A CDP is the Best Contextual Data Layer for Agentic Systems

The power of quality, consented data in powering reasoning grade memory for agent decision-making.

2025 is all about agents. Oh, and vibe coding. I am still waiting for vibe agents, but maybe that's a 2026 thing. Let’s stick to just agents for now.

There are sales agents that score leads. Support agents that draft replies. Growth agents that buy media while you sleep. The space is too fast to keep up with. The best tools are changing weekly and hype cycles… Well, hype cycles continue to cycle. What I’ve noticed, trying desperately to keep as up to date on the topic, as much as a non-agentic human can, is that one foundational aspect does not change. Context. 

Research from the MemGPT project (an open source framework for building stateful agents with advanced reasoning capabilities and transparent long-term memory), now part of Letta AI, found that structuring an agent’s memory into separate blocks improves how well it responds. So, one block might hold a user profile with key traits and preferences. Another might store the agent’s own persona with its role and guidelines. 

By keeping critical details in fixed context slots, the agent avoids losing important information and does not have to reprocess the entire conversation history each time. This focused injection of context produces more consistent understanding and allows the model to reason more effectively than when it is given an unstructured, full record for every query.

The reason this works so well is actually quite simple. Context tells an agent who the person is, what just happened, and what it is allowed to do next. Miss that, and an assistant emails someone who opted out, recommends the wrong SKU, or contradicts finance rules. Nail it, and the same assistant fixes a failed payment, picks the right channel, and explains why. Context becomes state. It provides an accurate picture of the user’s wants, right now. In current state. 

An emerging use case for a Customer Data Platform (CDP), is to act as the contextual data layer providing agentic systems with good data. In fact, we’ve seen the early versions of this paradigm operating in production globally, right now. FOX used this contextual data layer to grow visitors to its “For You” feed by 376%. Creative Market cut data issue detection time by 93%. IBM increased billable usage on IBM Cloud by 17 percent and product adoption by 30%. It’s been used in various machine learning based systems for years. 

The same underpinning capabilities of a CDP unlock these abilities. However, building this layer is not an easy feat. Delivering effective context to the agentic system relies on an infrastructure for collecting, validating, profiling, activating, and enforcing data. We at Twilio Segment believe that a CDP is a great place to build a contextual data layer and that Segment CDP is a phenomenal option for this infrastructure. 

Here’s what you need to do if you’re interested in building the contextual data layer. I’ve separated out the foundational work, from Segment capabilities, to separate out my arguments for a CDP being the contextual data layer, and Segment CDP, being a great option if you choose to go down this path. 

Building the foundations

Start at the edge

Building the contextual data layer starts with a fundamental understanding of what a user is doing. To do so, you need to agree on an event collection system and framework. It's important to write a clear list of the events that an AI agent might want to know (the context), in order to understand what a customer has done previously. We recommend writing a clear list of events you collect. For each event, list the fields it must have. An example might look like this: 

Check every event as it comes in. If a field is missing or the type is wrong, block it and trigger alerts in the moment to the team that can resolve the issue. Map near-duplicates to one name. For example, order_complete, OrderComplete, and Order_Complete all become Order Completed. 

Ensure that bad events do not get stored. Even better if they can get fixed. It's these clean inputs that will save the data teams weeks of cleanup and keep your agents from acting on junk data. 

If you're a Segment CDP user, Segment’s Connections SDK automatically sends these events to the purpose-built Protocols functionality to check them in real time and quarantine anything off spec. As a user, you get a violation report you can act on and Segment can transform the event, so that bad data never arrives as context. 

Resolve the human and keep their choices with them

Once events are flowing, anchor them to a person. Store a live profile for each customer so your agents can answer two questions before they act. 

1. Who is this? 

2. What are you allowed to do? 

Put the identifiers you trust in one record, for example, a user_id, an email hash, and a device identifier. Store consent on the same record, for example, consent_email and consent_sms. Keep the profile fresh as people log in, switch devices, or change their preferences.

If identity and consent aren’t in one place, agents act in seconds on data that’s minutes old. In other words, agents will end up moving faster than your data. With Segment, the Profiles functionality merges identifiers into one record. In fact, Segment has built a complete object around the consent model on a profile. Consent Manager captures preferences in app and on web, writes the flags to the profile, and integrations like OneTrust keep policy aligned. Put simply, your agent checks one place before it acts.

Optionally mirror to your warehouse 

If your company uses a data warehouse, treat your CDP as the place where the live story lives, and your warehouse as the longer, more curated view that your finance and analytics teams trust. We recommend letting the CDP assemble a real-time profile, then mirroring a small set of high-value traits into your warehouse on a scheduled sync. The goal is that the agent and the dashboard end up operating off the same data. 

Frequencies of syncs also matter. We believe it’s important to define sync schedules for the various types of data being synced. We recommend the following schedule:

• Profile updates in seconds

• Warehouse mirror of priority traits every 15 minutes

• Tier two traits, maximum 24 hours

This might be a challenge if you stitch collectors, queues, merges, and cron jobs across teams.

With Segment, Profiles builds the live record and Profiles Sync keeps Snowflake, Databricks, BigQuery, or Redshift current on a schedule you control, no custom pipelines. The payoff for agentic AI is simple, your agent reads context from Segment to act, retrieval-augmented generation (RAG) grounds on the mirrored state, and refunds, credits, inventory, and consent match everywhere.

Feed the agents traits, not lists

Once profiles are in place, it's time to decide what your agent actually needs to see before it acts. Building an effective contextual data strategy realistically starts with being comfortable with the fact that agents do not need whole tables. They need a short list of traits that change a decision. We recommend starting by writing five or six per use case and give them clear names everyone can use in a sentence. 

For support-based use cases, you might choose entitlement tier, last app version, last failed payment, preferred channel, and consent to email. For growth use cases you might choose lifecycle stage, churn risk, and discount sensitivity. 

Keep the list small so prompts stay readable. Getting this right is imperative. The goal is to avoid ad hoc exports that grow over time, big payloads that slow everything down, and prompts full of noise the agent has to parse. 

If you use Segment, the live profile already holds these traits. An agent can pull directly from the Segment Profile, or given that Segment orchestrates data activation across various platforms, any agent that operates on platforms like Zendesk, Braze, or Salesforce, for example, has the right data, in the tool it needs to operate in. 

A final note

Start small this week. Write a one-page event list your agent should know. Add consent fields to the profile and verify they show up in the tools that send messages. Pick five traits for your first use case and name them clearly. Set freshness targets you can hold. The recommendation is profiles in seconds, and priority traits in 15 minutes. Turn on validation so bad data never becomes context. Then ship one agent that reads the profile, checks consent, and acts in the permitted channel.

Segment helps you get there faster, but the pattern stands on its own. A consistent collection path, one resolved profile, a small number of traits, consent close to the decision, and a single place for agentic queries.

Your agents will either guess or know what to do. Choose to know. No customer will end up with an SUV in their driveway for a $1 because an agent said they could.