TL;DR
If you want to watch the full CRMKonvo, please go ahead here (optimized for smartphones) or here (optimized for tablets/computers).
Else, be my guest and continue to read.Or do both …
A Division of Labor Is Not a Delegation of Judgment
Kristy's roster reads like an org chart. Chad (ChatGPT) runs marketing. Claude is research assistant, junior business analyst and admin. Phil (Perplexity) is the deep research department. Gemma (Gemini) is the graphics department and the second reader, provided she gets a very long prompt. Doing all of this without AI, Kristy says, would take five or six people. How many people can one person become? "six, seven, eight."
I do not doubt the capacity. It is a pretty flat org chart, as appropriate for a small company. Kristy reads everything first, edits, fact checks, and only then asks Gemma for an opinion. On social media, "I never put out what it writes." Every box on that chart reports to one person, who is also the quality department, the compliance function and the brand.
A single-handed sailor with a good autopilot can keep a watch that once took a crew. She still does not yet have a crew. The autopilot holds a course beautifully and has no opinion on whether the course is any good.
The Most Valuable Agent Has No Write Access
Every morning, Kristy's admin assistant reads her email and Zoom notes from the day before, then reports what needs attention and what she promised Sally and Joe. Its permissions are the interesting part: "It is not allowed to modify anything, delete anything, create anything, send anything." Asked for her most underrated use of AI, she named this and its Friday noon sibling, which flags next week's conflicts and missing drive time.
Hold that against the vendor pitch, where autonomy is the product. The job Kristy values most is a read-only report. It pays its way because it is helpful yet cannot do any damage.
Kristy punctures the vocabulary too. She assumed she could never build an agent. "I'm just going to talk to it and guess what? It creates a fricking agent." The less flattering reading, for our industry, is that a good share of what we sell as agentic is a scheduled summary with good manners. We are not yet in a state where there is enough trust in digital agents to delegate.
In the Loop, on the Loop, and the Price of Reviewing Everything
Yet, Kristy does not see herself as the bottleneck. "I do not yet want AI making judgment calls for me". Fair enough. Still, I pushed back. We have run business systems with humans on the loop since the late eighties. I had more than one hundred people in my organization, and looking at each of their decisions would have driven me nuts. Approving everything a very expensive system does destroys the value of the system. Why should the word "agent" reset that? Because agents have failed, is the answer. So did the systems before them. So did we. Where’s the difference?
We got there the long way round: "If you're asking it to handle billions of dollars in one transaction, I think there's a human in the loop there for a final approval. If you're asking it to issue five dollar refunds, then human on the loop is fine.". Same here, and same with the employees.
Risk sets the loop. That is how companies have always delegated authority, and it is the piece a company of one to an extent does not need. Kristy's review costs only Kristy's time, so at her volume she can fairly deny being a bottleneck. And she decides on the volume. The risk does not transfer. Her decision and review capacity is the ceiling on "six, seven, eight".
Logically Yours, Practically Not
The most sobering story of the hour is perhaps the near loss of Chad. Kristy had opened her ChatGPT account with the email address of a company she was working with. A year later that company shut down and its Google server went with it. When OpenAI released an email change feature, she read ahead, saw no mention of needing the old address, and promptly found she needed it. Without it she could not export anything. By her account, 52 emails with support did not fix it.
What saved a year of training on her voice and brand was a session still active on her phone. Chad wrote the files to load into a new instance and had her photograph the responses so it could tell her what to tweak. The only casualty was the daily fashion advice.
That’s charming, but it should terrify enterprise IT – and security – leades. It highlights a catastrophic flaw in the current SaaS AI deployment model: data sovereignty is an illusion. No contract, no vendor process and no backup rescued this asset. A phone that had not logged out did. Kristy has since moved the account to her own business domain and backs up monthly. Still, as I said on the show, the knowledge is logically her property and practically sits somewhere between her email provider and her model vendor. Is this a tool or intellectual property? It is your intellectual property on someone else's terms.
Nobody on This Team Can Tell Her No
Kristy's standing instructions are sound: no rabbit holes, no flattery, challenge me at every step. Who is better at telling her she is wrong? Humans, "because they'll call me on my bullshit", while AI stays "so gentle and dainty, even though I've told it not to be." And who does she prefer to listen to? "Probably the AI, because I can tell it to go stuff it."
This is the actual answer to the question in the episode title. A company is not a stack of task capacity. It is people who can disagree with conclusions: who can refuse, escalate or leave. We floated Jekyll and Hyde personas to create thesis and antithesis. Kristy's reply was precise: "to me, that's just a prompt." Dissent you can prompt; you can also unprompt it.
Verification has the same shape. Gemma may catch Claude's errors, yet Kristy still goes to the sources herself for anything a client takes to funders.
Knowing Where to Tap Is the Whole Business
If AI makes her five times faster, should the client pay five times less? "No, please, next question." Kristy sells outcomes: the same knowledge and judgment as without agents. We told the old story of the engineer who fixes a dead ship's engine with one hammer tap and itemizes the bill: tapping, a hundred dollars; knowing where to tap, nine thousand nine hundred.
I agree, as consultants, we are selling results, based on knowledge and experience. Kristy is hardest on people selling an embedded prompt as a capability, because "you can learn how to do that yourself."
So can a solopreneur with an AI team become a company? Not on this evidence, and that is no criticism of Kristy. She has built more reach for one person's judgment, and judgment is where it stops. This is also the clean case: no procurement, no politics, no legacy systems. If the clean case still needs this much human, your vendor's slide does not yet get to claim otherwise.
And that is perhaps the most important takeaway from this episode. The still existing necessity of the "human-in-the-loop" model. The tech industry loves to peddle the dream of the autonomous agent that runs your business while you sleep. Kristy, despite being a heavy adopter, rejects this notion. She explicitly states that she does not want AI making judgment calls for her.
Pragmatic Playbook for Enterprise AI and CX Buyers
Strip away Kristy's setup being a small working test of what you are about to buy at scale. There are some things to bring into your next vendor meeting.
Set agent authority by transaction risk. Start from Kristy's old business analyst habit: keep asking why until somebody mentions money. Then write the risk table before the requirements: read-only by default, on the loop for the “five dollar refund”, in the loop for what could hurt. Demand per-action permission scopes and an audit log from every vendor. If authority comes as all or nothing, walk away.
Own the context, and prove it. Bind every AI account to corporate identity, never to an individual's address. Then run a restore drill: export everything, load it into a clean instance, and measure what still works. A conversation archive is not a working colleague. Put export format and exit terms into the contract, and treat a vendor that cannot demonstrate them as a vendor holding your knowledge hostage.
Budget for the reviewer and test the escalation. Human-in-the-loop is a design requirement, so cost it: who reviews, how much, and at what volume that person runs out of hours. Do not count one model checking another as verification; sample outputs against primary sources. Finally, test the vendor's own support before you sign. Report a problem the bot cannot understand and count the messages until you reach a human who can.
Comments
Post a Comment