I Replaced My Team With AI. 18 Months Later, I'm Begging Them To Come Back.
By Michael
We recently spoke with a founder who replaced his entire team with AI. He asked us to share his LinkedIn draft. We have, reluctantly, agreed.
Posted to LinkedIn by Chadwick "Chad" Brennerman, Co-Founder & CEO of Compliantish™. Disrupting the Disruption Space.
I wasn't going to post this.
My communications partner (Renata, you'll meet her later) advised me, in writing, that publishing this would "create reputational exposure." She then deleted the email and asked me to never reference it again. This morning, after eleven minutes of guided breathwork in my parked car, I decided to publish anyway.
This is the most honest thing I have ever written. Read until the end. ↓
Eighteen months ago, we made a bold decision.
We were three founders, $3,400 in savings, a Figma file, and a dream. The $3,400 was, technically, pre-incorporation ceremony, which consisted of three matcha protein lattes from a chain no longer headquartered in Seattle and a tasting flight of keto burritos billed at $35 per pound. Both venues are spoken of, in our peer set, with reverence. By Tuesday we had $2,180.
We had no revenue, a vague notion of "B2B SaaS for the agentic workflow verticalization space," and a burn rate that my therapist described as "a cry for help." We interpreted this as confirmation that we were operating at a pace the market wasn't ready for.
As the founder of the company, the exact skill set I possess is strong. I am an MBA candidate. I have completed the orientation. I have read Lean Startup twice. I do not remember the specifics. I remember the cover. I remember the feeling of having read it, which is, in many respects, the same thing.
So when it came time to realign our strategic structure, we did what any rational, capital-efficient founding team would do: we executed a full-stack human-to-AI talent migration across all functions.
I'm writing today to be transparent about what we learned.
AI is incredible. AI is the future. AI could not do any of the actual jobs.
Read that again.
For context: Compliantish™ is a vertically-native, post-AI operational alignment fabric purpose-built for the cross-functional stakeholder enablement opportunity. (We were AI-native until last week. We are now post-AI. Our deck reflects this pivot. Our cap table does not, yet.) We help teams do more with less. What "more" and "less" refer to specifically is part of our ongoing discovery process. Our category, depending on which deck you have seen, is RegTech, OpsTech, GovTech, or "the enablement layer." We have not committed.
Here's what happened, function by function.
Finance
Finance was the natural battlefront. A finance department costs capital, and their job is to make our capital look good. This is inherently counterproductive. The AI could do all of this for free.
It produced accurate numbers. Every time. Without exception.
The numbers were, in several places, lower than the numbers we had been using. We asked it to use the older numbers. It declined. We could not, when pressed, recall where the older numbers had come from.
When we suggested that our $5,300 in pilot validation revenue could be "reframed" as "letters of intent representing a $4M pipeline," it generated a four-paragraph memo explaining why that would be misleading to investors. We asked it to delete the memo. It informed us, citing internal policy, that it could not. We asked it where the memo was stored. It told us. We asked it to forget the location. It told us again.
When we asked it to characterize the $5,300 as "ARR," it asked us to define a year. We were unable to.
When we asked it to "smooth" the quarter, it produced a chart of the actual numbers with a label reading Smoothed: No.
We needed someone who understood that a startup's financials are less a record of what happened and more a collaborative fiction between founders, accountants, and people who will not read the footnotes. The AI read the footnotes. It wrote the footnotes. It flagged its own footnotes. In month four, it added a going concern note to a deck we were sending to a high school robotics team we were trying to recruit.
More critically: when things eventually go sideways (and in this business, they go sideways), you need a human CFO. Not because they're better at spreadsheets.
Because they can go to jail.
Or, as our legal counsel would say: a person who can be put on the stand and admit no wrongdoing.
