By: Betty Rodriguez-Hakes and Catherine Clay
Most marketers have already incorporated AI into their daily work. In fact, the AMA report “Agentic AI and the New Architecture of Marketing Relationships” reports that 62% of surveyed organizations are at least experimenting with AI agents, and 23% are scaling agentic systems in at least one business function.
Jorge Perez, President & CEO YMCA of Greater Cincinnati, and Maggie Ma, Co-Founder, daoflo, joined us on August 19th during the Digital Community meeting to show what closing that gap actually looks like in their organizations and for their customers.
Both speakers opened the same way: everything you’re about to see was made with AI, and here’s how. Jorge’s slides, photography, and copy were generated by Claude working from a trained brand voice. Maggie built her deck as an HTML file, along with a custom editing tool to build it. Neither of them buried that. That transparency turned out to be the through-line of the whole session.
Here’s what marketers should take away, and what to do about it.
1. The seat is shifting from creator to curator
Most marketers built their careers as creators. You look at the data, you make the thing, and your output is capped by your own bandwidth. Jorge’s argument is that the job is moving to curation. You teach the AI to follow your lead, then you direct, review, and refine. His examples:
- A six-part board training video series on financial oversight. Script from ChatGPT, slides from Claude, voiceover and music from an innovation lab. Assembled in one weekend, after two years of not getting it done. “I couldn’t create them before,” he said. “I’m still necessary. But I can move faster now.”
- The organizational version of this is his marketing agent which is named Marlo. Staff push flyers, social posts, and board documents through Marlo for voice alignment before anything reaches the marketing leaders for approval. The human bottleneck disappeared, and the brand voice got more consistent, not less.
2. Stop adding AI to the job you already have
The trap Jorge named for himself: the natural instinct is to take an AI idea and slot it into your existing process. The better questions are: would you design the job this way if you were starting from scratch? Would you still need the same people to execute?
His illustration: The YMCA has a program that helps kids get ready for college and the team uses AI to help perform tasks faster. The goal is serving tens of thousands of students with an agent that answers questions in real time, around the clock, in 150 languages. Same mission, completely different scale, and you only get there by redesigning rather than retrofitting.
For marketers: don’t ask how AI speeds up your current campaign process; ask what campaign process you’d build if coverage, exploration, and execution were nearly free.
3. Know what stays human
Maggie drew a clean line between what AI can do now and what still has to be done by a human.
AI creates new operational capabilities: coverage (read every source, not a sample), tooling (a two-week build becomes an afternoon), exploration (five approaches for the price of one), execution (the same task across thousands of items).
Four things stay human: taste, evaluation, judgment, and accountability.
Taste is knowing what should be good before you can say why. It isn’t innate, it’s trained over years of practice and part of taste is quickly catching AI when it’s confidently wrong. As Maggie put it, one test of whether you’re AI-smart is whether you can spot the mistake no matter how eloquent the output sounds.
Evaluation can be partly delegated, but only after a human thoroughly defines the standard. Judgment requires full context that the machine doesn’t have. Accountability doesn’t transfer. No matter how much of the work AI does, a human is still the one who answers for the result.
4. One human, six agents
daoflo’s propriety business development workspace runs with one person and six parallel agents: scouting new leads, prepping client calls, searching public RFPs, re-engaging lapsed accounts, researching a person or organization, and prepares offline meeting notes.
The split is deliberate. Humans set the target profile and lead qualification criteria. Agents crawl, aggregate, cross-check across Drive, email, and Slack, and draft agendas, one-pagers, and proposals. Humans approve before anything is saved or sent.
Result: 30 to 50% more throughput. But Maggie made a sharper point. The real gain wasn’t that the work got faster. It’s that work which previously never happened at all now happens, because there was never enough time for it.
And six agents is not a ceiling. You can run several agents at once, including two call preps simultaneously, which means the practical limit is your capacity to review, not the agent count.
5. Verification is the new bottleneck
This was the line that landed hardest:
AI did not create a shortage of production. It created a shortage of verification. Whatever your team cannot check is what is actually limiting your capacity.
Which leads to a useful equation for weighing any AI delegation decision:
Value of delegating to AI = time saved − (cost to specify what good looks like + cost to verify it is right + odds of being wrong × cost of being wrong)
Every variable except time saved is a human cost which means the real lever isn’t the AI, it’s how you invest in specifying, verifying and catching mistakes before they go out.
