Customer Experience Matrix: May 2025

Saturday, May 24, 2025

Testing the STIB Innovation Diffusion Model

The Substitution-Transformation-Infrastructure-Business Model (STIB) model of innovation diffusion that I presented in my last blog post is intriguing, but does it reflect reality? One measure is whether it provides a useful classification of industry products. To test this, I looked at the past ten weeks of product news in the CDP Institute Daily Newsletter, which came to about 50 items. Results were quite interesting.

But before we get to that, let’s flesh out the model a bit. It describes two core elements:

Innovation: the development whose impact is being measured. This might be electric motors or internal combustion engines. (More formally, innovation is a change in technology, where technology is defined as the tools, methods and knowledge used to perform a task. "Technology" in this sense isn't always "technical": a new marketing methods and financial tools can be important innovations. But I digress.)

Application: the process or product changed by the innovation being studied. This might be manufacturing in factories (a process) or transportation vehicles (a product). The application is assembled from components, which might be physical objects in a product or steps in a process workflow. The same innovation might affect many different applications but it's often more useful to study just one.

The model describes how an application changes in response to an innovation. This change comes in three stages:

Substitution: the innovation is deployed as an exact replacement for one component in the applications, without changing anything else. Electric motors were hooked to turn factory shaftwork instead of water wheels; internal combustion engines replaced horses for powering carriages.

Intermediate States: this is a period of experimentation. Multiple components are changed as the industry explores how to best take advantage of the innovation. The innovation itself continues to mature during this period, opening up additional possibilities.

Transformation: the industry settles on an optimal design incorporating the innovation. This "final" design is fundamentally stable although incremental improvements continue. Factory tools incorporate built-in electric motors; motor vehicles locate the engine and drive train most efficiently.

The model also tracks changes over time to technologies beyond the application itself:

Infrastructure: this includes products, suppliers, and processes that support the transformed application or are changed as a result. Electric motors require generating stations, power grids, and new machine tools, while they result in more efficient factory designs. Parts and materials manufacturers support motor-driven vehicles while auto ownership spawns new roads, filling stations, employment opportunities, legal frameworks, and more.

Business model: this describes business methods such as revenue sources, pricing models, marketing, distribution, customer service, funding, and ownership models. Larger factories lead to scale economies that favor large, national brands. Complex auto technology favors large organizations with deep resources for research, mass production, national advertising, and dealer networks.

The graphic illustrates relationships among these elements.*

In the initial state, the application, infrastructure and business model are all in their baseline (i.e., pre-innovation) configuration.

This is followed by substitution, where the innovation is applied to make a slight change in the application while infrastructure and business model are unaffected.

In the third stage, intermediate states, developers try variety of changes to the application and the infrastructure starts to mature in response to growing demands. Some will be a dead end (direct current electrical systems, steam powered cars). Others may hit on the transformed design before infrastructure is available to support them. Vendors may also fail if they don't change quickly enough to keep up with competition.

In the final stage, developers converge on a transformed configuration and application change comes more slowly. Some companies drop out or merge as the market consolidates. The infrastructure continues to evolve and new business models start to emerge.

The Future of Customer Management

Applying the model to a new topic requires specifying the application being changed and the innovation that is changing it. This might be tricky in some situations, but it’s pretty simple in the current context: the application is customer management while the innovation is artificial intelligence.

What’s not simple is predicting the form of the transformed application. Hindsight about factory equipment or auto design is easy, but we don’t know how customers will be managed in the future. This prediction isn't strictly necessary: we can simply observe changes as they occur without assessing their relation to the final state. But that assessment helps to organize the analysis and gives buyers and developers decide where to invest their resources.

Predicting the transformed state is a thought experiment: identify the constraints imposed by the current technology and imagine what the ideal design would look like if those constraints were removed (and replaced by whatever constraints the innovation imposes.)

As I see it, the dominant feature of customer management today is the number of separate steps that are performed by different people. There’s a creative team with a multi-step process to develop content, a targeting team with a multi-step process to select audience segments, a media team with a process to buy advertising, and multiple operations teams with multi-step processes to deliver messages via websites, email, social media, connected TV, digital video, games, out-of-home, podcasts, and elsewhere. The reason we have so many teams with so many steps is the limited capability of individual humans: each can be expert in only a narrow area, so many must work together to deliver a complete result. And, because they’re human, each person can only deliver a relatively small number of outputs over time. Thus, each output must apply to many people to deliver messages to everyone.

