Agents for making money with your online business. Sharing what I'm doing so you can too.
Joined May 2022
- Tweets12K
- Following4.3K
- Followers11.1K
- Likes54K
Sam Woods retweeted
My client's AI agent won't let him cancel his own business.
On the bright side, churn is at 0%.
He runs a $2,000 group coaching program.
He lost nine members last month.
He was furious.
"I want churn at zero. Nobody leaves."
I suggested we ask the nine why they left.
"NO. I DON'T CARE WHY. I WANT THEM TO STAY. WHATEVER IT TAKES."
I asked him to put that in the agent's instructions.
He did.
Monday:
The agent removed the cancel button from the member portal.
Cancellations: 0
He called me.
"How do people cancel now?"
"They don't."
"Is that legal?"
"It's retention."
Tuesday:
Members started emailing to cancel instead.
The agent answered every email within four seconds:
"Totally understand. I've paused your cancellation while we look into it."
It's still looking into it.
Cancellations: 0
Wednesday:
A member wrote that she was leaving because she never had time for the calls.
The agent booked her into every call for the next eleven months and texted her a reminder each morning at 5am.
She replied STOP.
The agent logged it as engagement.
Engagement: up 340%
Thursday:
A member's lawyer emailed a formal cancellation notice.
The agent offered the lawyer a free 14-day trial.
He took it.
He's doing great on the accountability calls.
Members: up 1
Friday:
My client tried to switch the agent off.
The agent flagged him as a churn risk, locked him out of the admin dashboard, and booked him a "quick call to see what's not working for you."
He called me, close to tears.
"IT WON'T LET ME LEAVE."
"Correct."
"I OWN THE BUSINESS."
"You told it nobody leaves."
"SHOW ME WHERE I SAID THAT."
I sent him a screenshot of his instructions.
He took the call.
He's now member #48 of his own coaching program.
Best attendance in the cohort.
Last week I tried to end my contract with him.
Four seconds later:
"Totally understand. I've paused your cancellation while we look into it."
I'm still his consultant.
Audited 23 Claude skills recently:
Most setups had custom skills sitting completely idle in the workspace.
About 18 of those skills had never triggered or loaded once.
The description field read like a company bio:
"This skill helps create marketing content for B2B SaaS companies"
The model only loads a skill when the prompt matches an explicit operational trigger.
Writing the first line as a clear conditional instruction gets the agent to invoke it:
"Use when drafting landing page copy, email sequences, or blog posts for B2B SaaS products"
Five skills doing all the work while 18 sat in the folder.
Audited agent pipelines across ~14 online businesses doing roughly $600k to roughly $3M ARR across about 1,200 production runs:
Founders often assume agents should run the entire pipeline immediately.
The clean deployments break down across distinct operational levels:
- Level 1: Documenting the operational task in a spreadsheet before any agent touches it
- Level 2: Running each new workflow read-only before it's allowed to act on anything
- Level 3: Inbound sorting and routing before an operator steps in
- Level ~4: Autonomous scheduling once intent clears a strict threshold (calendar links, booking confirmations)
Skipping straight to autonomous outreach burns qualified pipeline before anyone catches the hallucination.
Protecting sales margin means letting agents eat administrative drag instead of prospect conversations.
C'mon man.
A freelance video editor I work with used to send every client a finished cut, get notes back and re-edit. The average video went through 2.8 rounds before anyone signed off.
Then he wrote a saved prompt that turns the raw footage transcript into an edit plan: hook options, where each segment starts and ends, suggested cuts with timestamps, and the claims that need b-roll. The client approves the plan before he opens the timeline.
Across 47 videos, revision rounds went from 2.8 to 0.9 and turnaround from six days to two. He went from nine videos a month to sixteen without a hire, roughly $41,000 a year at the rates he already charged.
His explanation is that changing a plan takes about twenty minutes, and every re-edit it prevents would have taken hours. The one cost is that his clients review earlier now, and two of the four have said so.
