3 types of context every AI agent needs, and the one everyone skips

AI agents lean on a lot of data to hold a conversation: a model, a knowledge base, a set of rules, your CRM data, and whatever else seems relevant.

What's usually missing is product context: the real-time and historical signal of what people actually do in your product. It's been hard to access, but the teams that found a way in saw agent performance improve significantly.

With that context, agents answer with more specificity and skip straight to the solution that actually fits, instead of the closest match in a knowledge base.

Three kinds of context every agent needs

There are three types of data an agent should draw on:

1. Business context: the account’s annual contract value, segment, and firmographics.
2. User context: an account’s CRM record, their role, and tier.
3. Product context: the actions someone took within your product, how often they log in, and where they were in your product before they opened the chat.

The first two are common. Most CRMs will hand business and user context to any agent that asks for it and is properly set up. Product context, on the other hand, lives somewhere else. It comes from your analytics platform, the same system that’s tracking in-app behavior for your PMs and devs.

The knowledge base tells an agent what a feature does, while product context tells it that this specific user has navigated to that feature seven times across three sessions, gotten partway through setup each time, and returned to try again, which produces a different response than anything the knowledge base alone can supply.

For builders running an agent across more than one app, the same product data can follow a person’s path across all of their touchpoints.

Giving your agent signals that let an agent step in before a user asks for help

Rage clicks, dead clicks, repeated failed attempts, and workflow abandonment are product signals. Most agents don't have access to them, but the ones that do can stop waiting to be asked.

A support agent that detects a user has retried the same action three times in four minutes can surface a specific, contextual message at the point of friction before the frustration compounds into a support conversation or something worse.

The distinction between a reactive agent and a proactive one comes down to which signals it has access to and whether those signals arrive fast enough to act on.

Right now, the most impact product context has on agents manifests in the support space. Give your Fin, Decagon, Sierra, or homegrown support agent in-app behavior, and they can:

• See where users are dropping off and step in to guide them through the entire workflow
• Offer guidance when users are rage clicking or U-turning
• Walk users through setup and tasks that impact a user’s success in-app

Teachable plugged Pendo context into their Fin support agent and ended up resolving 67% of support tickets without involving a human at all. Learn more about their playbook.

Questions to ask before your next agent sprint

Most teams don't have a clean answer to what their agent knows about the user it's talking to. Before the next round of feature work, ask yourself:

1. Does your agent know what a user did in-app before they engaged with it?
2. Does it know whether they've been stuck on the same problem before? If yes, how many times?
3. Does it know whether previous conversations resolved the underlying issue? Or did a user just give up and leave your app altogether?

If the answer to all three is no, the next thing your agent needs is the product context to reason from.