AI removes this limit. At least in theory, a single AI agent can expertly execute all steps in the messaging process, combining content creation, audience selection, media buying, and delivery. In practice, a single super-agent is less likely than a master agent that calls on specialized subordinate agents. But so long as those agents complete their work almost instantaneously, the process will still function as a single step. Ideally, these agents would view all available data and collaborate with each other: content development would be informed by audience characteristics; media buying would take into consideration other channel opportunities; and so on. All these decisions would be coordinated to produce the highest total value: for a given advertising impression, the system(s) might simulate the results of several different creative treatments for several different offers for several different products, and then select the best creative/offer/product combination – or conclude that even the highest possible value for that impression is too low to justify the investment. This requires a sophisticated value prediction algorithm which would consider long-term as well as immediate results.

In addition to speed and coordination, the process would be completed at near-zero incremental cost.

In short, I see the transformed state as hyper-personalized, financially-optimized treatment of each customer during each interaction across all channels.

This is a radical change. (That's why they call it a transformation.) There are no more campaigns (defined as a predefined sequence of messages), no more audiences (defined as a group of customers who receive the same messages), no more content (defined as a fixed combination of text, images, and offers), and no more media buys (defined as group of impressions purchased together). What we have instead is a master agent orchestrating customer interactions across all channels to optimize use of budgets, data, customer attention, products, staff, and computing power. That agent would call internal and external data to guide its actions and would deliver messages across both internal (owned) and external (paid) media.

In the transformed world, today's creative, analytics, media, and operations departments have largely vanished, apart from a few human(?) experts left behind to monitor the AI. On the other hand, there might be more need for humans to do strategy and product development.

This vision also points to new infrastructure and business models:

Supporting infrastructure would include vendors building the base AI systems, which are too complicated for most companies to build for themselves (although non-technical users might use no/low-code tools to tune them); data tools to prepare, integrate, and expose data from all sources in real time; and analytical tools to measure interaction value.

Surrounding infrastructure would include new media that provide real-time access to their customers; low-friction integration methods to connect marketers with these systems; new billing and analytical methods; and social/legal frameworks to govern customer data collection, sharing and privacy.

New business models would be needed for companies and the systems, data, and media providers that support and surround them. With each interaction managed separately, providers might become on-demand services that charge for value provided in each case rather than billing based on labor, system use, or impressions.

I’m not sure this will happen. At most, I’d bet one of my later-born children on it. I’m laying it here out because the STIB model works best if you measure against a transformed state (and because it’s fun to think about.)

Mapping against the STIB Framework

Application

That said, let’s try mapping recent product news against the STIB framework. The application we’re analyzing is customer management, which is roughly what we cover in the CDP Institute Daily newsletter. This gives us a reasonable collection of announcements to work with, although I should stress it’s just a small sample based on items that happened to appear during a relatively brief period.

If the STIB model is correct, we should find clusters of products on a spectrum from business as usual, to substitution of AI within current processes, to complete transformation. Because transformation doesn’t happen all at once, we would expect to find several intermediate stages. Indeed, that’s the case.

Current State: The first cluster would hold products that execute the existing process without any change. AI-powered co-pilots might fit here. But co-pilots are so common that we don’t bother to write about them in the newsletter. Nor do we usually cover products that aren’t doing anything new. So this cluster is empty apart from one item about agentic helpers.

CommerceIQ Offers Agentic Teammates to Ecommerce Marketers

Substitution: The second cluster is simple substitution: products that replace a discrete task with an AI-generated equivalent. We see plenty of these, often offered as collections of (separate) AI tools for multiple tasks. We also see toolkits for users to build their own. So long as each AI tool executes one task separately from the others, this is still substitution. Recent examples are:

Intermediate Products: Now we move into products that change the underlying process but don’t reach the fully transformed state. The news items seem to fall into three clusters. It’s important to note that we didn’t define these in advance: they have emerged from the data itself.

Content Creation: This is a popular task to automate. It’s a single task but more than substitution because most vendors connect their content generator to response data and use this to automatically optimize content over time.

Goal-Driven Workflow: this cluster holds systems that have automated development and execution of a multi-step workflow, such as audience segmentation, journey design, or media buying. Like content creation, these usually let users specify a goal for the workflow and collect data to help measure results.

Customer Management: these products collect data to support optimization, help users to create messages, and select and deliver messages across customer touchpoints. They come the closest to the fully transformed state but don’t support all channels or create hyper-personalized messages in real time.

Transformed Products: This cluster would hold products that deliver the fully transformed process. There’s a good chance that some vendors have this mind, but we haven’t seen any products that deliver it. So the cluster is empty.

