The next martech consolidation wave is being framed as a software problem. Cut the licenses, merge the data, pick a platform, and the marketing machine should run cleaner.
That story is convenient. It is also mostly wrong.
The harder problem is that many marketing teams don't have a tool shortage. They have workflows nobody can explain end to end. An AI agent dropped into that environment won't create efficiency. It will automate the confusion, then make the confusion harder to audit.
Consolidation is arriving anyway
The pressure is real. CMSWire recently reported that 68% of CIOs planned vendor consolidation in 2026, as companies tried to reduce complexity around software, data, and AI. That number matters less as a forecast than as a mood. Finance wants fewer renewals. Operations wants fewer integrations. Marketing wants the same output with fewer people touching the system.
Klaviyo's August 5 announcement about acquiring the team and technology behind Agency makes the direction plain. Customer platforms want to own more of the intelligence layer, not just the messaging layer. The acquisition is expected to close in the third quarter, according to Klaviyo's investor announcement.
That doesn't mean every company should consolidate around Klaviyo, Salesforce, HubSpot, or any other major platform. It means the platform vendors see the same opening: marketing teams are tired of stitching together systems that each know one small piece of the customer.
The buying logic is understandable. The implementation logic is where things break.
The workflow is the product
A martech stack is usually described as a list of tools. CRM, CDP, email, analytics, media, experimentation. The list makes procurement easy to discuss and almost impossible to operate.
Customers don't experience a list of tools. They experience a sequence of decisions. A lead enters. Someone qualifies it. A segment changes. A message is approved. A suppression rule fires. A sales rep gets notified. Revenue is attributed, or it isn't.
The sequence is the real product.
If the team can't draw that sequence on one page, adding an AI agent is premature. The agent needs to know what event starts the work, which data is authoritative, what actions are allowed, where a human must approve, and what happens when the inputs disagree.
A single customer can be marked high intent in one system, opted out in another, and already contacted in a third. A human might catch the conflict because the campaign feels odd. An agent might select the most recent record and move on.
That is not automation. It's a faster route to an incident.
AI exposes the seams
Traditional automation can survive a surprising amount of mess. A scheduled journey runs on a fixed trigger. A dashboard shows a number with a known definition. A campaign manager compensates for missing context through memory and Slack messages.
Agents are different because they make choices across systems. They summarize, prioritize, route, personalize, and sometimes execute. Each choice depends on permissions, definitions, and exception handling.
That makes the old seams visible.
An agent can't safely optimize a funnel if the funnel has three competing definitions. It can't decide whether a contact is sales-ready if marketing and sales disagree about the threshold. It can't report lift if the control group is overwritten by a last-minute audience sync.
This is why martech consolidation often disappoints after the demo. The new platform may reduce the number of screens while preserving the same ambiguity underneath. A cleaner interface can hide a dirtier operating model.
The fix starts with a smaller question: which decisions should the system make without asking?
That question connects directly to the broader problem I wrote about in AI agents as operating infrastructure. The value isn't in giving a model access to more software. The value is in creating a controlled path from signal to action.
What to consolidate first
Don't start by ranking vendors. Start by ranking decisions.
A useful consolidation plan usually has three layers.
First, consolidate definitions. Decide what counts as a qualified lead, an active customer, a conversion, a suppressed contact, and an influenced opportunity. Put the definitions somewhere durable. If they only exist in a dashboard owner's head, they don't exist.
Second, consolidate permissions. An agent that can draft an email should not automatically be able to change a lifecycle stage, alter a suppression list, or spend media dollars. Separate read access, recommendation access, and execution access. Log every action that changes customer state.
Third, consolidate exception paths. The happy path is easy to demo. The valuable design work lives in the weird cases: conflicting consent, duplicate records, missing revenue data, a customer in two journeys, or a request that falls outside policy. Those exceptions should route to a person instead of disappearing inside an agent trace.
This is also where measurement needs to change. In the reporting trap created by AI marketing, I argued that a polished dashboard can make weak instrumentation look like control. Consolidation doesn't solve that problem. It can make the dashboard more authoritative while the underlying evidence stays fragile.
The platform decision comes later
Once the workflow is clear, vendor evaluation becomes less theatrical.
Ask each platform to show how it handles conflicting customer records, missing consent, failed actions, human approval, and a complete audit trail. Ask what the agent is allowed to do when the data is stale. Ask whether the system can export its decisions in a form another system can inspect.
Those questions are less exciting than a promise of autonomous campaign creation. They are also much closer to the cost of being wrong.
A platform deserves to win when it makes the operating model clearer, not when it makes the stack diagram smaller. Fewer logos on a slide can still mean more dependency, less portability, and a larger blast radius when the central system gets a decision wrong.
The best consolidation may not be one giant platform. It may be a small number of well-defined services connected by a shared decision model, with agents operating inside explicit boundaries.
That sounds less like a software purchase. It is. It is also the work most companies hoped the software would do for them.
The bill comes after the demo
Martech consolidation will continue because the old stack is expensive to maintain and the new AI layer needs cleaner access to customer context. But the companies that treat consolidation as license reduction will discover a familiar surprise: the tools were never the only source of waste.
The waste was in the handoffs. The duplicate definitions. The approvals that lived in private messages. The reports nobody trusted but everybody presented.
AI agents don't remove those problems. They put them on a timer.
Before signing the next platform contract, ask for a map of the decisions, not a map of the products. The answer will tell you whether you're buying leverage or just moving the mess to a more expensive room.