Summary
At the AI for Agencies Summit 2026, the Marketing AI Institute’s virtual event, presented by Screendragon, John Briggs, in a discussion with Cathy McPhillips, shared a simple reality check for agency leaders and their adoption of AI technologies.
This blog unpacks John’s session into a practical approach that covers:
- Why many AI pilots stall once they meet real delivery work
- What “workflow-first” actually means in an agency context
- How to start with one measurable use case (and prove ROI fast)
- How to build governance into delivery without slowing reviews or creativity
- How to manage cost as you move from proof of concept to production
Agencies do not get a sandbox: the AI reality check
When your team is juggling three live campaigns, and someone drops an “AI shortcut” into the middle of the workflow. The output might look fine at first glance, but someone notices the tone is wrong, and the facts are shaky. At that point, the review cycle starts again. Leading to more rework and more cost.
That is the point John Briggs makes early in his conversation: brands can sandbox, but agencies are running “a production line every day”, shipping work to deadlines. When an AI experiment breaks, it breaks in front of a client. That hits trust and margin.
Initially, your approach with AI was, “can we generate content?”. The real problem you will discover is whether you can operationalise AI across dozens of client workflows without chaos and the risk of messing up in front of the client.
Why agency AI pilots stall
Most failed AI initiatives look impressive for a week or two. Then reality shows up.
People lose trust. It generates more work, more review cycles, more approvals, and people quietly stop using it.
John describes a common pattern. Many pilots are built in isolation, disconnected from real work processes, and missing the foundations that make output consistent. The result is predictable. People lose trust, approvals multiply, and the tool creates more work than it removes. Quietly, the adoption of the tool dies. Even worse, new tools pop up “in the shadows,” and leadership loses control of what their people are using to get work done.
This is not a creativity problem. It is an operating model problem.
A quick self-test: are you creating friction, not flow?
If you are seeing any of the signals below, your AI “pilot” may be adding drag.
- Output needs heavy rewriting because it is not grounded in your data and rules
- Review cycles increase because nobody defined what “good” looks like
- People stop using the tool after early novelty fades
- Compliance and governance show up late as blockers
- Leadership cannot measure ROI, so the project becomes a faith-based investment
- Work happens outside the workflow because the AI tool lives in a separate “playground”
If that list feels familiar, the fix is not more experimentation. It is a workflow-first design.
Workflow-first AI: start small, prove it, then scale
John’s advice to agency leaders is blunt. If you try to find one common thread across every client and workflow, you will probably fail. Instead, pick a single, impactful use case that works inside the workflow, prove it, and build from there.
That “inside the workflow” point matters. He keeps coming back to the same foundations:
- A clear use case with realistic expectations
- ROI you can explain in plain numbers
- Integrations, so the agent can act, not just chat
- Human oversight and governance
- Data readiness, so the output is grounded
This is how AI stops being a side project and becomes part of the operating system.
What “small and measurable” looks like in practice
The best examples John shares are the bits on delivery that eat time and confidence.
He talks about teams going after friction and rework: brief quality checks before work starts, support for scope and estimates, risk flags when delivery slips, and brand compliance checks embedded in the flow.
John shares a particular example he recently worked through. An agency working with a large retailer had to validate product details on flyers and signage against the client’s ERP (Enterprise Resource Planning) system. It was a repetitive, high-risk time sink. They built an agent workflow that pulled the SKU (Stock Keeping Unit) data, checked the documents, annotated issues (even drawing boxes around them), and routed items back to creatives only when something was wrong. If it was fine, it moved forward. Tight use case. Data in place. Measurable outcome.
He also calls out a second sweet spot, involving summarising client feedback and converting it into clean tasks. This is how AI is reducing the messy handoff that slows delivery.
“Smarter AI, not more AI”: controlling cost as you go to production
There is another trap agencies fall into. They treat cost like an afterthought until the pilot becomes real.
John puts it simply: a proof of concept can be cheap, but production gets expensive quickly.
“the best teams don’t use more AI, they use smarter AI”.
In practice, that means matching the model to the job:
- Smaller models for repeatable, high-volume work
- Larger models only when you genuinely need a creative lift
- Clear choices that produce predictable costs and consistency
This is also where governance earns its keep. Done well, it should sit in the workflow, escalate only when needed, and act as the first pass that saves time.
When you do proper AI… governance can be invisible.
If you are making decisions about agentic workflows, it helps to evaluate each use case through three lenses: capacity, risk, and cost. That framing is familiar to agency leaders. It also makes AI discussions less abstract and more operational.
Who should lead the development of AI workflows
Near the end of the session, John answers a question many agency leaders are still debating. Who is best placed to spearhead this work?
In his view, operations and content operations. They control the flow of work and the resourcing. They can identify a step an agent can handle, run the task, return to the process, and keep work moving.
He also gives a clear warning. As more agents start “doing the work and coming back into the workflow”, lack of orchestration becomes chaos.
That is the real mindset shift. AI is here to stay, and you get the best results when an agent platform works alongside your people, not replacing your people. It’s vital to keep humans in the loop. Measure performance. Iterate.
The practical next step for agency leaders
If you want to move from experimentation to advantage, do this in the next week:
- Pick one workflow you already run at volume. Map the handoffs.
- Choose one measurable use case that removes friction or rework.
- Define ROI in time and margin and decide how you will measure it.
- Put governance and human oversight inside the workflow, not at the end.
- Choose “smarter AI” by matching model size to the task and cost profile.
- This is the work Screendragon is built to support: orchestrating people, workflows, and AI agents at scale, so delivery stays controlled, visible, and measurable.
If you want your agency to move from AI pilots to AI that actually ships work faster and safer, start with a workflow-first build. Explore Screendragon AI Foundry and see how you can design, govern, and scale practical agent workflows inside real delivery processes.