What Changes for CDPs in 2027

Your CDP vendor shipped an agent this year, or an MCP server, or both. The feature works. The question for 2027 isn’t whether it works. It’s what it does to how your team uses the platform, how the platform is priced, and what your consent and identity data has to look like for the agent to be allowed to act.

This is a follow-up to CDP vs CRM in 2026. That piece drew the line between the data layer and the workflow layer. This one is about what happens to the data layer when a third kind of user shows up: not a marketer, not an integration, but an agent with a token.

Five shifts. None of them is a prediction about a feature that doesn’t exist yet. Each is a consequence of something that shipped in 2026.

The agent is another user of your CDP

Fig 1 · A third kind of user · Marketer, integration, and now the agent, and what it needs from you

In 2026 the marketing platforms opened a second door to the account.

Bloomreach calls Loomi Connect its interface layer for AI agents. It exposes the platform as tools an agent can discover and call: customer data and identifiers, campaigns, automation scenarios, Asset manager content, audiences and segmentations, consent and frequency settings, and analytics. MCP is the only integration method, through a remote server. Most tools are read-only; the write tools create campaigns and scenarios as drafts and don’t send or activate them. Deletes work only on draft or inactive objects. The documentation says it plainly: there’s no automatic rollback, so you’re accountable for what your users and agents change (Bloomreach documentation).

Klaviyo went further in the same direction. On September 9, 2026 it announced that more than 260 MCP tools are callable from Claude, ChatGPT, or any AI system a team already uses. We covered what that means for a small team in the Klaviyo piece. Brevo released an MCP server in October 2025, in early access, for reading campaign performance and updating contacts from Claude, Cursor, or Perplexity.

Here’s what that changes. Until now the user interface was the permission system in practice. The person who knew where to click was the person who could act. A new colleague took weeks to learn which filter meant what and which scenario was safe to touch. An agent connected over MCP gets the same view of the account on day one, reads the schema as it’s named, and acts on what it finds.

“Knowing the tool” stops meaning knowing the clicks. It starts meaning knowing what the data means and having written it down somewhere the agent can read. That’s the first shift, and the other four follow from it.

Packaged or composable stops being a question

Two years ago the architecture debate was a choice: a packaged CDP that copies data in, or a composable setup that activates from the warehouse without copying. In 2026 both camps moved toward each other. Packaged CDPs exposed SQL and warehouse connections; Klaviyo put SQL in its data platform as a preview, Bloomreach Marketing (Engagement) exports its data to BigQuery. Warehouse-native vendors now market real-time activation and identity resolution, which used to be the packaged side’s argument.

In 2027 the question is no longer which camp. It’s which system is the source of truth for each field. The warehouse holds the record: orders, products, finance’s version of revenue. The CDP holds what the warehouse can’t: the identity graph across devices, the last ten minutes of behavior, the consent state as of now, and the ability to act on all three before the batch runs.

That split has to be written down per field, because the agent in the first section will read the CDP’s copy of the field and assume it’s right. When the CDP says a customer’s lifetime value is 400 and the warehouse says 650 because returns are only in the warehouse, the agent builds the VIP segment on 400. Nobody told it otherwise.

Consent and identity become runtime rules

Loomi Connect applies the account’s IAM permissions: a user without rights to PII sees masked values, and the agent inherits the same limit. That covers who may see what. It doesn’t cover what the data may be used for.

Consent isn’t a permission on the account. It’s data on the profile. Say a customer agreed to a newsletter in 2023, and that consent is stored as email consent. An agent asked to build a win-back campaign will find an email address with consent and use it. Is a win-back the same purpose as a newsletter? The agent can’t answer that unless consent is stored per purpose. The rule for which purpose covers which campaign also has to be written as a filter, not kept in someone’s head.

In 2027 consent becomes a constraint the agent reads at the moment it acts, not a checkbox a person remembers. The same goes for identity. If the same person exists as two profiles because the app and the shop use different IDs, a human marketer notices the duplicate in the audience count and shrugs. An agent counts two people, sends two messages, and reports two conversions.

