From initial decision to board-ready pitch. What AI handled, where it failed, and what I had to do that it couldn't.
Click any section to expand. Sections 05–11 contain the detailed session logs.
Yes, it compressed what would have taken a 6-person team two weeks (~480 hours) into 48 active hours. But it also compressed roughly 160 hours worth of the highest-level cognitive work — investigative reasoning, auditing financial models, catching what's wrong in something that looks correct — into those same 48 hours. The leverage is real. So is the fatigue.
Polished output is the trap. The skill that matters is critiquing, challenging, and finding holes — including reasoning about human behaviour. Nobody pays a $750 rush fee unless the standard offering is throttled. AI cannot model that incentive structure.
I suspect the productivity gap between people who combine these tools with the cognitive capabilities in points 01 and 02 above and everyone else will widen very quickly. Not the disappearance of the middle class — more like a 0.1% who compound their output exponentially, and everyone else competing for the same diminishing share. The separator is not technical. It is the ability to hold many data points in tension, derive the story, spot the lie in the clean model, and know which question to ask next (at least today).
A third of my instructions were teaching the AI how to think. Those drove 70% of the quality improvement.
Treating interest income as additive to BD/RIA fees. Revenue inflated by 61%. One correction forced a full revenue model rebuild.
The AI could calculate but not question. Every investigative question that uncovered a hidden clause or strategic insight came from me.
The AI would build a model then use the model's output to validate its own inputs. These errors compound silently.
The AI's default is to present everything it knows in the order it learned it. Editorial judgment is not computation.
The grandfathering clause was in the Form ADV. The January churn spike was in ChartMogul. The AI analyzed each document in isolation.
Does 205,200 + 12,700 = 218,800? No, it's 217,900. Small errors left unchecked erode credibility.
The AI can multiply two numbers. It cannot figure out which two numbers to multiply, or that the answer requires pulling data from three different documents.
It treated everything as ground truth. Ambiguity in source material is itself a finding worth flagging. The AI never flagged it.
It can tell you December is the biggest month. It cannot ask why, whether there are structural reasons beyond the tax deadline, or whether the pattern is accelerating.
It treated the $410K/yr advisory fee as current revenue when 76% of accounts are grandfathered at 0%.
When ChartMogul said 4,248 subscribers and the State of the Union said 4,052, the AI used whichever number it encountered last.
The AI's instinct is to show its work. A board memo needs the answer. A pitch needs the reason to act. The AI treats all three the same.
It presented a derived estimate with the same confidence as a confirmed data point.
Without explicit direction, every section would have centered text, purple gradients, and the same rounded card component repeated 15 times.
When asked for a competitive landscape, it mapped Solo OS providers but missed Carry's own competitors and partners like custodians and distribution channels.
It treated every document and spreadsheet as authoritative. It never flagged when column headings were misleading or a founder's verbal answer contradicted his own data.
It can research 14 competitors in an hour, build an interactive valuation calculator, and match a company's design system from screenshots. But it cannot tell you what matters, why the numbers don't add up, or how to frame a story for a board that needs to make a decision.
The work that mattered most was the work done between the AI's outputs: checking the math, asking what the numbers mean, reframing the narrative, pointing at data it already had.
Not one-time corrections. Standing instructions applied to every future output. These were the single biggest time saver across the project.
| Rule | Instruction | Why it mattered |
|---|---|---|
| No hyphens | "Remove hyphens from all copy." | Personal style preference. Applied globally. |
| Lead with insights | "Always lead with insights, then back them up with data." | Flipped the AI's default (data first, conclusion last). |
| Be concise | "If the calculation is in the table, don't repeat it in the text." | Cut word count by ~40%. |
| Kill list | Banned: delve, landscape, ecosystem, leverage, utilize, robust, seamless, and ~50 others. | Every draft scrubbed against this list. AI still slipped back. |
| No suspense framing | "Get to the point. Drop the suspense." | Stopped openings like "The answer surprised me." |
| One adjective | "Not two. If something is thorough, it's already comprehensive." | Reduced redundant description. |
| Vary sentence length | "Short for impact. Long for exploration." | Broke the AI's uniform 12-word sentence rhythm. |
| No false balance | "Give the critical insight three paragraphs. Give the minor caveat one sentence." | Stopped equal weighting regardless of importance. |
| Rule | Instruction |
|---|---|
| Study lettuce.co | "Match their design system exactly." |
| Think like the head designer | "Make a decision consistent with the brand." |
| No generic patterns | "No centered layouts, purple gradients, or Inter font." |
| Typography over containers | "Big numbers should be typography, not inside boxes." |
| Contrast rule | "Dark text on light. Light text on dark. Always." |
| Rule | Instruction |
|---|---|
| Lead with tagline | "A concise summary line that encapsulates the key takeaway." |
| One-pager over slides | "Prioritize simple, clean, sharp one-pager format." |
| Collapse the math | "Lead with conclusions. Keep calculations hidden." |
Questions the AI would never have asked. Each one turned a surface-level data point into a strategic insight.