Infrastructure

Infrastructure and Business Models won’t take their final shapes until the transformed process is fully deployed, but they do co-evolve with the application changes. This applies especially to supporting infrastructure, parts of which must be in place for some intermediate product to function effectively.

Supporting infrastructures: The key supporting infrastructures for transformed customer management are inputs (data access and quality), internal processes (agent cooperation and analytics), and outputs (media integration). We see announcements in all these areas.

Data Access: most data access announcements describe accessing data in real time without loading it into a separate database. So far, these developments focus primarily on reading data in the company’s internal systems. But remember that the full vision for transformation includes access to third-party data such as compiled customer behaviors and local weather. We do see some of that, although none is in the current sample.

Data Quality: these tools prepare data for AI use. Many data quality vendors have added features to support AI use and have added AI-powered offerings. These apply to customer data but we don’t usually write about them, so this cluster is fairly sparse.

Agent Deployment: this describes technology for building agents and helping agents work together. It’s another field with extensive activity that is largely beyond the scope of the daily newsletter. It will be critically important if the transformed state involves teams of agents that cooperate closely with each other.

Analytics: any goal-seeking agent will need internal analytics to guide its decisions. Still, there may turn out to be a market for independent agents that make their results available as a shared resource for teams of specialist agents. This would apply especially to customer value analytics, which are needed to compare opportunities across different channels.

Hightouch Adds Journey Simulations

Media Integration: this is the output infrastructure that deliver messages created by the hyper-personalization. The tools are steadily encompassing more channels.

Surrounding Infrastructures: changes to the surrounding infrastructure happen after the transformed application is in place. This makes them harder to predict than supporting infrastructure, which develops sooner. We do have some current developments that are likely to become more important as hyper-personalization matures. There will no doubt be others.

Integrated Commerce (Sales Agents): this describes shopping in non-traditional channels, including retail media, search engines, social media, video, connected TV, and mobile apps. These can work without hyper-personalization but are vastly more effective when offers are tailored to the individual and context. In many cases, the interface will be a chat-style, AI-based sales agent that has a real-time dialog with the customer.

AI-Based Search (Buyer Agents): this describes marketing that is targeted at AI agents rather than humans. Today's most common implementation is AI search overviews, which do research for buyers. This changes the goal of search marketing from attracting traffic to appearing in genAI summaries. Other agents are evolving for other types of research and other stages in the sales cycle including making purchases on the customer's behalf.

Interacting with these buyer agents at scale and cost-effectively will require AI-based sales agents. In some ways, it won’t matter whether the sales agent is interacting with a person or a bot: to the sales agent, both are collections of data that must be analyzed and responded to appropriately. That said, the behaviors of humans and bots will be significantly different, so the sales bots will no doubt develop separate approaches to each. Ultimately, marketing systems may give buyer agents direct access to (some of) the product and promotional information they provide to sales agents, bypassing the sales agents altogether.

Because this is a relatively new development, our news coverage includes more research reports than product announcements.

Business Models

Few companies make announcements about changes in their business model, especially when those changes involve firing people. As a result, we have few newsletter items on AI-based business models. One change we do see is movement towards pricing based on value created rather than resources consumed. The transformed process will surely create other opportunities, perhaps including “no staff” companies run almost entirely by AI or “no product” companies that source products in real time as they find interested customers. It's likely the new business models will include ones we haven't even imagined.

What We’ve Learned

The main purpose of this blog post is to determine whether the STIB model provides a useful way to think about technology innovations. Of course I’m biased, but I believe it does. Not only was it reasonably easy to slot products into different categories, but the process generated several helpful insights:

Products that substitute AI for one step within an existing workflow are significantly different from products change the workflow.

It helps to envision the final, fully transformed product so we can assess intermediate products, identify the supporting capabilities and infrastructures it requires, and predict the surrounding infrastructure and business models it is likely to generate.

We should acknowledge that our prediction could be wrong. It may help to build scenarios around alternative outcomes.

Intermediate products appear before the final transformation. These become possible as new capabilities, infrastructures and business models appear. Observing the intermediate products helps to assess whether the changes are moving in the direction we expect or to adjust our prediction of the transformed state..

A minimum set of capabilities, infrastructures, and business models must be in place before the final design can succeed. Companies will fail if they offer the transformed design before supporting infrastructures are in place.

Capabilities, infrastructure, and business models continue to evolve after the (successful) transformed design is introduced. These developments make the product more effective and exploit the opportunities it creates.

