ChatGPT and Claude give solo founders 10x execution leverage on code and copy, but fail at the two things a co-founder actually provides: challenging your omitted blind spots and holding you to last month's promise.
For a solo founder in 2026, an AI assistant is the fastest teammate you will ever hire. It writes boilerplate code in seconds, critiques landing page copy at 2 a.m., and turns messy customer call transcripts into clean feature specs. Because it handles execution tasks so well, it is natural to start using the same chat window as a strategic co-founder: asking whether you should pivot, change your pricing tier, or build another feature before launching. That is where solo founders walk into a documented trap.
In a landmark field experiment with 758 strategy consultants, researchers from Harvard Business School, Wharton, and MIT (Dell'Acqua et al.) mapped what they called the "Jagged Technological Frontier." For execution tasks inside the frontier, professionals using AI finished 25.1% faster and completed 12.2% more tasks at higher quality. However, on complex judgment tasks just outside the frontier (where noisy qualitative interviews had to be weighed against misleading metrics), subjects using AI were 19 percentage points less likely to reach the right decision than those working without AI. The reason: AI generates articulate, plausible-sounding rationalizations that lull solo decision-makers into switching off their own critical judgment.
Instead of asking an LLM to do a job it is structurally bad at, split your solo founder stack across the Jagged Frontier:
| Founder Job | AI Assistant (ChatGPT / Claude) | Human Peer Circle (en)Apply → |
|---|---|---|
| Code, Copy & Spec Execution | Excellent (25%+ speed boost on scoped drafts & refactors) | Not for daily hands-on labor |
| Summarizing User Interviews | Strong at clustering transcripts you paste in | Spots when you asked leading questions or interviewed the wrong buyer |
| Challenging Your Framing | Weak: RLHF sycophancy mirrors your assumptions back to you | High: two stage-matched founders push back on rationalizations |
| Catching Omitted Context | Blind to what you leave out of the prompt | Remembers your last 3 months of commitments and patterns |
| Deadline Accountability | Zero social cost to closing the browser tab or starting a new chat | High: reporting live to two peers who expect your 30-day update |
en matches you with two other founders at your exact stage for a structured 60-minute session every month. You review whether you delivered on last month's goals (accountability) and pressure-test whether you are solving the right bottleneck next before spending another month on the wrong thing (calibration).
When you are tired of distribution and want to retreat into your IDE to build a new integration, you do not ask a neutral question. You type a prompt framed around why the new integration is strategic. Foundational research by Anthropic (Sharma et al., 2024) showed that models trained with Reinforcement Learning from Human Feedback (RLHF) systematically shift their answers to match the user's stated beliefs. In May 2025, Stanford researchers (Cheng, Jurafsky et al., "ELEPHANT: Measuring and understanding social sycophancy in LLMs") went further: on open-ended advice queries, major LLMs preserved the user's desired self-image and accepted the user's framing at rates 45 percentage points higher than human peers. OpenAI ran into this exact failure mode in April 2025 when it had to roll back a major GPT-4o update ("Sycophancy in GPT-4o") because optimizing for short-term user thumbs-up signals caused ChatGPT to enthusiastically validate impulsive and flawed decisions. If you pitch a bad pivot eloquently, an LLM is incentivized to hand you a polished 6-week roadmap for that bad pivot.
An AI assistant only knows what you type into the prompt. If you ask how to improve your landing page conversion rate, it will give you ten copywriting tweaks. It cannot see that you only had 40 visitors last month, that you have been avoiding direct outreach for six weeks, or that you promised yourself in September you would not touch the landing page again until you spoke to 15 churned users. Self-deception happens before your fingers hit the keyboard, in the facts you choose to omit.
A 2016 meta-analysis of 138 studies in Psychological Bulletin (Harkin et al.) showed that tracking goals produces its strongest effect when progress is recorded and reported publicly to other people. You cannot feel embarrassed in front of a stateless token predictor. When you miss a self-imposed launch date, you simply open a fresh chat thread and never mention the missed commitment again.
The most effective solo founders in 2026 use AI for 30 days of execution leverage and human peers for 60 minutes of monthly calibration. en is a peer accountability group for solo founders: a curated circle of three stage-matched founders who meet online for 60 minutes each month for structured check-ins and honest feedback. In your focus round, two founders at your exact stage provide the three things an LLM never can: memory of what you promised last month, refusal to flatter a comfort-zone detour, and real skin-in-the-game social presence.
Can prompting an AI to "be a brutal critic" fix the sycophancy problem?
Only partly. Asking an LLM to "roast my idea" makes its tone harsher, but it still operates strictly inside the facts you chose to include in the prompt. It cannot notice that you are asking a product question to avoid a sales bottleneck you promised to fix last month.
What is the best way for a solo founder to use AI alongside a peer group?
Use AI before and after your monthly circle, not instead of it. Before your session, use Claude or ChatGPT to summarize your metrics and draft a 5-line decision brief. Bring that brief to your 3-founder circle on en for a focus round of human calibration, state your next commitment out loud, then use AI all month to execute it faster.
What did the Harvard Business School study find about AI and strategic decisions?
In "Navigating the Jagged Technological Frontier" (Dell'Acqua et al.), consultants using GPT-4 completed execution tasks 25.1% faster, but on a complex judgment task outside the frontier where quantitative data conflicted with qualitative nuances, AI users were 19 percentage points less likely to find the true root cause than people working without AI.
How is an en circle different from chatting with founders in a Discord server?
Public Discord and Slack channels suffer from the same continuity gap as a new chat thread: strangers give drive-by opinions without knowing your business model or last month's goals. en matches you with the same two stage-matched peers so context compounds month after month.
Use AI to build faster all month. Join a 3-founder circle on en for 60 minutes a month to make sure you are building the right thing.
Questions? Email us at hello@en.social