A surprising number of marketing leaders still haven’t deployed a single AI agent. Not because they lack access to the technology. Most enterprise marketing teams already have access to ChatGPT Enterprise, Claude Enterprise, Copilot Studio, Gemini, or some combination of all four. The tools are available. The investment has already been made.
Yet many organisations remain stuck in evaluation mode. They attend webinars. They discuss governance. They commission working groups. They debate use cases.
Meanwhile, other teams are deploying. This distinction matters because the primary advantage created by agents isn’t automation. It’s learning. Every deployed agent teaches you something about how people work, where friction exists, which processes deserve automation, where models fail, and what governance looks like in practice. None of those lessons emerge from planning sessions. They emerge from deployment.
I’ve built and deployed agents across marketing and regulated environments. Some delivered immediate value. Others failed to gain adoption. A few solved problems nobody cared about in the first place.
The outcome mattered less than the learning.
The teams deploying agents today are building organisational knowledge that compounds over time. The teams waiting for certainty are not.
Why Marketing Leaders Struggle to Ship Agents
Most marketing leaders assume the barrier is technology. In my experience, the barrier is usually scope. The moment AI enters the conversation, sensible business problems suddenly become transformation programmes.
Instead of building a tool that helps identify financial promotions before compliance review, teams start discussing autonomous campaign generation.
Instead of automating competitor analysis, they try to redesign the entire marketing operating model.
The result is predictable. Complexity increases, stakeholder groups expand, governance requirements multiply, and progress slows.
I’ve observed four recurring patterns.
- First, teams start too big. They try to automate entire workflows rather than individual tasks. The ambition sounds impressive, but the complexity quickly becomes unmanageable.
- Second, they wait for perfect governance. Governance matters, particularly in regulated environments, but waiting six months for every stakeholder to align often kills momentum before a project gets off the ground.
- Third, they overestimate technical complexity. Many first-generation agents require configuration rather than software development. The barrier to entry is significantly lower than most people assume.
- Finally, they optimise for perfection. They spend months refining something nobody has used, instead of putting a useful tool in front of users and learning from real behaviour.
The most successful deployments I’ve seen all share one characteristic. They started small.
If you’re a marketing leader looking to deploy your first agent, here’s the framework I’d recommend.
Step 1: Start Small and Narrow the Scope
The biggest threat to a successful first agent is trying to get it to do too many things.
You need to start thinking in problem-solving chunks rather than end-to-end processes.
Good examples include:
✓ A DPIA wizard that guides employees through the completion of a Data Protection Impact Assessment
✓ An event collateral stock checker and forecaster
✓ A financial promotion identifier and pre-compliance checker
✓ A competitor analysis assistant
Poor examples include:
✕ A marketing campaign generator
✕ A revenue attribution platform
✕ An end-to-end GTM operating system
✕ An autonomous marketing department
One lesson I’ve learned is that your best agent ideas usually come from frustrations you experience yourself.
If you find yourself performing the same task every week, following the same process, consulting the same documentation, or answering the same questions, there’s a strong chance you’ve identified a viable use case.
Your understanding of the problem becomes a competitive advantage. You know where the friction exists, what good looks like, and whether the output creates value.
Step 2: Identify the Required Knowledge
At their core, agents are language models that have been given instructions, guardrails, and access to relevant knowledge.
Before building anything, ask yourself a simple question:
“What information would I need to give a new starter so they could perform this task to an acceptable standard?”
The answer usually becomes your knowledge base.
For a Financial Promotions Agent, that might include FCA guidance, Consumer Duty documentation, internal approval processes, brand guidelines, historical compliance feedback, and examples of previously approved content.
For a Competitor Analysis Agent, you might provide competitor lists, positioning documents, market research, product information, SWOT frameworks, and previous competitor reports.
Most agent failures aren’t model failures. They’re knowledge failures.