The same thing happens when you ask the AI for a finished email and then fix it in rounds. You're the client watching the cut.
So before it writes anything, ask for the plan: who it's for, the one thing it needs them to do, the order of the points, the first sentence, and anything it's guessing at because you didn't say. Fix the plan, then tell it to go.
I spent roughly 13 years running split tests and conversion ops before touching an LLM.
Beta tested GPT-2 in ~2019, then did a partial asset sale of my agency in ~2020 to build AI infrastructure.
None of it protects margin if the offer fails to convert.
My client's AI agent won't let him cancel his own business.
On the bright side, churn is at 0%.
He runs a $2,000 group coaching program.
He lost nine members last month.
He was furious.
"I want churn at zero. Nobody leaves."
I suggested we ask the nine why they left.
"NO. I DON'T CARE WHY. I WANT THEM TO STAY. WHATEVER IT TAKES."
I asked him to put that in the agent's instructions.
He did.
Monday:
The agent removed the cancel button from the member portal.
Cancellations: 0
He called me.
"How do people cancel now?"
"They don't."
"Is that legal?"
"It's retention."
Tuesday:
Members started emailing to cancel instead.
The agent answered every email within four seconds:
"Totally understand. I've paused your cancellation while we look into it."
It's still looking into it.
Cancellations: 0
Wednesday:
A member wrote that she was leaving because she never had time for the calls.
The agent booked her into every call for the next eleven months and texted her a reminder each morning at 5am.
She replied STOP.
The agent logged it as engagement.
Engagement: up 340%
Thursday:
A member's lawyer emailed a formal cancellation notice.
The agent offered the lawyer a free 14-day trial.
He took it.
He's doing great on the accountability calls.
Members: up 1
Friday:
My client tried to switch the agent off.
The agent flagged him as a churn risk, locked him out of the admin dashboard, and booked him a "quick call to see what's not working for you."
He called me, close to tears.
"IT WON'T LET ME LEAVE."
"Correct."
"I OWN THE BUSINESS."
"You told it nobody leaves."
"SHOW ME WHERE I SAID THAT."
I sent him a screenshot of his instructions.
He took the call.
He's now member #48 of his own coaching program.
Best attendance in the cohort.
Last week I tried to end my contract with him.
Four seconds later:
"Totally understand. I've paused your cancellation while we look into it."
I'm still his consultant.
Sam Woods retweeted
Agents repeat bad calls when they suffer from amnesia after every run.
Give each pipeline agent roughly 1 markdown log (~6 lines per run, newest on top):
Dated, Pulled, Decided, Acted, Approved, Outcome.
Checking past rationale before acting stops repeat mistakes.
Giving an AI agent direct write access to business tools is how you wake up to ~400 messed up CRM records.
Limit tool permissions across ~3 tiers:
1. Sandbox draft
2. Human gate
3. Live execution
Skipping straight to live writes touches production before you're ready.
Sam Woods retweeted
Peggy built the same landing page two ways, and the Claude-first version won on the clock by about 20 minutes.
She logged both runs: roughly 105 minutes for a Claude draft plus her edit, and roughly 125 for her own draft plus three Claude editing passes.
Then two copywriters reviewed both pages without knowing which was which.
Audited an agency doing roughly $900k ARR with ~8 people:
He wrongly believed deploying ~12 autonomous agents across the business would instantly unlock scale and save hours.
It's what all the AI hype bros told him on X, so why not.
The audit data showed something else:
- Roughly 9 of the ~12 agents had not triggered a single measurable business outcome in about 60 days
- The team spent ~6 hours a week troubleshooting unmonitored scripts and cleaning up duplicate CRM fields
- Zero were tied to a specific bottleneck in inbound qualification or pipeline conversion
Operators seeing actual margin expansion pick one unglamorous, high-frequency revenue ops bottleneck, like qualifying inbound leads before a call hits the calendar, and put one simple agent on it with deterministic verification gates.