What's the missing value between members with funded accounts vs. not? What does that mean for customer inheritance for us?
3,995 unfunded accounts (54.6%) are empty shells. Real customer base is 2,951, not 4,052.
Every downstream number recalculated on the correct base.
What are some reasons there would be a Dec/Jan cliff outside of the tax deadline?
December surge amplified by Black Friday promos. January cliff was promotional hangover, not a fee reaction.
Reframed from advisory fee churn crisis to structural post-promotional hangover.
Is there a grandfather clause or material fee change in the last 3 years?
Form ADV Part 2A: accounts opened before June 20, 2025 pay 0% advisory fee until January 1, 2027. 76% are grandfathered.
Turned the advisory fee from current revenue into a future cliff event. Created the single biggest negotiation variable.
Research the exact window where Carry is the only option in December.
Carry does same-day through Dec 30. Competitors cut off Dec 15–24. 10-day monopoly window.
Turned "December is a big month" into "Carry has a 10-day monopoly window (Dec 22–31)."
Who are Carry's own competitors and current partners?
Mapped 14 potential acquirers. Collective became the primary threat. Existing Lettuce partnership at 90% discount surfaced.
"If you don't buy, someone else will" became the urgency argument.
What does 'self-directed' actually mean in the AUM columns?
$205M managed (94.2%) vs $12.7M self-directed (5.8%). Ankur said 'most are self-directed' but by dollars, 94.2% is managed.
Confirmed advisory fee applies to $205M. Revealed Ankur may not fully understand his own AUM composition.
The ARPU went from $79 to $24 for new customers. What's driving that?
New customer ARPU collapsed 70% in 10 months. Higher value customers leaving, lower value joining.
Revealed the subscription business is in a value erosion spiral.
Is there a retention cliff at month 12?
Month 12 churn halved but months 2–6 got worse. Despite 2x more customers in 2024, 12-month survival was lower.
Shifted from annual renewal problem to leaky middle problem.
What does the DriveWealth contract actually say about transferring the relationship?
Section 37.2: requires written consent for assignment. Without it, the assignment is void. ACAT transfer cost of $65/account.
Made the DW contract a binary deal-breaker question.
What's the actual advisory fee revenue collected to date? Not the potential, the actual.
Only ~982 accounts currently pay 0.20%. Actual current revenue: ~$15K–$29K/yr. Not the $410K/yr potential.
Separated current advisory fee revenue (tiny) from post-2027 potential (transformative).
Not corrections to individual numbers. Each one installed a way of thinking that changed every downstream calculation.