Continued evolution creates rapid growth and expands the value of the final design, giving it an increasing advantage over alternatives. This advantage ultimately locks other configurations out of the market, even though some may be technically superior.

Applying the STIB model to customer management offers additional insights. Our sample of product news is enough to show:

Many current products do simple substitution. These are the easiest to deploy and can show immediate improvements in cost and quality. (There are also many products that support the existing workflow without any changes, but those don’t show up in our sample.)

Some intermediate products are already available. Most of these automate a single stream of tasks within the customer management workflow, such as content creation, campaign design, media buying or analytics.

Intermediate products usually work within a single department. This makes them easier to drop into the larger workflow and reduces the number of users whose work is disrupted.

Some products aim to automate an entire workflow, such as campaign design, development, and execution. Doing this with autonomous AI agents is the leading edge of the industry today.

Some companies offer components that support the final vision. These include application capabilities such as content optimization, response simulation, and advanced attribution, as well as infrastructure and business model changes including agent coordination, cross-company data sharing, touchpoint integration, and performance-based pricing.

I haven’t seen any products that promise the predicted final state of “real-time, hyper-personalized, omni-channel messages.” This has surely occurred to many smart people, so I’ll guess they have decided (correctly) that the capabilities, infrastructure and business models aren’t ready yet.

The final state may be delivered by a single integrated product or by multiple agents working in concert. Remember that the final state requires close cooperation between advertisers and media companies and between different departments within the same company. This makes a multi-agent solution more likely and reinforces the needed for data- and process-sharing technology..

Implications

Here are some practical implications of what we’ve discussed.

Substitution products can be valuable, but the benefit won’t last. It’s tempting to argue that substitution is a poor investment because it offers only incremental improvements on today’s current processes and that companies, and software developers should instead focus on more profound solutions. But the plain fact is that substitution creates substantial benefits and is easier to deploy than changes that require process change. Product developers and business users shouldn’t shy away from substitution but they do need to recognize it has a relatively short shelf life. Product developers should also realize that nearly all incumbent vendors will add substitutions to their current products or have already done this. That makes it hard to convince users to change to a new system on the basis of substitution alone. Attracting new clients will require solutions with more substantial advantages, which is what the intermediate product designs can offer.
The goal of "real-time, hyper-personalized, omni-channel messaging" is not yet widely discussed. What leads me to expect this as the transformed state is the growth of direct sales in social media, search results, connected TV, podcasts, and pretty much every other channel. This "instant commerce" collapses the multi-step "awareness, interest, desire, act" cycle into a single moment when the customer is presented with an opportunity to buy. Taking full advantage of this moment requires marketers to connect with all message opportunities, to gather all data so they can assess the potential value of each opportunity, and to deliver the most effective possible message for each opportunity they purchase. Only AI can do this effectively at scale, most likely by presenting intelligent agents to interact with customers who engage.**
The transition can be gradual. Instant commerce can be deployed in one channel at a time, can work with limited data, and doesn’t require advanced message optimization. Companies and vendors can move towards the fully transformed state in stages, building experience, product, infrastructure, and business models along the way.
Instant commerce depends on long-term relationships. That may seem like a paradox but customers will only engage with companies that they trust. There’s no time to build trust during the interaction moment, so trust must be built in advance. The good news is that the importance of trust is widely recognized and the methods for building trust (among humans) are well understood, if not always well executed. The industry may need new lessons in building trust among AI agents.
Infrastructure may offer the greatest opportunities. The biggest gaps between what’s currently available and what’s needed in the transformed state seem to be deep connections to collect external data and interact with touchpoint systems.

Shallow connections already exist, but real-time, hyper-personalized, omni-channel messaging implies real-time queries of external data sources about individual customers to collect up-to-moment information on behaviors. The touchpoints where those behaviors occur must capture, identify, assess, expose, and charge for that data in real time in a privacy-compliant fashion. Bear in mind I’m talking about touchpoints outside the company that wants to use this data. Very little technology exists to do that today. Data clean rooms are a start.
Beyond sharing information, those same touchpoints need to receive ad messages and deliver them to their visitors, again in real time and with feedback on response. Most of the messaging technology will reside with the ad buyer or middlemen, who will need to receive notification of contact opportunities, gather data and assess those opportunities, select the appropriate message, bid on delivering the message, transmit the message on bids they win, and measure results. Again, the existing technology to do all this in real time, at scale, and across many channels is limited at best. (Today’s programmatic ad system is a partial model.) In addition to data movement, this process requires sophisticated evaluation models so marketers can accurately bid on the projected value of each interaction.
Because all this interaction is happening between different companies, the infrastructure technology must be widely shared. This could imply broadly accepted standards implemented by many developers, or, more likely, proprietary technology built and sold by a few major suppliers. Competition to be one of those suppliers will be fierce but the rewards are likely to be huge. The rewards might diminish over time once the process is well enough understood to create open standards that are a viable, cheaper alternative.

