Sit through a composable CDP pitch and the framing is set for you before any real question gets asked. On one side: the packaged CDP, described as old, rigid, a closed box that duplicates your data and charges you for the privilege. On the other side: the composable CDP, described as modern, open, cheaper, the obvious direction. The choice is presented as already made. You just have not caught up yet.
That framing is a sales tool, not a decision. The honest version is messier and far more useful. Composable is not automatically modern, and it is not automatically cheaper. For a large share of teams it is the wrong call, and the reason has nothing to do with how forward-thinking they are. It has to do with five concrete factors: how much data-engineering capacity you have, whether you genuinely need real-time, how mature your warehouse is, how hard speed-to-value is being pushed on you, and who is supposed to own the data. This piece lays out what each model really is, the tradeoffs nobody puts on a slide, and a decision tree you can walk for your own organisation.
Two models, plainly
Strip away the marketing and the difference is about where customer data physically sits and who assembles the pieces.
A packaged CDP is one vendor's product that does the whole job. It collects behavioral events, stores a copy of the customer data in the vendor's own platform, resolves identity into unified profiles, and gives marketers a place to build segments and push them to channels. Collection, storage, identity, segmentation, activation: one contract, one interface, one company to call. Salesforce Data 360, Adobe Real-Time CDP, Treasure Data, and Tealium AudienceStream sit here. The deal is simple. You hand the vendor the hard parts and pay for them to be solved.
A composable CDP, also called warehouse-native, does not store a separate copy. It leaves customer data in the cloud data warehouse you already run, Snowflake, BigQuery, Databricks, or Redshift, and adds the CDP functions as a layer on top. Identity resolution, segmentation, and activation run against the data where it lives. The activation step uses reverse ETL: traditional pipelines pull data into the warehouse, reverse ETL pushes the modeled result back out to email tools, ad platforms, and the rest. Hightouch is the best-known name; Census, acquired by Fivetran in 2025, and RudderStack work in the same space. The deal here is different. You keep control of the data and assemble the stack, and you carry more of the work yourself.
The term composable is recent. Reverse ETL as a labeled category dates to roughly 2020, and Hightouch built the composable CDP pitch on top of it, with a rallying cry that amounted to telling friends not to buy a CDP at all. So the model is young. That matters, because "modern" gets used as if it settles the argument. It does not. Newer is not a tradeoff. It is a date.
The tradeoff, honestly
Every real choice here comes down to one exchange: control and cost on one side, speed and self-sufficiency on the other. Both directions are defensible. Neither is free.
Packaged buys you speed and a single throat to choke. The vendor handles pipelines, identity logic, uptime, and the upgrade path. A marketing-operations team can run it without a data engineer in the room. The price is real money and real dependence. You pay a platform fee, your data lives in someone else's system, and the vendor's roadmap is now partly your roadmap.
Composable buys you control and, in theory, lower licensing. The data never leaves your governed warehouse. You are not paying a vendor to re-store information you already store. You pick best-of-breed pieces and swap any one of them. The price is that you now own the assembly and the maintenance, and that is not a small line item. It is the line item the pitch leaves off.
Two myths deserve to be retired here.
The first is that composable is always cheaper. License fees can be lower. Total cost often is not. Running CDP workloads against the warehouse generates compute charges, because building audiences and journeys means querying large tables, and warehouses bill for that compute. A composable stack also tends to be several products with separate contracts: warehouse, transformation layer, reverse ETL, sometimes a separate identity tool. Add the people, and the picture shifts hard. Composable vendors counter that warehouse compute keeps getting cheaper and that good tools cap runaway queries, which is fair. But the composable advantage is real and not automatic. When you net out build cost, compute, and headcount, composable frequently carries the higher three-year cost, not the lower one.
The second myth is that packaged means locked in and behind. The packaged vendors watched the composable pitch land and absorbed it. Most now offer zero-copy or warehouse connectivity, querying and activating data in Snowflake or Databricks without duplicating it. David Raab, who named the CDP category, went as far as arguing in late 2024 that "composable CDP" is dead as a distinct category, because packaged vendors split their products into modules and the once-clean line dissolved. You do not have to agree fully to take the point: the binary is softer than the slide claims.
What actually decides it
Here are the five factors that should drive the choice. Walk each one for your own organisation before you let any vendor walk it for you.
1. Data-engineering headcount. This is the single biggest filter, and it is the one the pitch is quietest about. A composable CDP is warehouse-native, which means warehouse work, modeling tables, maintaining pipelines, monitoring syncs, fixing them when they break, is your work. Industry estimates put the ongoing need at two to three data engineers, sometimes more, dedicated to keeping a composable stack healthy. If you have a data team that already runs a warehouse and maintains dbt models, that capacity may exist. If your martech is run by marketing operations with little or no engineering support, it does not, and a composable CDP will quietly become a stalled project. There is a reason for the gap between hype and behavior: an Acxiom survey found only 3 percent of marketers planned to move to a composable approach within twelve months, against 17 percent moving toward an integrated suite. The hype is loud. The hiring is not.
2. Genuine real-time needs. Be precise about what real-time means for you, because the word is abused. Reverse ETL is batch by design. Many syncs run every fifteen minutes to several hours, so there is a lag between a customer's action and the moment downstream tools learn about it. For a cart-abandonment flow that lag is usually fine, you want to wait a while before sending that email anyway. For in-session website personalization, or an AI agent that must act, see the result, and adjust within seconds, batch lag breaks the use case. Composable vendors have answers, Hightouch caches subsets of data for in-session work, and rightly note that many teams confuse near-real-time with real-time when they do not need true real-time at all. So the test is honest scoping. If your important use cases tolerate a few minutes, composable is fine. If they truly need single-digit seconds end to end, packaged platforms built for streaming have the easier path.
