts at runtime. This keeps the graph maintainable while still giving the agent enough structural knowledge to reason well.

This will evolve. As retrieval systems improve, pushing more instance-level data into the graph becomes viable.

What's Next?

The context graph is just one piece of a complete context system. Stay tuned and give me a follow you want to hear more about:

• Memories

• Skills

• Keeping a graph current

• General purpose AI vs a full context system (why connecting your data to Claude and ChatGPT has limitations)

• Evaluations Frameworks