Back to School with Copilot Pete: Our Mission to End Slop Grenades and Signature Tests

At Greenough Communications, we prioritize training and development. We send team members to major industry events—including Unbound, Advertising Week (up next!), HLTH, and RSA—to learn and contribute to the conversation. We also invest in deeper team learning: each quarter, the entire agency meets in groups for an eight-part training on an assigned topic, and every other month, we host an external guest speaker. Our executive team has completed executive coaching sessions with Elise Mitchell focused on leading and working effectively with the teams we support. We have also held ongoing AI workshops led by Women Applying AI co-founder Pam Boiros, giving our team a shared foundation and common vocabulary as the tools evolve.

This back-to-school season, we added a six-week AI certification with Pete Pachal, former journalist at Mashable who now runs The Media Copilot. Around here, he is known as Copilot Pete. 😉 The course builds toward a demo day where each team member will present a workflow they built for one of their programs or repeated activities. It’s exciting, and we can’t wait to get to the good stuff!

We keep investing in these areas because our brand partners hire us for our judgment, and judgment comes from experience and ongoing practice. When the tools, algorithms, and upgrades change every few months, a team that trains together holds a common standard for quality. That standard is part of what clients buy from us, and I consider it one of the clearest differentiators.

We are one session into our AI certification course and want to share some insights.

“The best AI is the one that knows you” – Copilot Pete.

For years, most AI training centered on prompt frameworks, the formulas that tell you how to structure a request. They still help when a tool knows nothing about you. What matters now is what the tool already knows before you type a word.

Pete showed this by running the same short media pitch request several times, adding context between rounds while leaving the prompt itself untouched. The first draft was generic and full of bracketed blanks. By the last round, it read like something an account lead could review and send, and the prompt was no smarter than it had been at the start. The difference was everything around it.

For communications and marketing teams, this shifts the work left. Investing time in teaching your preferred tool your brand voice, approved messaging, and examples of effective work will pay dividends. Job-specific context, such as a reporter’s recent coverage, should stay with that assignment so it does not influence the next one. A concise brief explaining how you write and what matters to you will also help first drafts sound more like you.

Any output can be wrong.

The software most of us grew up on is predictable. Word does the same thing every time you ask. AI works on probability, so the same request made twice can come back two different ways. That is what makes it useful for creative work, and it is also why a well-grounded answer from the most advanced model can still be wrong or venture outside the brief.

In our work, numbers hurt most. Pete warns that models often transpose digits inside statistics, and his own demo draft carried a percentage he said he had not yet checked against the study. A wrong figure in a pitch costs a reporter's trust, and in a client's earnings materials it costs a lot more.

The habit we are building is to treat every number, name, and link that came through AI as unverified until a person has checked it against the source.

The model's confidence tells you nothing about whether it is right.

Audiences have their guard up, and AI-generated work costs trust.

The public is no longer neutral about AI. Readers, reporters, and clients have learned the tells, from the overused em dash to the "it's not this, it's that" contrast to words like "actually." Once someone spots one, they discount everything around it, even when a person wrote the line. Platforms like LinkedIn and Substack have built ways to flag AI-generated content, and the tells keep shifting as models get tuned away from the old ones.

The bigger risk for agencies is what researchers at BetterUp Labs and Stanford's Social Media Lab call workslop: AI output that looks like finished work but doesn't move the task forward. It reads polished and confident while saying little, runs longer than anyone asked for, or carries a statistic nobody can trace. Shopify CEO Tobi Lütke recently gave the same problem an even better name, "slop grenades," for AI-drafted emails and code sent without being read, which leave colleagues to find and fix the damage.

The cost is measurable. In the BetterUp survey of nearly 1,000 U.S. desk workers, 38% said they had received workslop, and they spent an average of 3.4 hours a month fixing it. What’s alarming is, as an agency, we’re often between two Claudes. Every reworked draft – most of which were not created or regenerated by our own team – cancels out the time AI was supposed to save, and the value is not far behind.

PSA: Anyone want to take Between Two Ferns off the cutting room floor and rename it?! Would be a great YouTube series! 👀

Keeping the human in human-led

Our AI Charter commits Greenough to AI-powered, human-led work. The first session gave us some practical ways to maintain this promise:

• Break the regenerate loop.

• Apply the signature test.

• Keep sensitive information where it belongs.

The rest is still unwritten – Natasha Bedingfield

The tools will keep improving faster than any single course can keep up with. What lasts is a team that knows our brand partners well enough to work as an extension of their own, and that holds a standard for quality that no tool can set on its own. That is why we keep investing in our people and, through them, in the value we deliver to our brand partners.

‍