The next wave of martech is agentic.

Having sold my small agency to Alloy last year, I’ve been grateful to join a culture that encourages its leadership team to do things like writing blog posts. Unfortunately, one side effect is that I... occasionally get asked to write blog posts. And so, I hastily signed up to write about “the top ten things to consider when evaluating martech tools.” That title, don’t get me wrong, was my own suggestion — but when I mooted it, I hadn’t necessarily thought all the way through about what I’d actually have to write.

I had a general sense of it, of course, based on my years in the field: don’t buy based on buzzwords; make sure you have a good understanding of your requirements and let them shape your decision-making process rather than being swayed by snazzy-looking demos that might not suit your use cases; consider all of the costs; avoid vendor lock-in, and so on. Good advice, I think, but hardly revolutionary — and the truth is that there are plenty of handy checklists already out there.

Luckily for me, there is one big thing happening in the world as I write this that means I can add a little bit of extra color on top of a standard selection checklist. Any guesses? Two words. Been around a long time but showing a real resurgence. That’s right: balloon pants.

I kid. Yes, it’s AI. And so, rather than give you another bulleted evergreen vendor-selection checklist, I’m going to try on my prognosticator’s hat and make three predictions about how AI will change the martech selection process.

Prediction 1: Agent-readiness — not bolted-on AI — will be key.

I know, I know: “agentic” is already a bit of a tired buzzword. Still, it has legs. The premise of having AI tools work not merely as chatbots but as helpers that can use memory, knowledge bases and tools to go and do things — both prompted and autonomously — is too enticing to pass up, and even in its relative infancy is proving genuinely disruptive. And yet a lot of tools in the martech stack today are still very “hands-on-keyboard,” requiring a human in the loop to build — in most cases — nearly everything. We’re getting much, much closer to being able to reliably use AI to ideate things like sophisticated campaign briefs that consider available tools and data and recommend appropriate KPIs. But actually bringing those campaigns to fruition — from building journeys in the marketing automation platform, designing and maintaining a consistent tagging system, configuring a viable A/B test or implementing integrations to get data from across systems into polished, executive-ready reports — often still proves frustratingly manual.

The last year or two have seen a lot of martech providers rush to bolt on user-facing AI functionality. “Describe what you’re looking for,” they’ll ask, and then move happily along building something that, to paraphrase Douglas Adams, is almost but not entirely unlike what you were trying to do. And while this can no doubt improve, I think it’s the wrong model. Having a separate AI (each with its own context and quirks) in every tool can start to feel like a Rube Goldberg machine: overly complex, inefficient and fragile.

The best path forward, instead, borrows from the lessons of decades of open approaches to technology: open, well-documented APIs, common data structures like XML and JSON and patterns such as webhooks have long enabled the different parts of the martech stack to work better and more seamlessly together, even when their inner workings are very different. That interoperability didn't come from individual tools each bolting on some cool, well-liked piece of functionality. It came from making sure different tools could talk to and hand off to other tools in a common, expected way, while individually doing what they do best.

This is the role that MCP — Model Context Protocol — is designed to play; it’s “an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools.” In very simple terms, it gives platforms a standardized way to describe the data and actions they can make available to an AI application — and gives that application a consistent way to discover and use them, subject to the permissions and controls put in place.

What that increasingly enables is a shift in how the individual tools are used: from having a human essentially doing the grunt work of implementing a strategy across the tools in the stack to letting the AI of choice coordinate and execute that work (with a human’s supervision).

At Alloy, our web team is using MCP-enabled technology to move much more quickly from creative brief to design. A lot of the martech stack isn’t there yet, but I predict that the vendors that are thinking in terms of how they enable Claude (or ChatGPT, or DeepSeek or whatever tomorrow’s AI agent-of-choice might be*) to work with their platform will be those best positioned for success.

Takeaway: Don’t just ask what AI features a platform contains. Ask how easily, safely and comprehensively an external AI agent can use it. Does the vendor support MCP? Which capabilities does it expose? Can an agent both retrieve information and take action? How granular are the permissions? Are those actions logged and auditable? And is the vendor genuinely committed to an open, interoperable future?

Prediction 2: Unexciting controls will matter as much as exciting functionality.

As AI applications grow more powerful and approaches like MCP allow for greater orchestration and automation of the martech stack, boring old concepts like governance, security and accountability are going to become ever more important. That means granular permissions, approval gates, detailed audit logs, meaningful testing environments and reliable ways to pause, reverse or recover from an agent’s actions.

All of this exposes a bit of a paradox: how to give AI-powered martech stacks enough autonomy to actually be useful, without (as a seminal work on the subject of AI warned us) accidentally enabling them to launch a nuclear strike on Los Angeles because that’ll help boost home sales in Las Vegas.

I’m being somewhat glib — and, thankfully, most of us in the marketing world aren’t operating systems capable of launching nuclear weapons (no matter how urgent the task of selling more widgets might seem); still, the more we’re willing to relinquish control over the details of the process, the greater the risks and the need for guardrails and audit mechanisms. Today, that should typically mean having a human in the loop at key moments to verify and approve higher-risk actions — large-scale database updates or mass email sends, for example — before they take place. The trouble, of course, lies both in requiring approvals often enough to manage the risk without inducing approval fatigue and in making sure the human in the loop sufficiently understands what they’re actually approving.