3. Warehouse maturity. A composable CDP rests entirely on the warehouse beneath it. If you already run Snowflake, BigQuery, or Databricks, the data is reasonably modeled, and someone owns it, the foundation is there. If you have no warehouse, or one that is half-built and poorly governed, composable does not skip that work. It just makes it a prerequisite. And a CDP of either kind will not clean your data for you; it unifies whatever it is handed. Composable on an immature warehouse means you are buying a CDP and a warehouse program at once, and the warehouse program is the longer of the two.
4. Speed-to-value pressure. How hard is someone pushing for results? Packaged CDPs are sold as fast, though the CDP Institute's own milestone guidance is sober: figure on a couple of months to a working pilot and six to twelve months to full deployment. Composable timelines vary more by team. With a mature warehouse and a defined use case, a composable build can reach a first activation in three to four months. Without that foundation it runs longer, because the foundation has to be built first. The point is not that one model is always faster. It is that packaged moves at a more predictable pace because the vendor owns more of the path, while composable speed is a function of the warehouse and the engineers you already have. Under heavy speed-to-value pressure with a thin data team, packaged is the safer bet.
5. Who owns the data. This is part principle, part risk. With packaged, a copy of customer data lives in the vendor's platform, governed partly by their controls. With composable, data stays in your warehouse, inside your governance, and that is a genuine advantage for organisations with strict privacy requirements or a firm policy that customer data does not leave the house. But composable adds its own wrinkle. Every reverse ETL sync pushes data out of the warehouse to downstream tools, so personally identifiable information still crosses vendor boundaries at the activation step. A composable stack of four or five tools also means PII passing through four or five systems, which is more surface area for a compliance review, not less. If your security team has a strong opinion about where the master copy of customer data sits, that opinion may settle this factor on its own.
The decision tree
Walk it in order. Each step narrows the answer.
Start: do you operate a cloud data warehouse today, reasonably modeled and owned? If no, your near-term answer is packaged. You can revisit composable once a warehouse exists and is healthy. Building one only to host a CDP is a far larger project than the CDP itself.
If yes, do you have data engineers, two or more, who can take on CDP pipeline work as an ongoing responsibility? If no, lean packaged. The architecture can be elegant, but with nobody to run it the elegance does not survive contact with production.
If yes, do your priority use cases need true in-session or sub-ten-second activation, or do they tolerate a few minutes of lag? If you need genuine real-time everywhere, scrutinise composable hard or lean packaged with streaming. If a few minutes is acceptable for most of what you do, composable stays in play.
If composable is still standing, is anyone forcing a result in under three months? If yes and your warehouse is not already CDP-ready, packaged will hit that deadline more reliably. If timelines are sane, continue.
Final check: does your security or governance team require that the master copy of customer data never leave your environment? If yes, that pushes toward composable, with eyes open about PII still moving on every sync. If there is no such hard line, decide on the first four factors.
If you cleared every gate, composable is a strong, deliberate choice rather than a fashionable one: a mature warehouse, real engineers, tolerance for some latency, sane timelines, and a control requirement that rewards keeping data in place. If you tripped on the first two gates, packaged is not the dated option. It is the correct one for your situation. And the tree allows a middle path: start packaged to get value moving, build warehouse and team maturity in parallel, and move composable later when the gates actually clear. The decision is not permanent, and treating it as a one-way door is its own mistake.
Where this is heading
The packaged-versus-composable line is blurring, and a third option is rising behind both. Most packaged platforms now offer warehouse connectivity, so the question is shifting from "copy or no copy" toward how well a vendor activates data in place. The 2026 Gartner Magic Quadrant for Customer Data Platforms marked the moment: Hightouch, a composable vendor, landed as a Leader on its first appearance, a clear signal that warehouse-native architecture has reached the enterprise mainstream rather than the fringe. The CDP Institute reports that more than one in four CDPs now support warehouse-native architecture.
The third option is the agentic CDP. Both Gartner and Forrester now frame the CDP's future around AI agents reading the customer profile and acting on it. That raises the stakes on the real-time factor above, because an agent that must act, observe, and adjust in seconds is the use case batch reverse ETL handles worst. Some warehouse-native vendors, Hightouch among them, have repositioned around agentic marketing for exactly this reason. The likely near-term picture is not packaged or composable winning outright. It is the architecture mattering less than two things underneath it: a well-governed warehouse and an honest answer about latency.
None of which removes the need to choose well now. The cost of choosing wrong is documented. eMarketer, citing CDP Institute survey data, reported that only about 64 percent of deployed CDPs deliver significant value, and that figure has been sliding. A meaningful share of that disappointment is an architecture decision made on vendor framing instead of the five factors. Run your organisation through the tree honestly, and you will know not just which model fits, but why. That "why" is the part the pitch never gives you, and it is the part that has to be yours.
Council summary
This post argues that the packaged versus composable CDP choice is sold as a settled story and is not one. The real decision rests on five concrete factors: data-engineering headcount, genuine real-time need, warehouse maturity, speed-to-value pressure, and who must own the data. We verified the load-bearing claims against primary sources: the Acxiom survey finding 3 percent of marketers planning a composable move against 17 percent toward a suite, David Raab's October 2024 argument that "composable CDP" is dead as a category, Hightouch's first-time Leader placement in the Gartner Magic Quadrant for Customer Data Platforms published in early 2026, the CDP Institute finding that more than a quarter of CDPs now support warehouse-native architecture, and the eMarketer figure that roughly 64 percent of deployed CDPs deliver significant value. The reader takeaway is plain: walk the decision tree honestly, and packaged is often the correct call rather than the dated one.
Comments