Building AI around the way marketers work

For one of our pharmaceutical clients, developing campaign concepts took weeks of avoidable iteration. Concepts would either land too broadly or too narrowly, and valuable research performed upstream disappeared as teams condensed it with each round of revision.   

This meant marketers had increasingly less confidence in where to focus their creative energies.  

To minimize ongoing revisions and keep decisions informed by insights, we built a workflow of three AI agents, each with one job, inside an interactive workspace within the client's secure environment. Marketers entered the brand's inputs, watched the agents work, and reviewed the results.  

1. The first agent read the brand's consumer research and identified the attitudes and habits that stop consumers from acting, so the upstream insight stayed in play instead of shrinking with each round. 

2. The second developed campaign concepts around the brand's strategic priorities, the long-term objectives every initiative must advance. 

3. The third scored each concept against the client's own evaluation criteria: alignment with the brand's core positioning; strategic fit; creative strength; and feasibility across paid, earned, shared, and owned channels. Every score came with a written rationale, so marketers could see why a concept was strong or weak. 

The agents did the analysis, while the marketers made the calls. This meant marketers could shape and refine the direction of the concepts as they worked, using the analysis to focus on the opportunities they judged most promising.  

To gauge the effectiveness of this novel approach, we encouraged the client to judge the output the way they would any agency's work: blind.   

During a live ideation event, the AI-generated concepts were evaluated alongside ideas from five major advertising agencies, with no indication of which came from where. The client's marketers chose two of the AI-generated concepts to take forward.   

The ideas held up because the marketers had shaped the system that produced them. We studied how they judged a strong concept, which research they trusted, and what each brand priority meant in practice, and then designed the agents to work from that knowledge.   

The quality of the output is the quality of the thinking you give it.  

In a previous article, “Why AI pilots hit a wall,” we explored why useful AI outputs so often fail to become useful work. Here we’ll consider the importance of designing AI around how marketers work, because simply adding a generic AI tool won’t yield the desired result.    

Leveraging your team’s expertise 

An experienced marketing team may be starting from zero with connected AI, but it certainly isn’t starting from there with expertise. Team members now their customers, their brand, their standards, and what makes an idea worth defending.   

Instead of asking marketers to relearn their work around a new tool, it’s better to first understand how their work happens:   

• The inputs they depend on (where context is acquired) 

• The internal steps they take to perform their tasks (where information is processed/refined/enriched) 

• The downstream handoffs (where context can disappear) 

• The delays, exceptions, escalations, and decisions that can sometimes arise (where governance and policies need to be enforced) 

Too often, teams are presented with a false choice: Fit the work into a standard tool or commission a custom build with a long list of things to invent.   You don’t have to invent this from scratch, and you should be cautious of anyone who says you do. Two things need to exist before a first AI workflow is useful:   

1. An honest picture of how the work runs today 

2. A foundation the agents can run on safely 

For the first, we use Credera Mirror™, a simulation environment we built to take the process documentation and timing data a team already has and show where effort goes and where work waits.   

Then we test alternatives before anyone changes a job. Could research happen in parallel? Could an agent prepare the materials a reviewer needs? Would removing one handoff create more value than speeding up five other tasks?   

Any team can ask those questions of its own process, with or without a tool. The point is to ask them before building.  

For the second, we use Credera AgentOS™, the reference architecture we built for running multi-agent workflows and the one we run our own work on. We built it because we needed to prove the method on ourselves before asking a client to trust it, and because starting from a tested foundation gets a team to a working workflow in weeks rather than quarters.   

Whatever foundation you use, every task an agent takes on needs the same five things:  

1. The right context 

2. Governed access to tools and data 

3. Guardrails on what it may do 

4. Checks that evaluate its output 

5. A clear decision right that hands the call to an accountable person when judgment is required 

 If you can’t point to all five for a given task, that task isn’t ready to be delegated.   

A start: One workflow you want to make better 

You don’t need a finished AI roadmap to get started. Bring us a recurring piece of marketing work, the materials your team uses today, and the people who know where it gets stuck. Together, we can identify a first workflow worth building, what it needs to connect to, and the result that would justify expanding it. 

Then measure that result in the work: useful ideas taken forward; fewer corrections; or a brief the next team can act on sooner. Then extend what proves valuable. 

You bring the knowledge of your brand, and we’ll bring the foundation and the team to turn it into a way of working you can use. 

Contact our AI specialists to start designing an AI system around your brand. And stay tuned for our next piece in this series, an ebook that will provide an in-depth look at our approach.