Walk into any marketing organization right now and you'll hear some version of the same ambition: "We're going to use AI agents to do more, and do it faster!" That's a good instinct, because agents really are good at the repetitive, data-heavy work that eats marketing teams alive.
But what usually doesn't come along with the ambition is an operating model to go with it. So a few months in, a lot of teams find themselves with a folder of impressive demos, a pilot channel that's gone quiet, and no clear line back to business value.
I've come to see that as a process gap more than a technology one, which is actually good news, because process is something we already know how to fix. If you treat an AI agent as a tool you buy or a feature you launch, you've skipped the part that matters most. It's much closer to a hypothesis, and hypotheses need somewhere to be tested.
For me, that somewhere is an incubation team: a small, cross-functional pod whose job is to figure out quickly whether a given agent is worth building at all.
Here's the model I use to run that lab, which I call the PROVE framework. It pulls from three disciplines I keep coming back to:
- The ROI-first innovation cadence of "think big, start small, scale quickly"
- The human-centered rigor of design thinking
- The data-layer discipline of journey mapping
Establishing ROI before you build an AI agent
Before any of the framework matters, there's one thing I won't budge on: you establish business value and ROI first, and then you build.
I know how obvious that sounds, but I also know how tempting it is to skip this step. A shiny new model has real gravitational pull, and honestly, prompting is a lot more fun than sitting down to define what "good" would even look like. But if you can't tie a proof of concept back to a metric, you don't really have one.
The three gates of an AI agent pilot
The model moves through three gates, and each one asks something different of you. Together, they're what keeps a pilot from tipping too far toward ambition on one side or too much caution on the other.
- Gate one, think big: Diverge widely across the problems worth solving, then converge hard on the single highest-value opportunity.
- Gate two, start small and learn fast: Build the smallest possible version that can test the bet, because speed of learning beats polish.
- Gate three, scale quickly: Once value is proven, move fast in the other direction by integrating, governing, instrumenting, and operationalizing before the momentum fades.
If you scale before you've proven value, you'll burn through the budget. But if you stay in learn-fast mode indefinitely, you'll never ship anything at all. The key to success is knowing which gate you’re in and behaving accordingly.
eBook | Get your marketing data AI-ready
Transform your marketing strategy with our eBook on building an AI-ready data foundation, essential for leveraging AI in today’s landscape.
The five stages of the PROVE loop
Inside those gates is a five-stage loop you run over and over. Each stage produces one clear output, and you don't move on to the next one until you have it in hand. I use "PROVE" as the shorthand for:
- Pinpoint
- Ready
- Orchestrate
- Validate
- Expand
Pinpoint the use case worth testing
Start with the customer journey, not the technology. Where in your customer's experience (or in your team's own workflow) is there a moment that's repetitive, judgment-light, data-rich, and expensive in time or money? When you find something with all four of those traits, you've very likely found your first agent.
It helps to know what the capability stack underneath an agent needs to look like before you commit to a candidate, because that's what tells you whether "in weeks rather than quarters" is realistic.
Diverge across candidates, then converge on one by running it through four lenses:
- Desirable: Does a real user actually want this, and will they trust it?
- Viable: Is there a business case, and can we measure it?
- Feasible: Can we technically build a first version in weeks rather than quarters?
- Responsible: Can we do it safely, within privacy, consent, and brand-safety limits?
The output of Pinpoint is a single sentence, your value hypothesis: for this user, at this moment, we believe an agent that does X will produce outcome Y, and we'll know we're right when metric Z moves by a set amount.
Ready your data and guardrails
This is the stage where pilots need the most attention, because an agent is only as capable as the data and context it can reach.
Ready is a deliberate inventory of two things, borrowed from journey-mapping discipline:
- Data to be collected: What will the agent need to capture as it works?
- Data to be used: What can it draw on from the stack you already have, including your analytics platform, your customer data platform, your media platforms, your first-party data, and your knowledge base?
For each of those, you're confirming that access is genuinely wired up, that the data is clean and current, that using it squares with your consent and privacy commitments, and that you've been explicit about what this agent isn't allowed to do or say. If the foundation underneath is shaky, that work comes before the agent rather than after it. It's the same sequencing problem I've written about with identity architecture, where skipping rungs produces use cases that quietly underperform.
I call this whole inventory the data and trust spine, because it's what holds up everything else. It decides what your agent is able to do, and it decides whether you can trust what it hands back.
Orchestrate the minimum viable agent
Now you can build, as long as you keep it to the smallest version that can actually test your value hypothesis. You're not building the full vision here, just a slice of it you can put in front of someone.
Orchestrate is where the model, the data connections, the prompts or workflow logic, the tool calls, and the human-in-the-loop checkpoints get wired together. Anything you add beyond what the test requires could potentially slow your learning down. The output is a working agent, in a sandbox, ready to be tried on real tasks.
Validate against value and trust
Put the agent in front of real users doing real work, and measure it against two bars:
- Value: Did the metric from your value hypothesis move?
- Trust: How does it score on the things that sink AI in production, including accuracy, hallucination rate, latency, cost per task, and brand safety?
Keep in mind, though, that neither bar tells you much if your measurement can only show you what happened at the surface. Getting to analysis that explains why a number moved is what separates a pilot result you can act on from one you have to take on faith.
Expand into production
Expand is the deliberate work of moving from sandbox to production, which means integrating the agent into everyday workflows, hardening governance and monitoring, training the people who'll use it, and setting up ongoing ROI measurement so the value you proved doesn't disappear six months later.
Agents making real decisions inside live workflows are already doing this, and what that looks like in B2B marketing and decisioning is a useful picture of what "in production" means.
Expand also closes the loop. A mature incubation practice always has a next hypothesis queued, so the pod moves from one proven win to the next bet without losing tempo.
Who belongs on an AI incubation team
In my experience, the right unit for this work is the pod itself, small enough to move and complete enough to ship. One person can't carry it, and a whole department will move too slowly to learn much of anything.
Here's an example lineup to consider against your own org chart:
- Pod lead or value owner: Usually a marketing or analytics leader, this person owns the value hypothesis and the ROI story end to end.
- Builder: An AI or data engineer who wires up the model, the tools, and the data connections.
- Data and measurement strategist: Owns the data and trust spine, including what to measure, whether the data is ready, and whether the result is real.
- Domain expert: The marketer who will actually use the agent, keeping every decision anchored to a real workflow instead of assumptions.
- Trust partner: A part-time or shared role covering privacy, consent, and brand and responsible-AI review.
Above the pod, an executive sponsor protects its time and clears organizational roadblocks while letting the team manage the day to day.
The goal of an incubation team is twofold: driving adoption and innovation, and finding, analyzing, and building the use cases worth scaling. The organizations that win with AI won't be the ones with the most demos. They'll be the ones that built the muscle to prove, through repeatable and collaborative processes, which transformation projects are actually worth it.
I'll be at TAG's //SHIFT AI summit in Atlanta on September 18, speaking about this exact topic, so come join me if you’re in the area! And if you’ve got questions about what your first marketing agent pilot should be, or what your data needs to look like before one can run, we'd love to talk.