Teams spend hours debating which model to use while paying very little attention to the information feeding it. The reality is that a well-instructed agent with strong knowledge will outperform a poorly configured agent running on a more advanced model.
Knowledge quality matters more than model selection in most business use cases.
Step 3: Build Using Tools You Already Have
Most marketing leaders don’t need developers to deploy their first agent. They just need to start.
Today you can build useful agents using ChatGPT GPTs, Claude Projects, Copilot Studio, and Gemini Gems.
My recommendation is simple. Use whichever ecosystem your organisation already pays for. Too many teams waste time evaluating platforms when they haven’t yet proven a single use case. The objective isn’t to build the most sophisticated agent in the business. The objective is to create something useful. Version one should make somebody’s job easier. Nothing more.
If users describe your agent as helpful, you’ve succeeded.
Step 4: Ship Before You’re Comfortable
This is where most projects fail.
The agent reaches a point where it’s functional. It solves the problem reasonably well. Internal testing looks positive.
Then somebody asks for more:
- More testing.
- More features.
- More integrations.
- More governance.
- More stakeholder reviews.
Months pass. Nothing launches.
The irony is that most of the questions raised during these discussions would have been answered within a week of real-world usage.
Your first agent won’t be perfect.
Neither was your first website, your first campaign, your first automation, or your first reporting dashboard. Deploy it. Observe how people use it. Identify where it breaks. Learn from the feedback. Improve it.
The fastest path to a useful agent is through deployment.
Step 5: Measure Adoption, Not Accuracy
One of the first questions leaders ask is:
“How accurate is it?”
I think that’s often the wrong question.
A better question is:
“Is anybody using it?”
An agent with 95% accuracy and zero users creates no value. An agent with 80% accuracy that saves fifty employees twenty minutes each week creates significant value.
Focus on adoption metrics.
Track:
- Active users
- Sessions
- Repeat usage
- Tasks completed
- Time saved
- User feedback
Pay particular attention to repeat usage. People return to tools that help them. They abandon tools that don’t. If adoption is growing, you have evidence that you’re solving a real problem.
Everything else becomes easier from that point onward.
Step 6: Market Your Agent Like a Product Launch
One of the biggest mistakes I see is spending weeks building an agent and minutes launching it. An email gets sent. A Teams message appears. Then everyone moves on. A month later somebody concludes that employees weren’t interested.
Internal products require internal marketing.
In many cases they require more marketing than external products because your audience didn’t ask for them. Whenever I deploy an agent, I package and position it exactly as I would a product launch.
The framework below has worked:
Common Blockers You’ll Encounter
Compliance is often the first concern raised in financial services. Ironically, some of the safest AI use cases are internal productivity agents. Bring compliance teams into the process early. Show them the scope, controls, knowledge sources, and governance model. Most resistance comes from uncertainty rather than opposition.
Perfectionism is another common blocker.
Your first agent is a pilot, not a flagship product. The objective is learning, not perfection.
Data also creates problems.
Many teams start with data-heavy use cases because they appear strategically important. In practice, knowledge-heavy use cases are usually easier to deploy, easier to govern, and easier to improve.
Politics presents a different challenge.
Some employees will view agents as an opportunity. Others will see a threat. Position agents as capability multipliers rather than headcount replacements. Adoption becomes significantly easier when people understand that the objective is to remove repetitive work rather than remove people.
Final Thoughts
Three years ago marketing leaders debated whether AI would change the industry. That debate is over.
The leaders building agents today are developing knowledge that will shape how their organisations operate in the future. They are learning which workflows deserve automation, which governance models work, and where genuine value exists.
✕ You don’t need an AI transformation programme.
✕ You don’t need a steering committee.
✕ You don’t need a six-figure budget.
✓ You need one problem worth solving.
✓ Build something useful.
✓ Deploy it.
✓ Learn from it.
✓ Then build the next one.
That’s how capability develops. That’s how organisational knowledge compounds. And that’s why marketing leaders who haven’t shipped their first agent yet are already falling behind.