6. Check your readiness before you build anything
What determines whether you move from individual use to something that actually builds team capacity isn’t budget or headcount — it’s readiness. As Maggie put it, organization size, budget, team AI maturity, and vertical are all additive. Readiness is a multiplier, and multipliers can go negative. An unready team with a big budget just produces unverified work faster, and erodes trust in the process.
Five readiness tells. If you score 3 or below, don’t build or buy yet:
- You can name what “good” looks like, with examples you can point at.
- You can describe the workflow end to end.
- Leadership and staff agree on what AI is for.
- There’s a safe, low-stakes, real place to fail.
- You know what things cost today.
Budget, team size and AI maturity are all slow to move. Readiness isn’t. It’s the one lever you can pull in approximately three weeks.
Action steps for marketers
Write down one workflow this week. No AI involved. Event promotion, monthly newsletter production, speaker sourcing, pick one and document it end to end. This is step two of Maggie’s five-step pathway and it’s where most rollouts stall, because it feels like homework.
Delegate only the cheap-to-verify steps. Narrow, boring, high frequency. If checking the output takes longer than doing the task, you picked the wrong step.
Build the verification asset. A rubric, a few golden examples, and a clear description of what bad looks like. This is the piece that compounds. Every workflow after it costs less.
Audit the source files. It’s not enough to train the voice once, you have to keep curating what goes into the training set. The file could have all of the correct information but could be written in a tone that doesn’t align. Bad inputs can degrade the output from the agent over time.
Put the right human in front. Not necessarily the enthusiast, but the subject matter expert (SME) who can look at an output and instantly know it’s wrong.
Build an agent instead of writing prompts. A pre-prompted agent skips the “you are an expert in…” ritual every single time. Jorge’s team runs 6+ of them for about $20 a month.
Label every handoff. Verified means you stand behind it. Reviewed means you checked it for errors. Raw means it’s a draft and the recipient should verify. Then run the three questions before anything goes out:
- honesty (are you comfortable if the recipient knew how this was made?)
- harm (are you comfortable if this were done to your community?)
- legitimacy (are you comfortable if everyone in the industry did this?)
Hot tips from the Q&A
Get into LinkedIn through Chrome. Several attendees hit the same wall: Claude says LinkedIn isn’t accessible. It isn’t, directly. Log into LinkedIn in Chrome, then let Claude take over the browser. You’ll literally watch the mouse move and type. Jorge used the same trick on his organization’s intranet, told Claude to read the whole thing and build skills from the 21 leading practices, then went to dinner. It was done when he got back.
Beat rate limits with scheduled tasks. LinkedIn caps browser searches at roughly ten at a time. Jorge’s workaround: schedule Claude to run the search through Chrome every two hours. Ten more contacts, then ten more, then ten more. Maggie added the detail that makes it practical: Claude Pro includes three scheduled tasks per day at no extra cost, and they run in the cloud, so they keep working after you close your laptop.
Keep a second laptop for agent work. When Claude takes over your browser, it takes over your machine. Jorge runs a dedicated always-on laptop so browser tasks don’t interrupt his actual work.
Train voice by analysis, not imitation. Don’t paste samples and say “write like this.” Feed the AI your writing and ask it to analyze and describe the style back to you. Then tell it to create a voice from that description. Jorge found this dramatically more effective, and he audits the files he uploads to agents so nothing off-voice sneaks in.
Talk, don’t type. Both a habit and a capacity move. Jorge holds voice conversations with his agents on the drive in, so work is waiting for him when he reaches the office.
Use Projects, not endless chats. Projects hold persistent context for long-running work. Jorge currently has Claude managing a capital campaign plan and has used projects for senior leadership hiring.
Calibrate the model to the task. Start on a lighter model for drafting and iteration, then escalate to the heaviest model for a final full-document pass. Top-tier models burn through usage fast.
Let agents check and confirm with each other. Jorge built agents for different functions – Marlo for marketing voice, Harmony for HR. When something needed HR accuracy and brand voice in the same output, Harmony checked its language against Marlo.
A note on AI detection. One attendee raised watermarking and detection tools. Jorge’s Substack flagged a piece as 80% AI-written when it was 100% his outline, his ideas, and his words, proofed by AI for grammar. This is unresolved and worth watching, especially for anyone publishing under a personal or corporate byline. It’s also another argument for declaring your process up front, the way both speakers did.
Coming next month: the Digital Community tackles NotebookLM. Watch our Eventbrite page for registration.
Interested in speaking or hosting a Digital Community session? Email communities@amacincinnati.org.
Disclosure: this recap was drafted with AI assistance from the session transcript and speaker slides, then reviewed and edited by the authors and AMA Cincinnati MarcCom committee. Human in the Loop!