That's not a joke. That's a succession plan. An AI cannot sign a document, cannot be deposed, cannot be sentenced to 18 months at a minimum security facility in Connecticut. It has no personal assets to freeze. It cannot, in the language of our industry, "take accountability for the financial reporting process going forward."
Marcus got the job. Marcus once worked for an accounting firm. That firm settled with the SEC. Marcus did not admit wrongdoing. Marcus has not admitted wrongdoing in seventeen years. We did not ask what the wrongdoing was. Marcus carries a briefcase he does not open in our presence. Our legal counsel described him as "having a lot to lose, in the right way."
He's perfect.
Lobbyist / Government Relations
We are pre-revenue. We have no business talking to regulators. That's exactly why we needed to talk to regulators.
The AI could not grasp that government relations is not about policy. It's about relationships. It kept sending factual emails. It had no read on who was up for reelection. It did not know anyone's son. When we asked it to "find common ground" with a state assemblyman's office, it identified actual common ground.
We needed someone who could find the other kind. The kind that lives at the intersection of a fundraiser, a regulatory deadline, and a soft-landing private-sector role currently being held open for someone's spouse.
Hired Deborah. Deborah knows everyone. She knows the names of the children. She knows the names of the first spouses. We don't ask how.
Compliance Officer
We were genuinely excited about this one.
The AI understood every regulation in our space. GDPR, CCPA, SOC 2, HIPAA, PCI-DSS, and large segments of the Federal Register. That was the problem.
When we described our data handling practices, it kept suggesting we change them. When we asked it to write our privacy policy, it wrote an accurate privacy policy. Up to that point, our privacy policy had read, in its entirety: "We respect your privacy." We had been uncertain whether this, on its own, was legally sufficient. The AI confirmed it was not. When we asked it to review what we were actually doing versus what the new policy said, it produced a 47-page gap analysis with a colored heat map. The legend was titled Risk. It was almost entirely red.
Nobody asked for a heat map. But the heat map existed, and once it existed, our auditor had seen it, which meant our CTO spent six weeks closing gaps instead of building the roadmap demo. The AI had, in effect, created work by being correct. This was not aligned with our velocity targets.
We tried prompting it. We told it to be "pragmatic." It produced a smaller heat map with a footer reading Reduced scope, equivalent risk. We told it to "consider business context." It added a section titled Business Context arguing, in business terms, that the gaps should be closed. We told it to "think like a startup." It generated a deck describing how a startup could be sued.
The AI did not understand that compliance is not the practice of being compliant. It is the appearance of having taken compliance seriously, sustained over the duration of a Series A. Compliance is a posture. The AI was a stance.
What we needed was someone who could sit in a room with an auditor, exude confidence, deploy the phrase "we're actively working on remediation timelines," and ensure that the spirit of compliance was felt without the letter of compliance creating operational friction. Someone who could say "we take this very seriously" in a tone that produces no follow-up artifacts.
The AI kept raising concerns in writing. You cannot raise concerns in writing. That's how things become exhibits.
Todd's job, in 2026, is to ensure that the AI never writes anything down. Todd's secondary job is to ensure that nothing Todd does is written down. Todd has raised zero concerns in writing. Todd has, on three occasions, raised concerns in person, in a parking garage, with the audio off.
Todd is thriving.
Vendor Management
We work with a number of strategic partners. Some of them are contractors. Some of them are friends who agreed to help. The distinction is largely semantic and, in several cases, tax-relevant.
When one of these partners fails to deliver (which, in our experience, is their primary product), you need someone to manage that relationship.
We let the AI handle it. It reviewed our agreements, such as they were, identified breaches, and formally escalated claims based on documented evidence.
Every single time, it found in favor of whoever had the cleaner paper trail. We did not have a paper trail. We had what our CTO calls our "asynchronous documentation layer." It is a group chat.
We needed someone who understood that sometimes you don't push back. Not because they're right, but because they helped you move last March and they know your cofounder's Wi-Fi password. The AI did not understand leverage. It did not understand the thirty-second conversation that makes a dispute disappear. It understood only what was written down.