There are interesting ideas around things like synthetic test environments — with synthetic leads, personas, users, inboxes and CRM records — that can be used by an agentic martech stack to run an entire campaign end-to-end, at large scale, thousands of times, so marketers can watch and measure what it does… all without a single real customer ever being touched. Still: how do you validate that an agentic system will behave well under real conditions, without exposing it to real conditions? Test it only in the sandbox, and you never really know how it'll behave once actual customers, actual edge cases and actual chaos show up. Test it for real, and you've accepted the very risk you were trying to avoid in the first place.

Some of these questions start to get into genuinely philosophical territory. Thankfully again, I think we’re still quite a long way — no matter what solution vendors might say — from a truly AI-driven-and-optimized martech stack in which campaigns are being designed, built, tested and executed end-to-end with limited human interaction. But we are on that path, and vendors, working with marketers, are going to have to find better ways to increasingly enable intelligent autonomy while managing the risks of a rogue agent wiping out your sales database or (as could have happened with one of Alloy’s clients, had it not been for a manual verification step) publishing an AI-augmented blog post about how self-storage customers might want to “install some security fencing around that storage facility perimeter” — having apparently confused the customer with the facility operator...

Having the means to implement guardrails and manual verification steps, with granular permissions structures and access to detailed audit logs, will help mitigate both the reputational risk and the damage that these systems can do. But then, of course, the temptation will be to get AI to monitor the logs and the permissions structures and the guardrails. Quis custodiet ipsos custodes?

Takeaway: Don't just ask whether a vendor lets you build in guardrails: ask whether they've actually thought about where the guardrails need to sit. Can actions be tiered by risk, so a routine report pull doesn't require the same sign-off as a mass email blast or a database write? Can you assign approval thresholds by action type, not just an all-or-nothing toggle? Are there audit logs detailed enough to actually reconstruct what an agent did and why, after the fact — not just that something happened? Has the vendor thought about approval fatigue at all, or will their approval flow train your team to rubber-stamp everything within a month?

Prediction 3: Pricing models built for humans are about to make less and less sense, but figuring out what does make sense is going to be messy.

Most martech pricing still assumes a human is the unit of work. Per-seat licensing, tiered by number of users, exists because a "seat" was always a reasonable proxy for a chunk of a person's time and attention. But if an agent can do the work of five people's worth of campaign execution, traditional seat-based licensing models break down quickly.

The obvious alternatives are some flavor of outcome-based, activity-based and consumption-based pricing: paying per campaign executed, per lead qualified, per ticket resolved or per workflow completed — rather than per person logged in.

All of which sounds logical… until you discover that one apparently simple workflow consists of 47 separately billable agent actions. For those of us who have quickly become eagle-eyed watchers of Claude token utilization, for example, a more complex, AI-driven pricing future doesn’t feel great. It makes costs harder to predict, doesn’t align with traditional expectations or budgeting models, can discourage experimentation and becomes opaque and frustrating for users. And figuring out the right approach to consumption-based pricing is deceptively complex. Just look at the Salesforce Agentforce example. The company’s introduction in late 2024 of an AI usage-based model, charging $2 per AI conversation, led to a lot of backlash, not least because it charged the same whether the AI solved anything or not. Salesforce has since bounced through several different models, landing, for now — and not without a lot of pain and frustration in the process — on a menu of pricing options that includes a hybrid seat-and-consumption model. That evolution suggests that even the largest vendors are still working out how best to price agentic labor.

As vendors open up their platforms to be used with and by AI tools and not just by humans, there’s going to be increasing complexity around pricing models — and just handling that, from a budgeting and renewal standpoint, is going to be its own headache. It’s not going to just be about which specific pricing model a vendor decides to use, but also about whether that model aligns with your business’s needs — or, indeed, whether they’ve actually thought about the mismatch or are still pricing as if a human is going to be the only one doing the clicking.

Takeaway: Ask a vendor directly how their pricing changes if an agent, not a person, ends up doing most of the day-to-day work in their tool. A flat "per seat, unlimited agent use" answer sounds great today, but treat it as a honeymoon rate, not a permanent one. Vendors notice when their pricing model stops capturing the value they're delivering, so be prepared for repricing at renewal rather than assuming you’ll be grandfathered indefinitely. Lock in favorable terms contractually while you can, rather than assuming today's math holds in two years.

Conclusion

A reasonable question at this point — and one I've been quietly asking myself throughout — is where an agency like Alloy fits into all of this. If the machines are going to handle the grunt work we've historically been paid for — building the journeys, wiring up the integrations, wrangling the tags, assembling the reports — am I merrily typing my way towards my own redundancy?

I don't think so, and not solely because I'd like to keep my job. Every one of these predictions describes work that doesn't do itself: someone still has to work out what a client actually needs before pointing an agent at it, see through the vendors whose "MCP support" turns out to be a press release, design the guardrails so the thing stays useful without deciding that the fastest route to more home sales in Las Vegas runs through Los Angeles and make sense of pricing nobody's quite figured out yet. The more capable the agents get, the more the value shifts to the judgment around them — and spotting the patterns across a lot of clients and a lot of stacks is exactly what an agency does before any single in-house team racks up the same reps. Less hands on the keyboard; more architect of the stack and designer of the guardrails.

And none of what I’ve predicted means the traditional selection criteria disappear. You still need to understand your requirements, account for the full cost of a solution and resist being dazzled by a slick demo — AI just adds a new layer to each. Increasingly, the right platform won't be the one with the most impressive AI features; it'll be the one your agents of choice can use effectively, your organization can govern safely and your finance team can pay for predictably. And one thing's for sure: the balloon pants thing can't possibly last.