The regulatory side is in motion too. The European Commission’s Digital Omnibus proposal, tabled in November 2025, would have moved cookie consent rules into the GDPR. The Council’s negotiating mandate of 21 October 2026 keeps them in the ePrivacy Directive and sets a four-month wait before consent can be asked for again after a refusal. Parliament hasn’t adopted its position, trilogues haven’t started, and none of it is law. Don’t build on it. Build per-purpose consent regardless, because the agent needs it whether or not the regulation lands.

The bill moves to usage

Fig 2 · From flat fee to metered · Agentforce pricing before and after Flex Credits

Loomi Connect carries no extra charge until what Bloomreach calls the Credit Start Date, in the first half of 2027. After that, jobs draw from a monthly credit allowance, and the documentation leaves room for paid tiers later. Salesforce set the pattern a year earlier: Agentforce moved from a flat $2 per conversation to Flex Credits, $500 for 100,000 credits, 20 credits per agent action, so $0.10 every time the agent does something.

The pattern is the same across vendors: compute is metered, and an agent is compute. An agent that re-reads a profile every minute to decide whether to send is a cost line. A poorly written analytics query run by an assistant on a schedule is a cost line. The event volume that was a storage question in 2025 is a billing question in 2027.

A CDP cost model has lines for license, implementation, operations, and internal time. It needs one more line now: agent and AI actions, with a baseline measured before the meter starts. The teams that know their monthly event, API, and query volumes today will read the 2027 credit allowance as a number. The teams that don’t will read it as a surprise.

Decisioning moves next to the profile

You can see the direction in the products. The decision about what to send to whom, and when, is moving out of the campaign builder. It’s moving into something that reads the profile at the moment of the request. A prediction score, a scenario with a condition, an external decision engine, an agent with a brief.

That makes the CDP the place where everything a decision needs comes together: identity, eligibility, recent context, constraints like frequency caps and quiet hours, and the outcome event that tells the next decision whether the last one worked. Whether the CDP is also the place where the decision is made is a separate question, and a bigger one.

What to do in the last quarter of 2026

Four things, each an afternoon to start and a quarter to finish.

Name the events and write the dictionary. The agent reads the event taxonomy as it’s named. “purchase” and “order_completed” existing side by side was an annoyance for a human and is a wrong segment for an agent. Prune the taxonomy, then document what each event and property means in a place the agent can be pointed at.

Store consent per purpose. Not per channel. Write the rule that maps each campaign type to the purpose that covers it, as a filter in the platform, not as a note in someone’s head.

Decide the write scopes. Loomi Connect creates drafts; someone has to approve them. Decide who, decide which tools may write at all, and turn the rest off. Treat the agent’s token like a new employee’s login: least access first, more when earned.

Measure the volumes now. Events per month, API calls per month, analytics queries per week. Put the numbers in the TCO sheet before the meter starts, so the first credit statement in 2027 compares to something.

How we prepare clients for this

At Asteroad we spent 2026 doing the same cleanup for clients who wanted the AI features: event names, consent purposes, one profile per person. The checklist is in AI readiness and it hasn’t changed. What changed is the reason it matters. In 2025 a messy taxonomy made reporting slow. In 2027 it makes an agent confidently wrong, and the meter runs while it’s being wrong.

The takeaway

The platform didn’t change its job in 2026. It still holds the unified profile and makes it available to other systems. What changed is who the other systems are. One of them is now an agent that reads your account the way a new hire would, with no onboarding and a token that costs money per action. Everything on the list above is the onboarding you’d give that hire, written down once so every agent gets it.

Read next

• CDP vs CRM in 2026. The data layer and the workflow layer, before the agent showed up.

• Klaviyo’s autonomous CRM: what Composer and Headless mean for a two-person team. The same shift on one platform.

• AI readiness: what your customer data needs first. The cleanup list, platform-neutral.

• Event taxonomy pruning: fewer events, better data. The first item on the list, as a method.