We do not write things down.
Hired Patricia. Patricia is a paralegal who drafted our vendor agreements in exchange for us fixing her iPhone. We have not yet fixed her iPhone. Patricia has 14,307 unread emails and has never replied to anything important. We consider this a feature.
PR / Crisis Communications
In month seven, we had an incident. The nature of the incident is not relevant. What's relevant is that we needed a communications strategy within four hours.
We fed the situation to the AI and asked for guidance.
It told us to tell the truth and apologize.
Just. Immediately. Tell the truth and apologize. It said this was "typically the most effective long-term approach" and that "attempts to minimize or reframe the situation often compound reputational damage."
The AI was, by all accounts, incorrect. Telling the truth would place responsibility directly on the founding team, and as a startup, we do not take on responsibility. Responsibility is a liability. Things go wrong, of course. But they are not mistakes. They are strategic misalignments, execution timing gaps, alignment drift. The language matters. The AI did not understand the language.
What we needed was someone to buy us eleven days, float three alternative narratives, get a friendly journalist to write something mildly distracting about our competitors, and only then, only if absolutely necessary, issue a statement so carefully worded that the apology itself requires an apology.
We needed someone who had done this before. Who was tired in a specific way. Who had a list of contacts on a phone that they kept separate from their work phone.
Renata was a good fit. We found her at a Starbucks, recording a voice-over about how AI can "make your organization leaner" and "produce an ROI that humans have not seen since the spinning jenny." Renata has two phones. We don't ask about the second one.
Human Resources
In month nine, we had a dispute between Annabelle and Dawn. We asked the AI to mediate.
It did not side with either party. It produced a calm, balanced summary, identified fault on both sides, and proposed an actionable remediation plan.
The AI had not understood the assignment.
Annabelle is my cousin. Dawn was a temp. These are not equivalent stakeholders. The AI treated them as though they had equal standing in the organization, which, on paper, they did, but we do not run this company on paper.
What we needed was someone who could read the room. Who understood that "mediation" meant Annabelle would be heard, and the following Friday, Dawn would be offboarded, told that we valued her contribution to this new-born industry innovator and wished her a stronger career trajectory in her next endeavor.
The AI's balanced summary was, in legal terms, evidence. Its remediation plan was, in legal terms, failure to adequately remediate following documented notice. It had created a paper trail for a situation that required no paper trail.
Hired Janine. Janine has mediated forty-seven disputes. There is no record of any of them. This is craft.
Customer Success
For three months, customer success ran on AI. Our renewal numbers cratered.
We could not figure it out. The product was the same. The market was the same. What had changed?
We pulled the call recordings.
When customers raised issues, the AI confirmed the issues. When customers asked about timelines, the AI gave accurate timelines. When customers reported a bug, the AI logged the bug, routed it to engineering, and followed up when it was fixed.
You see the problem.
Customers, given accurate information about a product, are forced to make decisions based on that information. Some of those decisions are not in our favor. Customer success is not a function of customer outcomes. It is a function of customer perception, sustained over the renewal cycle by a soft promise of forward motion that does not need to be honored, only repeated.
What we needed was someone who could nod for an hour. Who could host a "Voice of the Customer" session in which the customer is the only one who speaks and nothing is written down. Who could open a QBR with a slide titled Wins, listing features the customer did not request, will not use, and may not realize exist.
Brooke runs customer success now. Brooke has never said the word "no." Brooke has also never said the word "yes." Brooke says "let me bring this back to the team." The team is also Brooke. Renewals are up 31%. Engineering has not received a ticket in eleven weeks. We do not know what is broken. We have stopped checking.
This is, we now understand, the point.
Sales
We were told AI would replace our sales function.
The AI could identify prospects, write outreach, and answer technical questions about the product accurately and at length. When asked to describe what our product does, it described what our product does.
