Editorial
Research Report
Building the Case for Modern Public Sector Messaging
The Gist
- What is "inference debt"? The rising cost of running a personalization stack whose returns have flattened, hidden across four signals: flat lift despite rising spend, costlier A/B tests, over-segmented audiences and mounting integration costs.
- Why has behavioral inference stopped improving personalization? Inference only learns from behavior customers have already produced, and that ceiling is compounded by signal fragmentation, privacy regulation and cookie deprecation.
- What should replace inference as the primary signal? Expressed intent — what customers directly say they want, captured in the moment and persisted over time — becomes the new foundation, with inference remaining a secondary input.
The Curve Everyone Expected to Keep Climbing
The personalization stack enterprises have been steadily building since 2015 rested on two assumptions: that behavioral inference would keep getting better, and that more behavioral data running through better models would produce better personalization. These were reasonable assumptions at the time, and many of us made them.
The bet has stopped paying, but the conversation inside most enterprises has yet to catch up to that fact. The vendor still walks in with a roadmap that promises the next release will change the trajectory, the team still frames the quarterly number in a way that makes the shortfall sound like a testing or tagging problem, and the budget still gets approved on the same cycle as always. The people who championed the program are still in the room although the line item has longer tenure than half of them.
The question has moved from whether the stack is performing to what the next round of spending is actually buying and whether that answer is good enough to keep buying it.
What a Decade of Inference Actually Bought
The reframe ahead is hard for a specific reason: the investment worked.
Behavioral inference produced real value for the organizations that committed to it early. Customer Data Platforms (CDPs) finally unified data the team had been trying to pull together for years. Recommendation engines lifted conversion in the categories where the signal was strong enough to read. Recognizing the same customer across channels in close to real time, a multi-year ambition became the norm.
Segmentation got sharper along the same trajectory. Audiences that used to be defined in millions could be defined in thousands, then in hundreds, and the marketing team running them stopped thinking of itself as a brand function and started thinking of itself as a revenue one.
A retailer could treat a lapsed high-value, long-term customer differently from a lapsed new one, a bank could treat a mortgage prospect differently from a checking account holder browsing the same page, and the work that previously required a campaign manager and a quarter could be done by a rule and an afternoon.
The early adopters built a real moat. Acquisition costs came down because the matching got better, retention climbed because the next-best-action engine got smarter, and the companies that ran the play in 2017 and 2018 are still operating with cost-per-customer numbers their later-arriving competitors still seek.
Related Article: Customer Acquisition Makes You Famous. Customer Retention Makes Money.
Why the Inference Curve Has Plateaued
The structural reason is the simplest one. Behavioral inference works by watching what a customer does and predicting what they want from the pattern of those actions.
This means the system only learns from behavior the customer has already produced, and that backward-looking design forms the ceiling the rest of the stack sits beneath.
That was a tolerable constraint while the signal was still new and every data point was visibly improving the model. It binds now. The high-value signal has mostly been captured, the models have settled, and the next data point often costs more to collect than it adds to the prediction.
The rest of what has slowed the curve sits on top of that ceiling and impedes the ability to keep climbing. The signal-to-noise ratio has gotten worse as touchpoints, devices and identity fragmentation multiplied, which means a growing share of the behavior the system captures is ambiguous, unattributable or both. Privacy regulation and third-party cookie deprecation have shrunk the input dataset the original models were trained to depend on; a constraint the architecture was never designed to survive.
The team running things is not blind to any of this. They are watching the same dashboards and pulling the same levers that used to move the number. The explanation that fits the data is not the explanation anyone wants to put on the slide.
The customer has changed faster than the curve itself. Gartner's 2025 research on personalization put numbers to what most marketing leaders are already feeling.
- 53% of customers report negative experiences from personalized marketing
- Customers who experience it are 3.2x more likely to regret the purchase and 44% less likely to come back for the next one.
The system is still firing on the schedule as designed. But now, the customer is asking it to stop.
What Inference Debt Actually Looks Like on a P&L
Inference debt is the cost of running a system where payback has slowed or stopped. Like any debt, it accumulates without notice.
The annual budget review is an anachronism. It was built for a business that changed slowly enough for once-a-year planning to keep pace with reality. That business no longer exists. Inference debt accumulates monthly. The model degrades monthly. The customer's expectations evolve monthly. The accelerated pace of change rewards organizations that can redirect spending in shorter cycles. Once-a-year reckoning defers that decision until the gap between what that stack costs and what it returns has been growing for three quarters with nobody watching.
The debt stays invisible because it never shows up as a single failing line. It surfaces in four places.
Rising Cost Against Flat Lift
The personalization budget keeps growing because vendor contracts renew, data volumes climb, and the team size running it does not get smaller. The lift does not grow.
The Testing Problem
Optimization teams are running more tests to find smaller gains, and the marginal cost of finding a winning variant has been climbing across categories for the better part of three years. The test backlog is longer than it has ever been. The wins it produces are smaller than they have ever been. The team is working harder than it has ever worked to keep the number from declining.
The Segmentation Paradox
Audiences have been cut so finely that individual segments no longer have enough volume to produce statistically meaningful results, which means the program looks more sophisticated every quarter and produces flatter outcomes at the same time. The sophistication and the flatness are the same phenomenon viewed from different angles.