Data quality doesn’t get the attention it deserves. Survey after survey shows that data issues are the top roadblocks to marketing, personalization, and AI success. Yet companies rarely make data quality an investment priority. There isn’t much to say about this except that real-time, hyper-personalized, omni-channel messages make data quality more important than ever. This is yet another piece of infrastructure that’s ripe for improvement.
Buyer bots might change everything. Buyers have limited attention, which is why gaining attention has always been the first step in successful marketing. This still applies to real-time, hyper-personalized, omni-channel messages -- so long as they're being sent to humans. But if customers delegate their purchasing activities to bots, the fundamental truth is no longer true: buyer attention will no longer be limited. Developments such as search engine optimization and AI search overviews as early examples of marketing to bots: marketers must target the algorithms, not attract human attention. But search marketing of any kind is still aimed at putting messages in front of eyeballs. Truly automated purchasing will remove humans from the entire process. The change won't happen overnight but it seems plausible to expect a mix of human and bot buyers in the near future. A STIB model with bot buyers as the final transformed state would be quite different from the one I’ve presented here.

Summary

This post has explored whether current developments in customer management technology can be effectively analyzed with the STIB model of innovation diffusion and whether the results provide useful insights into industry trends. I believe the answers are Yes and Yes. I've also presented a specific vision for the industry future, of "instant commerce" delivered through real-time, hyper-personalized, omni-channel messages. I can't promise that this is correct, but it's an interesting starting point for discussion.

_________________________________________________________________________________

* The graphic is conceptual but could be quantified by counting the number of components in each product that match the final transformed design. Fun!

** Things are admittedly a bit more complicated for non-impulse purchases. But I’d argue you’re still trying to motivate an action in the moment, even if it’s only saving an offer in a wallet for future consideration.

Sunday, May 11, 2025

Predicting the Future of AI-Based Customer Management: Lessons from Machine Tool Motors

As someone who has spent his career tracking new technologies, I’ve long been fascinated by how those technologies spread over time. The standard model of technology diffusion, developed by Everett Rogers and based on New Deal studies of how farmers adopted new crops, offers the familiar segmentation of users into innovators, early adapters, early majority, late majority, and laggards.

It’s probably best known in the tech world in its Crossing the Chasm incarnation, which adds the all-important chasm between the second and third stages.

But there’s another model I find more helpful. This looks not at who adopts new technologies but at how they’re deployed over time. There's a consistent pattern:

Substitution (the new technology is deployed to exactly replace the old technology without changing the surrounding processes or workflows)
Transformation (the process is adapted to take advantage of the capabilities of the new technology)
Infrastructure change (new infrastructure is developed to support the transformed process)
Business model change (a new business model emerges to take advantage of the transformed processes and infrastructure)

I haven’t found any single source that presents this model in quite these terms, which may mean I could name it after myself, although for now I'll just use STIB. Some close matches include Ruben Puentedura’s SAMR Model for educational technology (substitution, augmentation, modification, redefinition), Deloitte’s Three Horizons Model (process optimization, process flow and quality, new business models) and Carlota Perez’s theory of technological revolutions (radical innovation, initial optimization, incremental innovations, maturity).

The paradigmatic example, at least for me, is deployment of electric power in U.S. factories. This started with connecting electric motors to the shaftworks that were previously driven by waterwheels or steam engines (substitution). After several intermediate changes, the endpoint was attaching a separate motor directly to each machine tool (transformation). This enabled an infrastructure change: no longer constrained by the limits of mechanical power transfer or proximity to water or coal, factories became cleaner, larger, single story, and located near other resources or markets. The new infrastructure resulted in a new business model featuring centralized mass production and standardized national brands.

Perhaps you can't relate to machine tools. Fair enough. Consider the automobile instead. The first cars were literally horseless carriages, pretty much the same as horse-drawn carriages except that they used a mechanical motor instead of a horse. That was substitution. Over time, the design of the cars was modified to take better advantage of having an internal combustion engine. That was transformation. Once it became clear the new model was going to be a success, manufacturers developed new, mass-production business models that better suited the new products and processes. Ultimately, this was supported by a new infrastructures of roads, filling stations, repair shops, traffic rules, licensing requirements, insurance products, and ultimately an entire, auto-based suburban landscape.

