How can comms leaders measure and prove the ROI of AI?
Define a reach goal that maps to your company's OKR. Say the OKR is to increase revenue, so your communications goal is to increase awareness and top-of-funnel leads. The goal and KPIs then become increased media coverage and the work needed to achieve that coverage.
Step 1: Tie your AI plan to a company OKR and a marketing KPI
- Pick one result to lead with: more coverage, less brand risk, faster turnaround, or new capabilities.
- Write a hypothesis: if we have better visibility into trends and reporters, we'll land more coverage, and more coverage means more top-of-funnel awareness.
- Name the jobs that test it: scan and trendjack, draft better pitches, find the right reporters.
Step 2: Map your PR workflows and their dependencies before you automate
Before you automate anything, write down what every step leans on. For pitching, that's templates and media lists for the email, monitoring for the feedback, a framework for the insight, and context for the next pitch. The details matter. If your media list doesn't mark who's a feature writer, the agent doesn't know you only want feature writers.
Step 3: Use AI evals to test and improve your prompts, templates, and context
An eval is a fixed test you rerun every time you change something. I could spend a full session just on evals, but in short:
- Collect real examples of past work and how it landed.
- Write a yes/no checklist. A machine is binary, so you have to take something contextual and make it binary.
- Change one thing at a time (a prompt, a template, a style guide) and rerun. Did it get better?
Step 4: Scale AI agents in stages and measure ROI against a day-one baseline
Record your baseline first. Then move up one level at a time, and only when the level before it passes:
- Human-led: agents draft and find reporters, and you review everything.
- Agents: a reviewer agent grades the work, and you check that its grades match yours.
- System: the agents hand work to each other, replies feed back in, and you spot-check.
At month six, show each number next to where it started: placements and reply rates, off-brand drafts caught, hours per pitch, new things launched. If each stage got to its quality bar faster than the last, say so. That's your evidence of compounding.
And you don't have to remove yourself completely. Knowing when to review and when to automate is part of your judgment and the learning process.
What does a working AI system look like for a comms team?
Once you have one system, other things start to become possible. At my consultancy, Scribes, the downside of being a consultant was always that when I'm done, I'm gone. Nobody can tap me on the shoulder and ask what the executive said. So I built a second brain from my conversations with the executive, which the whole client company can now query. It's living context. What we haven't talked about in a year gets pruned away.
That's the real answer to “what's your plan for AI?” It's a plan for a stronger comms team, built one measured step at a time. And judgment is still at the center of all of it. It just shows up in what you choose to evaluate, and where you decide to stay in the loop.