Meet Segments: The Definitions That Decide What Your AI Agent Already Knows

A premium customer asks a billing question and gets an answer that already knows they are on the paid tier. A user whose account is flagged gets routed into exactly the right recovery flow, no fishing, no wrong turns. Someone contacting one brand in a large portfolio never hears a policy that belongs to a sister brand down the hall. None of those customers know why the answer fit so well. That's the point.

What makes those moments work is Segments. It's one of the least visible parts of the Maven platform and one of the most load-bearing. So this first entry in our product series goes where announcements usually don't: the audience definitions that determine what knowledge documents your agent can access before it says a word.

What Segments actually is

Segments are reusable, named definitions that describe a kind of conversation or a kind of user. Premium accounts. Banned accounts. Conversations on the email channel. Users who've already verified their identity. You write it once, give it a readable name, and attach it wherever it applies. The same definition drives a knowledge base, an action, or an entire workflow, without being rewritten each time.

A Segment is a definition of an audience, who they are, what channel they’re on, what they’re allowed to see. The easy-to-miss part: Segments is a deterministic piece of the product, and runs before any model reasoning. It's not a suggestion in a prompt the model might honor on a good day. It's a gate. The agent evaluates the definition first, decides what's eligible, and only then reasons over what made it through. Knowledge or actions that fall outside the segment aren't just discouraged. They're not in the room.

That ordering is the whole design. A model asked nicely to ignore the wrong brand's pricing will usually comply. "Usually" isn't a standard an enterprise can stand behind. A segment removes the question entirely, because the wrong content was never eligible to be retrieved in the first place.

Where your customers run into segments without realizing it

Segments rarely get named in a release note because they don't look like a feature. They look like the absence of a problem. A few of the places they're quietly working:

Tiered accounts

A standard user and a paid user can ask the same question and need different answers. A segment keyed to account state makes the right knowledge eligible for each, so the agent isn't guessing the customer's tier from context clues. The customer just gets an answer that fits their plan.

Brand and business isolation

One Maven customer runs a portfolio of consumer brands on a single agent. Each brand's knowledge has to stay sealed off from the others, because a brand's voice, policies, and product facts can't leak across the wall. Segments enforce that separation deterministically, so the same agent can serve many brands without any of them ever borrowing another's answers. No cross-brand bleed, no separate agent to maintain per brand.

Internal knowledge versus customer-facing knowledge

Support teams often want their agent to draw on internal runbooks when it's assisting a human agent, but never when it's talking to a customer. A Segment tied to whether the conversation is in copilot mode keeps the agent-facing material available where it should be and invisible where it shouldn't. The customer never sees the internal note; the agent always does.

Intent-based routing

When the platform can tell what a user is trying to do, a Segment can send that conversation down a purpose-built path. Create-account intent and recover-access intent look similar in the first message but need very different handling. Segments let the agent commit to the right workflow with a deterministic rule rather than a hopeful guess.

Why “before the model reasons” is the entire argument

This is where Segments connect to something larger. Maven's agents reason over a Graph of Record, the unified picture of a customer, their history, and the systems of truth behind them. The graph holds everything the agent could potentially use. A segment decides which slice of it is in play for this conversation, before reasoning starts.

Think of it as access control for intelligence. The Graph of Record is the knowledge; Segments are the deterministic definitions of who gets which part of it, and when. That division of labor is what lets a single agent overlay an entire enterprise without becoming a liability: it can know a great deal in aggregate while staying strictly scoped in any given moment.

It's also why this fits the overlay model rather than fighting it. Maven sits on top of the helpdesks, CRMs, and knowledge sources a company already runs. Segments are how that breadth gets governed instead of dumped into every conversation indiscriminately. The agent connects to a lot and is permitted very little at a time, on purpose.

Why this matters for autonomous resolution at scale

Autonomous resolution is the goal: the agent doesn't just route or deflect, it resolves the issue. But resolution you can trust depends on the agent reaching for the right knowledge and the right action every time, not most of the time. Mastermind reached 93% autonomous resolution within six weeks on this kind of foundation. A high resolution rate is only safe when the agent is reasoning over a tightly scoped, correct set of inputs.

The stakes are concrete. Get scoping wrong and the failure isn't abstract: it's a customer handed the wrong policy, a banned account treated like a new one, a brand answering in another brand's voice. Each of those is a small erosion of trust, and trust is the thing an AI agent is supposed to protect. Segments are the unglamorous, but critical, reason that does not happen.

The takeaway

The most reliable AI customer service isn't only the product of a smarter model. It's the product of disciplined control over what the model is allowed to see. Segments are that control: the definition that was already made, before a single word is reasoned, about who your agent is talking to and what they’re there to know. Your customers will never ask about them. They'll just keep getting answers that fit.