This may seem academic, but it applies directly to something you probably care about: the growth of AI. Seen through the STIB framework, most of the AI applications we see today are substitutions: an AI copywriter replaces a human copywriter in an unchanged workflow. The omnipresent co-pilots are another, even less disruptive type of substitution: they help humans perform the same tasks more efficiently, again without changing the workflow.

But it’s clear that these substitutions don’t take full advantage of AI’s potential. Again referring to the STIB framework, the question is what the transformed applications will look like. To answer this, we conduct a thought experiment: what would the process look like if it were redesigned to make best use the new technology? For electric motors in factories, the final form was motors attached directly to machine tools. For most customer data and customer-facing operations, the final form of AI is likely to be a single process that completes all the previously-separate steps at once. This is because AI is not limited by the need to have different specialists perform each step in the workflow, a constraint which results from the inability of mere humans to master more than one specialty, and from the need for experts to check the output of each step before moving on to the next.

But I don't think unified execution is the final form of the AI transformation. A deeper change is likely to remove discrete units such as customer segments, content pieces, campaign flows, and maybe even standardized products. Those exist because humans could only manage small numbers of segments, messages, campaigns, and products. An AI could handle more-or-less infinite numbers of these, which in practice would mean treating each customer and each interaction individually.

This leads to an end-state of “hyper-personalized” messaging, where content is custom generated on the fly for each customer and context.

Imagine an all-knowing, all-seeing bot that listens to what’s happening in the market and jumps into action each time it sees an opportunity to do something useful. The action will be optimized using all relevant data, including the company’s own information about the customer; second-party, third-party, and public information about the customer; behaviors of other customers; and market conditions, inventory, and who knows what else. In another dimension, this listening can extend beyond company-owned systems such as websites and contact centers, to include appearance of customers on third-party sites (already available to some degree through programmatic ad bidding) and even in walled gardens (which already receive lists of customers to watch for; the change would be to open a channel that lets the company assess the situation, generate the optimal message, and send it back for the walled garden for delivery.)*

It should immediately be clear that this vision requires infrastructure and business model changes from what’s available today. A much-improved data sharing infrastructure is needed to monitor behavior and access data outside of company-owned systems. This implies new business models to compensate external data owners for access to their information. Perhaps the data owners would charge a fee for letting companies monitor their data streams or query their data stores; or maybe they would only charge for data that a company uses; or perhaps the fees will be based on outcomes such as clicks or sales. Most of these schemes will ultimately require some way to estimate the value contributed by a particular piece of data.

Hyper-personalized message delivery requires more infrastructure and business model innovations. One of the most important developments in marketing today is the emergence of new channels that allow direct customer interaction: these include interactive TV, social commerce, online games, retail media, and even interactive podcasts and out-of-home advertising. All are alternatives to common web display ads, social media ads, and search ads, which are also becoming more interactive. As with data, the key change made possible by AI is the ability to monitor vastly more opportunities at once, to evaluate the potential of each opportunity in real time, and to take advantage of the opportunities offering the greatest value.

I certainly hope that everyone reading this realizes that what I’m describing is far beyond the capabilities of today’s AI systems. The data access process requires AI to continuously ingest, clean, and integrate data from multiple sources and to automatically adapt as new sources appear and established sources change. Remember that AI is just beginning to address the bottleneck of incorporating new data sources into today’s CDP and warehouse systems. Similarly, we’re just beginning to see AI systems deliver intermediate steps on the way to hyper-personalized messaging. Today’s cutting edge is automated campaign design, which at best (and with much-needed human quality checks) could transform a user’s prompt into a complete campaign package of audience selection, content, and delivery rules and then execute that package. While impressive, this still uses the conventional structure of a few, discrete segments, content pieces, and campaign flows. That makes it closer to substitution than true transformation.

Another way to look at this is that the vision offers a roadmap for future AI development. The current frontiers in AI are goal-seeking agents, access to external data (Anthropic’s Model Context Protocol), and agent cooperation (Google’s Agent-to-Agent). (See this Medium post for a good overview of these.) If my vision is correct (which is by no means certain), steps beyond those frontiers will include proactive data gathering and integration, automated data value assessments, greater situation awareness, and better simulation of human behaviors. (I'd really like to say "understanding" of human behaviors but don't think we can quite attribute that to AI.) AI will also need more economical processing and reliable guardrails against hallucinations, biases, privacy breaches, and generally bad behavior.

