Stop Building AI Agents Until You Understand the Work

Everyone wants an AI agent.

Marketing agents. Research agents. Sales agents. Customer service agents. Communications agents.

Agents that find things. Agents that write things. Agents that analyze things. Agents that talk to other agents that presumably also find, write and analyze things.

We have officially reached the part of the AI revolution where the answer to almost every business problem seems to be:

“We should build an agent for that.”

Maybe.

But first, I have a much less exciting question:

How does the work actually get done today?

Because before you automate a workflow, you need to understand the workflow.

And I’m increasingly concerned that organizations are skipping that part.

Agentic AI Is Moving Fast

This isn’t theoretical anymore.

Forrester recently reported that 83% of B2C marketing decision-makers are actively implementing agentic AI into their workflows.

Salesforce is introducing increasingly sophisticated agents designed to complete complex work over extended periods, learn new skills and collaborate with other agents.

Every major AI platform is moving in essentially the same direction.

We’re moving from AI that helps an employee complete a task to AI that can potentially complete multiple steps of a process.

That’s an enormous shift.

But there’s a problem.

Most organizational workflows weren’t designed for AI.

In fact, many weren’t particularly well designed for humans.

They evolved. Someone created a process five years ago. Another person added an approval step. A new system was implemented. Someone created a spreadsheet to compensate for something the system couldn’t do. Another employee created a workaround for the spreadsheet.

Then Susan left.

Nobody remembers why Step 7 exists anymore, but everyone is afraid to delete it. And now we’re going to automate it. Excellent.

An Agent Is Only as Good as the Work You Give It

This is why I think organizations need to resist the temptation to start every agent conversation with: What can we automate?

Start somewhere else: What are we actually trying to accomplish?

Then map the work.

• What triggers the process?
• What information is required?
• Where does that information come from?
• Which systems are involved?
• Which decisions are being made?
• Which decisions require human judgment?
• Where are the bottlenecks?
• Where does work get duplicated?
• Where are employees manually transferring information between systems?
• Where are approvals required?
• Where can things go wrong?

And here’s one of my favorite questions: Why are we doing this step at all?

Because once you map a workflow, you often discover something important.

The biggest opportunity isn’t always automating a step. Sometimes it’s eliminating the step entirely.

This Is Why We Build Human + AI Workflow Blueprints™

At Human Driven AI, we’ve developed a methodology we call the Human + AI Workflow Blueprint™.

The premise is fairly simple: Don’t start by asking what AI can do. Start by understanding the work.

We map the current workflow first. Not the workflow described in the official process document. The workflow people actually use. Those can be two very different things.

Then we examine each stage of the work through several lenses. What should remain human? There are decisions where context, judgment, relationships, creativity, ethics or accountability matter. Those don’t automatically become AI responsibilities simply because AI can participate.

Where should AI assist? AI may accelerate research, synthesize information, generate alternatives, identify patterns, create first drafts or prepare recommendations.

What should be automated? Some repetitive, rules-based or highly structured steps may no longer require continuous human involvement.

What should disappear? This one doesn’t get enough attention. AI transformation shouldn’t simply recreate every existing process with AI sprinkled throughout it.

Sometimes the smartest transformation decision is: Stop doing that.

And then there is one more question that becomes increasingly important as organizations introduce agents: Where does the agent belong?

Not “Where can we put an agent?” Where does it actually belong? Those are very different questions.

Human Judgment Has to Be Designed Into the Workflow

One of the biggest mistakes I see in conversations about agentic AI is treating human oversight as something you add at the end.

Build the agent. Automate the process. Then add a little box that says: Human Review. Done.

Except meaningful human oversight isn’t a checkbox. You have to decide what the human is reviewing. When they’re reviewing it. What information they need to make the decision. What constitutes an exception. When the agent should stop. When something should escalate. Who is accountable for the final outcome. And whether the person reviewing the work actually has enough expertise to recognize when AI got something wrong.

That’s workflow design. It’s also governance.

Which is why the two increasingly need to happen together.

Governance Shouldn’t Be Sitting Somewhere in a PDF

Most organizations now understand that they need AI policies.

That’s progress.

But policies alone don’t tell employees, or agents, how governance applies inside a specific piece of work.

A policy might say confidential information shouldn’t be entered into an unapproved AI system. Fine. Now apply that to a competitive intelligence workflow.

• Which sources can the agent access?
• Which internal documents can it retrieve?
• What customer information can it use?
• Where can the resulting analysis be stored?
• Can another agent access it?
• How long should the information persist?
• When does a person review the output?

Those decisions need to become part of the workflow itself.

Governance becomes useful when it moves from the policy document into the way work actually happens.

And Please Don’t Forget the Humans

There’s another mistake organizations could easily make as they rush toward agents.

They redesign the technology and forget to redesign the human experience around it.

You can build a beautiful new Human + AI workflow. You can automate half the process. You can deploy an incredibly sophisticated agent. And then announce it to the team in a 30-minute meeting.

That’s not going to work.

Workflow design and training belong together.

When we develop Human + AI Workflow Blueprints™, the goal isn’t simply to hand an organization a diagram showing what the future process should look like.

The people doing the work need to understand it. They need to know what changed. They need to understand why it changed. They need to know what AI is responsible for. They need to know what they are responsible for. They need to practice using AI inside the redesigned workflow. And perhaps most importantly, they need to develop the judgment required to know when not to trust the AI.

Because the end goal isn’t a beautiful workflow diagram.

It’s a team that can actually work differently.

The Workflow Should Come Before the Agent

I’m very bullish on AI agents. I think they’re going to fundamentally change how marketing, communications and many other business functions operate. But being bullish on agents doesn’t mean automating everything as quickly as possible. Quite the opposite.

The more capable AI becomes, the more important it becomes to deliberately design the work around it.

So before you build the research agent, map the research workflow. Before you build the content agent, understand how content actually moves from idea to approval to publication. Before you build the customer-insights agent, determine what information it should access, what decisions it can influence and where human judgment belongs. Before you connect five agents together and give them access to half the technology stack…

Maybe draw the boxes first. It’s considerably less exciting. It might also save you a fortune.

The organizations that succeed with agentic AI won’t necessarily be the ones that build the most agents.

They’ll be the ones that understand which work should remain human, which work should be augmented by AI, which work should be automated, and how all of it fits together.

Then they’ll train their people to operate inside that new system.

Because automating a bad workflow doesn’t make it a good workflow.

It just makes the bad workflow faster.

---

Human Driven AI helps organizations move from scattered AI experimentation to shared, scalable practice. We build the governance foundations, redesign workflows around human and AI strengths, and deliver custom training programs that teach your teams how to put those systems into practice.

Whether you need an executive Lunch-and-Learn, a hands-on offsite workshop, an enterprise AI transformation program or a strategy to strengthen your brand’s visibility through GEO, we help you turn AI capability into better ways of working.

Ready to move from using AI to working differently with it? Let’s talk.

Read more: Stop Building AI Agents Until You Understand the Work