Part 3 of a series on AI and Corporate Affairs, by Andrea Christianson, Partner, Penta Group and Russ Grote, Managing Director, KPMG US
It’s 5PM. It’s been a day. Back-to-back meetings, a surprise media inquiry, and you just finished the launch plan for a tech partnership.
A draft press release hits your inbox. The news is buried, the connection to strategy is wrapped in jargon, and the executive quote is nowhere near their voice.
You were tired. Now, you’re done.
—
It’s 5PM and it’s now been three full days since you submitted a draft op-ed for a leader to your manager. Your inspired idea brought to life in a carefully curated op-ed — that merely represents who you are as a communicator. No big deal.
Email received. “Hey - sry for delay. pretty good. I flipped a few things around and caught a few grammar errors. Feel free to move forward once you clean it up. Thx.”
Thanks bro.
—
Every communicator has been on one, if not both sides of this at one point. Neither failure is about talent, caring, or professionalism. Nobody had the time for a more collaborative process. Now, AI has upended the time constraint, but judgment remains a bottleneck. General LLM AI-generated drafts — even with a human steer — aren’t good.
Building on the knowledge layer, one approach to this new problem is to build agents or skills embedded with best practices for various communications products and hook them to your knowledge layer, automating drafts with people reviewing on the back-end.
This is not the approach we’d take.
A good knowledge layer is not a black box of strategic knowledge maintained in an ivory tower. They are built and maintained through a collaborative and transparent process. Agents or skills should feel the same way.
The better approach puts the agent right in the middle of that process.
The author still writes the whole thing. The agent reads it the way a strong manager would on their best day: is the hook earning attention, does the logic hold, is the evidence real, does this sound like the executive and not a template.
And the experience is radically different. It’s closer to sitting down and talking a piece through — here’s what’s working, here’s what isn’t, and here’s why. It can also be done on an early quick draft to accelerate the iterative process.
Done well, this does two things at once. It tells the author their argument is off-message — a knowledge-layer catch. And it tells them the hook is buried on page two — a craft catch.
At KPMG and Penta, we’ve built several agents this way to improve the quality of press releases, op-eds, blogs, and more. They all have common features:
A detailed review rubric: Op-eds are reviewed for their thesis, hook, evidence, craft, and voice. Press releases for news value, clarity, and executive voice. Rubrics contain detailed expectations for what good looks like — not 101 best practices. For example, the KPMG rubric for press releases is ~4,000 words.
Coaching guides: Agents aren’t rewriting, they converse. Guides inform how an agent coaches, emphasizing iterative shorter rounds of feedback focused on one component of the product at a time. Feedback explains “the why” so the communicator can interrogate the feedback.
Examples: Examples of what good, okay and bad looks like anchor the agent’s reviews. These examples also include a similar range of feedback examples. These examples can help ground the agent.
Most importantly, agents all have one specific hard rule in the agent’s instructions — Never rewrite.
The mechanics will vary based on the IT environment. For example, an environment of CoPilot Studio with tighter context windows, these files — along with the knowledge layer — serve as the agent’s context and skills, while agent instructions enforce rules and build in triggers to tap into certain knowledge and skills. The art is tightening files, eliminating overlap, inconsistency and gray in the documents to keep the agent from drifting from your goal. For other set ups with larger context windows, the mechanics are different.
Agentic collaboration demands better person-to-person collaboration. Understanding agentic coaching can elevate best practices within a product, team members can spend more upfront time on strategy, audience, and messaging (all connected to the intelligence layer). Then with better drafts, coaching focuses on how that strategy and messaging nuance are applied in practice, reflecting how various stakeholders — an employee, investor or regulator — might read a phrase or argument. Coaching can also spend time on “what’s next” to give that content more legs.
This approach doesn’t just extend judgment with AI but facilitates a process that grows it faster.
We’ve covered the knowledge layer and agents but how does an organization actually transform? Who builds and maintains this stuff? And once you’ve transformed, how does the new organization loom and operate?
Stay tuned for our next piece.