The State of AI in Lifecycle Marketing

Ask an AI vendor what's possible right now and you'll hear a familiar story: agents that reason, adapt, and orchestrate the customer journey end to end with very little human intervention.

The research points to a more constrained reality.

AI is already very good at a growing number of marketing tasks. The problem starts when we confuse being good at individual tasks with being reliable across an entire workflow.

The model is often not the bottleneck. The workflow is. More specifically, the bottleneck is how much complexity we expect the model to carry at once, how many decisions are hidden inside what we call a single task, and what happens to reliability when those decisions start depending on one another.

That distinction matters a lot more than whether the newest model scores a few points higher on a benchmark.

What's currently being claimed

Where this is headed isn't in question. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of this year, up from under 5% a year earlier.¹ In BCG's 2026 global survey of 300 CMOs, 96% said AI is driving end-to-end transformation of marketing. Only about a third had actually moved to agent-led workflows.²

That gap is more interesting than the adoption number. Vendor platforms describe agents that reason, adapt, and orchestrate multi-step actions with little human involvement. The claimed transformation is running ahead of the operating model underneath it.

Execution is one thing. Judgment across a long chain of dependent decisions is another.

Complexity and reliability move in opposite directions

Researchers have documented a pattern sometimes called context rot: as the amount of information given to a model increases, its ability to reason consistently across all of that information tends to decline.

One 2025 study tested eighteen models across increasingly long contexts.³ Performance degraded as input length increased, particularly when the task required connecting several pieces of information rather than retrieving a single fact.

Other research has found a related problem.⁴ Information buried in the middle of a long prompt can receive less weight than information positioned near the beginning or end. In some tests, accuracy fell by twenty to thirty percentage points depending largely on where the relevant information appeared.

That creates a practical problem for marketing. The situations where we most want AI to exercise judgment are usually the situations with the most context.

A lifecycle decision may depend on persona, product usage, account status, prior messaging, sales activity, lifecycle stage, geography, entitlement, intent, campaign history, and business rules at the same time. We tend to treat that as one decision because the final action might be simple: send this message, route this lead, recommend this offer.

It isn't one decision.

Reliability compounds across the workflow

The second problem becomes visible when you stop looking at individual AI tasks and start looking at the full process around them.

Take something that sounds simple: turn a blog post into a LinkedIn post using AI and a workflow tool. In practice, a reliable version looks more like this:

1. Extract the core claim from the source post
2. Check it against brand positioning and voice
3. Identify the strongest supporting detail
4. Determine the appropriate LinkedIn hook
5. Draft the post
6. Verify that the draft hasn't changed the source argument
7. Check length
8. Check formatting
9. Confirm that no unsupported claims or numbers were introduced
10. Compare tone against recent published content
11. Verify links and mentions
12. Determine whether the post needs a visual
13. Route it for review
14. Incorporate edits
15. Schedule and confirm publication

Most of those steps are manageable individually. The problem is what happens when you chain them together.

A simple reliability model shows the problem. If every step has to succeed, failures are independent, and there are no retries or validation gates, end-to-end reliability is the product of the reliability of each step. A three-step process running at 70% reliability per step has about a 34% chance of completing perfectly under that model. A ten-step process running at 85% reliability per step falls to about 20%.

That's the 15-step workflow above, at 95% reliability per step under a simplified model. Even 99% reliability per step, a genuinely high bar to hold across fifteen separate steps, only holds up 86% of the time.

Real workflows are messier than this model. Failures aren't always independent, some steps can retry, and validation gates can catch errors before they propagate. But those controls are exactly the point: dependable AI workflows require architecture around the model. Chaining individually capable steps together doesn't automatically produce a dependable system.

This is why "AI can write a LinkedIn post from a blog" tells us very little about whether AI can reliably operate the workflow around publishing one.

I've run into this repeatedly while building AI-assisted processes. The difficult part is rarely getting the model to perform a task once. The difficult part is identifying all of the decisions that were previously being made implicitly by a person and turning them into something explicit enough for a system to reproduce.

A workflow that looked like one step turns out to contain twelve. That is where most of the work is.

Marketing has a decomposition problem

I don't think this is primarily an AI-literacy problem. It is a process decomposition problem.

Engineering and operations disciplines tend to spend more time breaking systems into components, defining dependencies, identifying failure states, and determining ownership before automating them. Marketing often works differently. A process gets described by its output:

• "Send a welcome email."
• "Build a nurture."
• "Route this lead."
• "Create a campaign."
• "Personalize onboarding."

Each one sounds like an action. Each one can hide a fairly substantial decision tree underneath it.

I saw this directly while rebuilding lifecycle systems. "Send a welcome email" quickly becomes a branching set of decisions, not one action.

Those questions existed before AI. Humans were simply absorbing many of them without documenting that they were doing it. AI makes the hidden structure visible because the system cannot reliably reproduce decisions you never explicitly defined.

That's closely related to the Setup Absorption Model. Adding automation does not necessarily fix the problem. First you have to determine where the work actually belongs.

The data underneath the workflow matters as much as the workflow itself

None of this works if the data feeding it isn't trustworthy in the first place.

Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data.⁵ The same research found 63% of organizations either don't have, or aren't sure they have, the data management practices AI actually requires. A separate 2026 survey of B2B marketers found only 44% rate their own data quality and accessibility as adequate for AI at all.⁶

That's not a technology gap. It's duplicate records, lifecycle stages defined three different ways across three different systems, and fields nobody clearly owns.

A welcome or onboarding series looks like the simplest possible automation candidate: one trigger, one audience, a handful of messages. But every branch in the decomposition above still depends on data that's accurate, current, and consistent across whatever systems feed it. If lifecycle stage means something different in the CRM than it does in the marketing automation platform, no amount of workflow decomposition fixes that. The workflow was never the actual constraint.

I saw this directly in a lifecycle rebuild that turned up years of accumulated CRM debt: duplicate records, conflicting properties, inconsistent lifecycle stages, cleaned up before any of the automation built on top of it could be trusted. That work rarely gets budgeted as part of an AI initiative. It's usually the actual initiative.

Where AI is reliable right now

The good news is that AI performs best when its role is bounded: narrow scope, clear inputs, limited judgment, observable outputs.

That doesn't mean every bounded task should use AI. If a decision can be expressed reliably as a rule, conventional automation is usually the better tool. A routing rule doesn't need a model.

AI earns its place where some interpretation is required but the boundaries of that interpretation can still be clearly defined: classification, summarization, first-pass content variants, intent interpretation, pulling structured information out of unstructured input. Lead routing, taxonomy enforcement, and suppression logic are usually better left to deterministic rules.

These are not insignificant use cases either way. In many organizations, both categories represent hundreds or thousands of hours of work, and they share something important: the boundaries of the task are relatively easy to define.

Reliability holds much better when AI is responsible for a bounded, interpretive task than when it's asked to own the system, or asked to do work a simple rule could already handle.

The useful question isn't whether AI is ready

Context degradation and workflow compounding look like two different AI problems. Operationally, they point to the same design issue.

In one case, we ask a model to reason across too much information at once. In the other, we ask too many dependent steps to remain reliable across an entire chain. Both get worse when something we describe as one task is actually several.

The same principle applies to sequencing a customer journey. Individual steps can all be reasonable while the system still fails because the dependencies between them are wrong. AI workflows behave the same way.

Don't start by deciding whether a workflow is "AI-ready."

• Map the decisions inside it.
• Count the dependencies.
• Identify what has to be true for each step to work.
• Automate the parts that can stand on their own.

The model is becoming less and less likely to be the limiting factor. Understanding the system you're asking it to operate is.