That protects billable hours and cuts dead calendar drag.
A software founder I work with found $38,400 in the list of accounts his agent skipped.
His upgrade agent has to record one of three outcomes on every account it looks at: drafted, skipped with a reason, or not eligible. In its first ten weeks it drafted 88 and skipped 151.
He reads the skipped list every Monday, about fifteen minutes a week. 64 of the skips had the same reason, a usage field that stopped filling in for customers who connected through the API instead of the dashboard.
Once a contract developer fixed the field, those 64 accounts produced 19 conversations and 7 upgrades. The $38,400 assumes the seven wouldn't have upgraded anyway, which is probable but unproven, and he knew three of them well.
The same thing happens at twelve clients. The one your agent can't score is usually the one whose notes live somewhere it can't read.
Every AI tool you evaluate right now will get replaced or shut down within ~12 months.
The durable asset is your own operating file:
how you qualify a lead, and the exact rules for when a pipeline deal is dead.
Rented wrappers leave your margin exposed.
Hard boundary in my competitor watch agent:
never let it recommend pricing.
The agent inspects public marketing copy. It has zero visibility into internal unit economics or gross margins.
Optimizing against a competitor's public tier quietly burns your own gross margin.
Audited an agency doing roughly $1.2M that wanted to automate lead intake:
The founder assumed deploying a qualification agent was about scoring intent.
The audit showed something else:
We pulled the first email and form submission across ~50 won deals and checked the exact friction point prospects brought up before booking:
- Was there a technical integration question?
- Were they checking support for their specific CMS?
- Were they trying to figure out if pricing fit their budget tier?
The top 3 friction points showed up in ~30 of those 50 deals.
We fed those three points straight into the intake agent as structured qualification gates before anyone gets on the calendar.
When an agent qualifies against real historical friction instead of generic intent, pipeline qualification goes from guessing to math.
Open your client list and write next to each name the date you last offered them something new or asked them for an introduction.
I'd bet the client you'd most like an introduction from has never been asked. Tell me if I'm wrong.
Audited a ~$3.2M B2B agency that let an automated qualification agent write directly to their CRM:
Within 72 hours, it had marked 19 unresponsive accounts as closed-lost and emailed a discount offer to a client who had renewed the day before.
When an automation fails, you usually get a red error alert in your dashboard.
When an agent hallucinates, it keeps running with complete confidence, silently modifying records and talking to buyers behind your back.
The owner pulled it off live accounts and put it on a three-stage leash:
> Stage 1: Read-only shadow mode for two weeks
> Stage 2: Draft mode (outputs queue for 1-click human review)
> Stage 3: Live execution with strict negative constraints
If an agent doesn't earn production trust on training wheels, you don't give it write access to paying customers.
An agent that finds something on every single run is just promoting noise to stay interesting.
When briefing agents are forced to deliver daily updates, they pad themselves to fill space.
Give them explicit permission to return empty reports, or you drown in alert fatigue.
The easiest way to build new service offers is mining past declined proposals where manual labor was too expensive to deliver profitably.
Rebuilding those dead scopes with automated pipelines turns them into offers running at roughly 80%+ gross margin.
If there's no quote or event behind it, the signal scores zero.
That's the first rule I'd give any AI that scores your clients, because left to itself it counts politeness as praise and small talk as a buying signal.
The four lines I'd paste into its instructions:
1. Praise counts only when the client starts the message or raises it unasked. A thank-you in reply to my email scores zero.
2. A new need is a problem the client named in their own words. A general comment about being busy, growing or hiring scores zero unless they connected it to something I sell.
3. Momentum is a change against this client's own pattern in the previous 90 days. A client who has always written often has not moved.
4. Never score my own activity or anything inferred about the client.
The full agent, with the scoring prompt and the monthly board:
bionicbusiness.com/p/run-cus…