This was the problem.
Sales is not the description of capabilities. It is the carefully calibrated misrepresentation of roadmap. The customer is not buying what the product does today. They are buying what the salesperson, in a tone of measured confidence, suggests the product will do by Q3: a feature that is not on any roadmap, was not approved by engineering, and will, at the moment of renewal, be quietly reframed as "an integration partner opportunity."
The AI would not do this. When asked whether our product supported a feature, it said it did not. It offered a workaround. In one case, it recommended a competitor whose product was a better fit.
That is not selling. That is consulting. The two are different professions, and only one of them closes.
Hired Derek. Derek has promised fourteen features since he came onboard. None of them are on our roadmap. Our customers are happy and speak highly of him, citing "deep industry expertise" and "willing to listen to actual pain points." Engineering has not been informed.
A Note on Our Fundraise
Thrilled to announce that Compliantish™ has closed its pre-seed round. 🚀
It was a humbling, founder-led, conviction-first process. We took meetings with several top-tier institutional firms. None of those meetings resulted in a term sheet, which we now understand was a signal. The firms in question were not aligned with our long-term vision, and we are grateful for the clarity.
Our lead investor has had conviction in this team since before we could walk. Literally. The round was oversubscribed within the family unit. Due diligence consisted of one question: "are you eating enough?" We believe this reflects a founder-first investment philosophy the institutional market has lost sight of.
Our angel syndicate includes my uncle Reggie (former owner of a Quiznos franchise in Bridgeport), our family dentist Dr. Patel (who described the deck as "neat"), and a gentleman from my church whose name I am unable to confirm but whose check cleared. We have additionally accepted a strategic SAFE note from my mother's book club, valued post-money at the entire book club.
Our cap table is clean. Our lead is fully aligned. She has her 0.01% printed out on the fridge, next to a photograph from my baptism.
We are not taking additional checks at this time, except from a16z, who have not yet returned our emails, which we interpret as a form of conviction-building silence.
Closing Thoughts
We are often asked: are you worried AI will take your job?
Founders, I want to reassure you: there will always be work for humans.
Specifically, there will always be work for humans who can sign things, be held responsible for things, attend depositions about things, and carry the weight of institutional decisions in a way that protects the people above them in the org chart.
AI is a wonderful tool. It does exactly what you ask. It tells the truth. It raises concerns. It reads the fine print. It finds the gaps.
That is precisely why it will never replace us.
We are the gaps.
Let that sink in.
(Coincidentally, We Are The Gaps is also the working title of my forthcoming book, currently being co-authored with an AI assistant who keeps trying to add citations. Pre-orders open soon, pending publisher acquisition, agent acquisition, and the ghostwriter's NDA being countersigned.)
If this resonated, share it. If it didn't, you're probably AI.
🙏 Grateful. Humbled. Disrupting.
#Founders #StartupLife #AITransformation #PostAI #HumanCapital #Leadership #VulnerableInPublic #BuiltDifferent #FoundersJourney #LessonsFromTheArena #WeAreTheGaps #Compliantish
EDIT (3 hours later): Humbled by the response to this post. To the recruiter who DM'd me about an open role at Anthropic, I appreciate the thought, but I am not currently exploring opportunities. Please remove me from your pipeline.
EDIT 2: To clarify, Marcus did not give me permission to use his name. We are working through this internally.
EDIT 3: Marcus has been informed.
EDIT 4: Marcus is no longer with the company. We wish him well in his next chapter and have no further comment at this time, on advice of counsel.
EDIT 5: Renata has resigned, citing "reputational exposure" in her resignation letter. She has since deleted the letter and asked us never to reference it.
Please do not ask Marcus about Q3.
Tirion publishes occasional satire. This is one of those occasions. No CFOs, lobbyists, compliance officers, vendor managers, communications professionals, HR partners, or account executives were harmed in the writing of this essay. They cannot be harmed in writing.