The table below offers a more detailed view of where I think things are headed. It looks at four major customer data processes: customer data management, people issues related to customer data management, customer data activation, and advertising. For each process, it lists the required capabilities for each of the four diffusion stages. You can think of these as requirements for new, AI-based products.

AI Applications for Customer Data (STIB Model)
	Data Management	People	Activation	Advertising
Substitute [execute via co-pilots and agents]	Data collection, ID resolution, connectors, metadata	Understand applications, requirements, training	Segmentation, analytics, prediction, sharing, privacy	Audience assembly, media buys, data buys
Transform [execute via unified AI systems]	Unified process, data as service	Management tools, define goals/prompts, explore opportunities	Hyper-personalize messages	Deliver best customer, data, channel/media; optimize spend
Infrastructure [required capabilities; many delivered via AI]	Automated data access, security, privacy, quality, transforms	Learning systems, training systems, process design systems	Efficient processing, attribution, instant commerce, buyer agents	Secure data sharing, consented data assembly, contextual targeting, marketplaces, fractional billing
Business Model [rely on AI for analytics and operations]	Value-based pricing	Pay for skill achievement	Goal achievement, Sales as a service	Goal achievement, audience as a service, value-based pricing

Where does all this lead? Here are the main points I hope you’ll take away:

1. The impact of AI on customer data is just beginning. We can expect AI to be deployed in the same pattern as other technologies. At first, it will substitute for humans or non-AI systems by performing the same tasks within existing workflows. Over time, it will transform those workflows into new processes that take full advantage of what AI can do. Ultimately, the industry will develop new business models and infrastructures to support the transformed processes.

2. For developers: consider which stage your AI project is targeting. While substitution is low-hanging fruit, bear in mind that current processes will soon be obsolete. Consider developing products that support transformed processes and their related business models and infrastructures.

3. For users: current self-service AI tools may enable you to build your own substitutions with minimal investment. Unlike system developers, you can afford to deploy these now and then discard them when something better comes along.

4. If you have a vision for a transformed process, you could try to build it. It won't achieve its full potential because the supporting AI capabilities, infrastructure, and business models are not yet available. But it might deliver enough value to justify the investment: imagine hyper-personalization based only on one company's own data and deployed only on the company's own customer-facing systems.

5. The final form of the transformed process will emerge over time. As companies experiment with different approaches, the industry will converge on an optimal design. Commercial developers will then build systems with this design and flesh out the supporting AI features, infrastructures and business models. As with most complex systems, commercial vendors will probably own most of the market because they can afford to invest more in their products than most individual firms.

6. Standardized integration mechanisms will play an important role in this new world. This assumes I'm right that the transformed process will rely heavily on connecting to external data and external delivery channels. This should make integration a particularly fruitful area for investment if you're a developer, or for building expertise if you're a user.

__________________________________________________________________________________

*As a practical matter, it's unlikely that thousands of companies would separately monitor thousands of data sources and delivery touchpoints. It's more likely that data sources and delivery systems will broadcast event information which marketers can read and react to as they see fit. Perhaps the sources and touchpoints will package their events into channels covering different customer groups or event types and let marketers subscribe to the channels they want.

Tuesday, May 06, 2025

State of Martech Report: Customer Data Platforms Are Evolving, Not Dying

Scott Brinker’s State of Martech report has grown from a one-page logo jigsaw to a small industry, co-authored with MartechTribe’s Frans Riemersma and apparently produced by a small army of elves reviewing thousands of products each year. The latest edition, released yesterday, provides the now-expected deep analysis of industry trends covering category growth, the impact(s) of AI, stack architectures, and more. It’s all interesting and too complicated (or maybe complex?) to recap here, so I can only suggest that you read it on your own.

In addition to data based on the 15,384 vendors now listed (up 9% from 2024), the report analyzes results from a survey of martech and marketing operations leaders. Again, there’s lots of fascinating information, but I was of course drawn to the sections related to Customer Data Platforms.

On its face, the report offers some pretty bad news for the CDP industry: the fraction of companies citing a CDP as the “center” of their martech stack fell from 15.5% in their similar 2024 survey to 12.5% in 2025. However, the 2025 survey is based on just 96 responses, meaning that’s a swing of three answers, so it’s not cause for too much alarm.* Still, it’s an interesting result, and even more intriguing when you separate B2B respondents (52% of the sample) from others (14% B2C and 34% mixed B2B/B2C). CDPs have never been widely adopted in the B2B world, and indeed, their share is unchanged from 2024 (7.9%) to 2025 (8.0%). The flip side to that is in the B2C and mixed group, the shift is larger: from 26.9% in 2024 to 17.4% in 2025. But, again, bear in mind the sample size: that 17.4% represents seven responses. A shift of three answers is well within the range of statistical noise.

