AI left this M&A broker too much in the loop

Christine McDannell has AI drafting her email, briefing her meetings and cutting her video, and she still ends the day supervising it. The owner of The Magnolia Firm, a sell-side M&A brokerage in Austin, ran five or more parallel AI tasks until exhaustion forced her down to two or three, and her sharpest complaint about the marketing work is that she is "too much in the loop."

For anyone building in this category, that is the interesting failure. Nothing she describes is broken. It is unfinished in a specific place: the last step, where the operator still has to stand.

Why did she build a voice agent and not deploy it?

Latency. She built one for the firm and left it switched off, because the pauses would cost her more than the answering would save.

"The latency though still is ... versus awkward long pauses. Like those are annoying."

Her decision was explicit:

"We didn't fully install it yet, because I'm like: people are gonna just get mad. So let's wait until it's better."

She knows better exists. She has been on the receiving end of a voice agent good enough that she had to ask — "it was so hard for me to tell" — and saw a zero-latency model at a hackathon that switched languages mid-call. Neither is what she can deploy. Her position on disclosure is firm regardless: announce the AI at the start of the call, and let the caller press zero for a person at any moment.

Where does AI marketing automation stop?

One step before it is useful. She can get AI to produce the content. She cannot get it to run the process end to end.

"I want it to go like from A to Z. I want it to actually — it knows my industry. So it does the content, it writes the prompts, it actually posts, or it writes the keywords. Like I have to prompt every time. Okay, can you make sure you research the proper keywords for M&A industry? The ones that are performing best? And then separately: okay, can you write a caption? So it's like piecemeal together."

The cost lands on her calendar:

"Which is not fun, because I'm the human in the loop — but I'm too much in the loop on the social media stuff."

Galina Fendikevich, who hosts Artificial Reality, the AI buyer-evidence podcast, described testing the same problem across many tools and finding no single one covering keywords, branding, carousel design, copy, video clipping, captions and posting — with posting itself full of platform gaps.

What is AI fatigue, and what causes it?

Supervisory load. McDannell went looking for an explanation when she noticed she was exhausted by tools that were supposed to be saving her time.

"The AI brain fry that people talk about — I went deep on it a couple weeks ago because I was getting exhausted. I'm like: AI is supposed to make my life easier, but I'm exhausted by the end of the day."

The diagnosis she got back, from the AI itself:

"It's like you have six amazing interns that are constantly coming to you like: oh my God, okay Chrissy, ready? Okay Chrissy, what do you think of this? What about this? It's constant."

The cause was her own working pattern — five or more parallel tasks and a pinned-chat list she describes as "ridiculous." The advice she took was to hold two or three at a time. Fendikevich raised the adjacent problem of volume: AI-generated reports running to tens of pages, all reasonable, all unreadable in aggregate.

What do these three complaints have in common?

Each tool finishes its own step and hands the join back to the operator. The voice agent produces speech but not a conversation a caller will sit through. The marketing stack produces keywords, captions and clips but not a published post. The parallel tasks produce output but not a queue that manages itself.

None of that is a model-quality problem, which is why more capable models have not resolved any of it. It is a completion problem, and completion is what she is still buying with her own attention at roughly $100 a month of subscriptions against the $40,000 a month of labour she says the setup replaces.

What would you have to build for her to buy it?

Three specifications come straight out of what she described:

• Voice with no perceptible gap, plus disclosure and a zero-out. She has already written the requirement and is waiting for something that meets it.
• Marketing from keyword research to published post, in one pass, industry-aware. Not twelve micro-steps she stitches together.
• A supervision layer, not another agent. Something that batches, queues and holds work until she can look, given that her own fix was to run fewer things at once.

The diagnostic for a founder reading this: your product may already be good enough at the task and still be losing to the operator's evening. Ask how many of your users are the last step in your workflow, and what it would take to remove them from it.

Where this comes from

S1E5: How she cut her payroll by 83% using one AI agent — the full interview with Christine McDannell. Listen or watch: YouTube, Spotify or Apple Podcasts.