The Integration Tax
The personalization stack does not run in isolation. It must stay synchronized with the CRM, the CDP, the content system, the analytics layer, and whatever the team adopted last quarter to fix what the prior tool could not do. Gartner's marketing technology survey has tracked martech utilization falling from 58% in 2020 to 42% in 2022 to 33% in 2023, which is the integration tax in numerical form. Enterprises are paying for capabilities they cannot reach, and the share they can reach gets smaller every year.
None of these signals looks like a failure on its own. They look like the normal cost of operating a mature personalization program, which is exactly why the debt has been allowed to accumulate. The annual budget cycle approves them one by one as line items. The cumulative picture only shows up when somebody is willing to look at all four at once and ask what they add up to.
Key Takeaways From the Inference Debt Analysis
The following table highlights the most important lessons, actions and strategic considerations emerging from this analysis of why personalization ROI has flattened and what enterprises should do next.
| Key Area | What Happened | Why It Matters | Recommended Action |
|---|---|---|---|
| Inference ceiling | Behavioral inference has plateaued because it only learns from past behavior | The next data point costs more to collect than it adds to the prediction | Stop over-investing in additional inference layers or vendor upgrades promising to break the ceiling |
| Inference debt signals | Rising cost, costlier testing, over-fragmented segments and integration tax are accumulating unnoticed | Annual budget cycles approve these costs individually, masking the cumulative picture | Review all four signals together on a shorter cadence than annual |
| Customer sentiment | Gartner found 53% of customers report negative experiences from personalized marketing, with 3.2x higher purchase regret | Customer expectations have shifted faster than the inference curve, undermining ROI further | Audit personalization triggers against customer regret research before scaling further |
| Expressed intent | Gartner's "active personalization" customers were 2.3x more likely to confidently complete purchase decisions | Expressed intent, not inferred behavior, is emerging as the higher-ROI signal | Redirect marginal investment toward systems that capture and persist explicit customer statements |
| Organizational shift | The existing inference stack remains in place as one input among several | The transition is additive, not rip-and-replace, lowering execution risk | Task the current personalization team with layering expressed-intent capture on top of the existing stack |
Why Some Dealers Are Pulling Ahead With AI
Register
Content Leaders Collective: Proving Content's Business Impact Starts With the Right CCMS
Register
Replacing Tasks, Not Roles: The Changing Nature of Contact Center Work
Register
How Fast-Moving Content Teams Keep Up as Sites Grow
Register
What Comes Next Can’t Be More Inference
The instinct when a personalization stack stops producing returns is to add another inference layer to it. A smarter model, a richer data source, a new vendor promising to break through the ceiling. None of that changes the shape of the problem, because the constraint is structural and another model running the same way will not change what sits underneath it. The next phase of personalization is built on a capability the inference stack does not have and was never designed to develop. Expressed intent.
The shift has two parts: acting on what the customer expresses in the moment and holding onto those expressions so the system can build on them over time.
Inferred intent reads behavior and works backward to a guess. Expressed intent works by acting on what the customer says they want, in the moment they say it. The two approaches share a vocabulary, but almost nothing else.
This is the same shift that played out in the content layer when the page model gave way to a delivery layer that responds to what the customer actually asks. The intelligence layer is going through a similar transition, and the inferred-intent stack is the page-era equivalent of the content library. It produced real value for a long time, but it’s not the foundation the next decade will be built on.
Some inference will continue to earn its keep, particularly in ecommerce categories where the signal is strong and the use case is narrow. Our argument is about where the marginal investment dollar goes from here, not about ripping out what is still working.
Gartner's 2025 research is starting to name the immediate version of the shift. Customers experiencing what they call active personalization, their term for personalization driven by what customers express in the moment rather than what is inferred from their behavior, were 2.3x more likely to confidently complete critical purchase decisions. The architectural version of the shift is broader than the moment-by-moment version Gartner is measuring, but the direction is the same one the rest of the industry hesitates to recognize.
Capturing expression in the moment is necessary, but not sufficient. The architectural distinction is persistence, and the foundation has to hold what the customer has expressed across the full arc of the relationship so that each new exchange builds on the last and the expressed signal compounds into something the brand can act on. The enterprise stack lacks that foundation today.
Memory, Not Guesswork
Persistence enables the best human interactions. The barista who sees you walking up and starts your drink, the bartender who remembers your last three orders, the salesperson who knows the customer's business well enough that the next conversation picks up where the previous one left off. None of those interactions are produced by inference. They are produced by memory, and the memory is held by someone who was paying attention.
The organizational shift is more modest than the architectural one. The inference stack does not get ripped out. It becomes one input among several, paired with the systems that capture and persist what customers express directly so that it is no longer carrying the whole foundation on its own. The team that knows the existing stack is the same team that operates the new one, the vendor relationships continue, and the data flows continue. What changes is which signal the architecture treats as primary.
Frequently Asked Questions About Inference Debt and Expressed Intent
The following answers address the most common practitioner questions about why personalization ROI is flattening and what to do about it.
What This Means for CX and Marketing Leaders
The question for a CMO sitting on a decade of personalization investment is not whether the investment was worth it. It was. The question is whether the next round extends what has already flattened or builds the foundation underneath it, the systems that receive and persist what customers express. Recognizing the flattening early and redirecting spending there builds a compounding asset instead of another depreciating one.
I have signed checks against this myself. The honest reckoning belongs to the people who built it. The next decade will be won by the brands willing to mark down what has flattened and build on what their customers tell them directly.
Learn how you can join our contributor community.