Let’s put sampling issues aside and assume there’s some drop-off in reliance on CDP. The question is, what has taken its place?

Did you just answer “cloud data warehouse, of course”? That’s not entirely wrong – the data warehouse share grew from 20.9% to 23.9%. But the big winner was MAP/CEP (marketing automation/customer experience platform), which grew from 19.4% to 26.1%. CRM grew from 17.9% to 19.5%, or nearly as much as data warehouses. Multi-product suites fell from 1.4% to zero, which hints quite strongly that the respondents heavily skewed away from large enterprises.**

If we combine the MAP/CEP, CRM, and DXP or ecommerce categories into “customer-facing systems”, the combined share of that group grew from 43.3% to 52.1%. So if I were to read any trend from this data, it would be that companies are centering their martech stacks on customer-facing systems, not on data warehouses.

This is actually consistent with the trends we’ve seen in the CDP industry itself, where the most recent major acquisitions (ActionIQ by Uniphore , Lytics by ContentStack, mParticle by Rokt) all involved merging a CDP into a customer-facing product, and where customer-facing vendors like MessageGears, Klayvio, Insider, Listrak, and Braze have added CDP (or CDP-ish) capabilities***. The CDP Institute classifies all of these systems as CDPs, in addition to whatever else they do. So, the way I see it, companies that list those products as the “center” of their martech stack are still building their stack on a CDP, even if they don’t call it one. That’s good news for the industry, not bad.

The survey also takes another look at the role of data warehouses in the martech stack, asking “Do you have a customer data warehouse/lakehouse integrated with your martech stack?” More than half the respondents (56.2%) said they did, a number that climbs to 92% for B2C vendors and 80% for enterprise companies. What’s interesting here is the relation of those answers to the previous question: many more firms have integrated a warehouse than are using their warehouse as the center of their martech stack. This is far from shocking but, again, suggests that the mere presence of a warehouse doesn’t mean that warehouse is the primary customer data store.

If you’re a “warehouse-centric” composable CDP vendor, you could read the relatively small share of warehouse-as-central system either as cause for alarm – the market isn’t as big as you thought – or reason to rejoice – the potential for growth is huge. Both could be true at the same time. But if the primary industry trend is for CDP functions to migrate to customer-facing systems (run by marketing or other business units), then a shift of CDP functions toward the (IT-controlled) data warehouse seems to be the wrong way to bet. (In this context, the recent sale of leading composable CDP Census to data movement platform Fivetran may signal incipient consolidation in the young-but-already-overcrowded composable CDP sector. That Census was bought by a data movement tool rather than a customer-facing system reinforces the notion that composable CDP products serve IT teams while customer-facing CDPs serve marketers and other end-users.)

Presumably all those non-central warehouses are acting as data sources to CDPs or customer-facing systems with a CDP inside. Indeed, the Brinker/Riemersma report positions CDPs (stand-alone or embedded) as intermediaries between data systems and activation systems (which they call "systems of knowledge" and "systems of context," respectively), responsible for "organizing or framing the data in a way that best serves more situational needs." I quite agree, and couldn't have said it better. I would expand on this by noting that CDPs within customer-facing systems will have direct access to the data those systems generate, so the role of the data warehouse is limited to supplying data that the warehouse collects elsewhere. A CDP with direct access to customer-facing systems overcomes one of the main drawbacks of the warehouse-centric approach, which is that warehouses often can't provide the real-time access to behavior data that's needed for many customer-facing applications of CDP data.

I should stress that CDPs are a bit player in the Brinker/ Riemersma epic. Definitely download the report (it’s free and ungated) and focus on whichever portions you find most relevant. It’s all good.

------------------------------------------------------------------------------------------------------------------------------------

*The 2024 survey had 168 responses, of which 60% were B2B, 9% were B2C and 31% were mixed. Doing that math, that’s 67 responses in the B2C and mixed group, of which 18 cited CDP as the center of their stack.

**The 2025 report says 30 of the 96 responded were from ‘enterprise’ organizations, but doesn’t indicate how these split between B2B, B2C and mixed.

***In addition, at least one unacquired CDP (BlueConic) has repositioned as a customer-facing system.

Customer Experience Matrix