| # | What I told it | What it was doing wrong | What changed downstream | AI credits |
|---|---|---|---|---|
| 01 | "The $2.7M is ($70K + $155K) × 12. Interest can't be on top of that. Ankur wouldn't undersell his company." | Adding $743K/yr interest ON TOP of $840K BD/RIA. Inflating revenue by 61%. | Complete revenue model reconstruction. Revealed the business is fundamentally an interest income business. | ~$18 |
| 02 | "BD/RIA fees compound with AUM growth. No new sales effort required." | Treating BD/RIA as a static $840K/yr line item. | Rebuilt the entire flywheel. $840K/yr became the floor, not the number. | ~$10 |
| 03 | "Cross-sell is opt-in. The 401k is sticky. Even at 5% conversion, that's $530K/yr at near-zero CAC." | Modeling cross-sell as a retention event. | $2.3M/yr base revenue is the floor regardless of cross-sell. Cross-sell is pure upside. | ~$8 |
| 04 | "Don't model the 1,101 content-only subs as part of the asset. They're gone." | Using 4,052 as the customer base for everything. | Recalculated everything on 2,951. Narrative shifted to 2,951 customers with real accounts and $218M in assets. | ~$8 |
| 05 | "Average funded account is ~$50K. That money sits on the platform for decades." | Applying SaaS churn logic to retirement accounts. Projecting 5–10% annual AUM churn. | Stickiness of retirement capital became a central pillar. Switching costs are structural, not psychological. | ~$3 |
| 06 | "Carry is the only provider after Dec 22. That's a monopoly." | Treating the December spike as a seasonal pattern. | The December Monopoly became one of three pillars of What We Are Buying. | ~$3 |
| 07 | "The advisory fee is a cliff, not a tailwind." | Framing Jan 2027 fee rollout as unambiguously positive. | Reframed as dual-edged: upside ($287K–$410K/yr) and downside (churn crisis). Became the single biggest negotiation variable. | ~$5 |
| 08 | "2nd year contribution is higher than 1st. Factor that into AUM growth." | Modeling AUM growth as new customer acquisition only. | Same 355 December customers generate more revenue each year without additional spend. | ~$5 |
Small errors left unchecked erode credibility in a board presentation.
"The $2.7M is ($70K + $155K) × 12. Interest can't be on top of that."
Full revenue model rebuild. Every downstream number was wrong.
"Cross-sell should be on 2,951 active members, not 4,052 total subs."
Recalculated from $1.45M to $1.06M. Cascaded through 6+ downstream numbers.
"205,200 + 12,700 = 217,900. You're showing 218,800."
Full AUM reconciliation across 3 data sources.
"The January MRR decomposition doesn't add up. +$4,942 − $16,256 = −$11,314, not −$7,314."
Rebuilt with exact CSV data. Found missing reactivation MRR.
"The churn table uses $24/mo ARPU but the text says $37.85 blended."
Audited every hardcoded number. Found stale Active ARPU ($41.54 vs $43.01).
"$500K–$900K doesn't match 6–8x on $211K revenue. That's $1.27M–$1.69M."
Recalculated enterprise value callout card.
"The grandfathered % uses 4,248 from ChartMogul but the valuation uses 4,052 from the State of the Union."
Fixed inconsistent denominators across the model.
"43% of total revenue doesn't match. $1.15M / $2.23M = 51.6%."
Corrected the percentage and all references.
The AI's default when it hits an unknown is to flag it as a question for the founder. Half the time, the answer was already in the data.
| # | Stuck on | What I told it | The derivation | Produced | AI credits |
|---|---|---|---|---|---|
| 01 | Content-only subscriber % — flagged as open question for Ankur | "You have the State of the Union. Do the subtraction." | 4,052 total subs − 2,951 active Solo 401k = 1,101 content-only (27.2%). | Separated the real customer base from the noise. | ~$3 |
| 02 | AUM growth rate per customer — modeled as a flat 7% market return | "Go back to the raw AUM data by month. Calculate the actual average contribution per account per year." | Pulled monthly AUM CSV, isolated new deposits from market appreciation. Found 2nd year contributions higher than 1st. | Replaced the generic 7% assumption with real contribution curves. | ~$8 |
| 03 | Managed vs self-directed split — action item to ask Ankur | "The AUM PDF columns already tell you. Stocks+ETFs is the DriveWealth managed column." | $205.2M managed (94.2%) vs $12.7M self-directed (5.8%). | Confirmed advisory fee applies to $205M. Caught that Ankur's claim 'most are self-directed' was wrong by dollars. | ~$3 |
| 04 | Advisory fee revenue — question for founder | "You have the fee schedule (0.20%) and the AUM ($205M). Calculate current vs potential separately." | Current: ~$15K–$29K/yr. Potential post-Jan 2027: $205.2M × 0.20% = $410K/yr. | Split advisory fee into two completely different numbers. Current revenue is negligible. Future revenue is the single biggest variable. | ~$5 |
| 05 | December Rush lock-in count — flagged as unverified estimate | "Look at the December sign-up data across all three years. Isolate the Dec 22–31 window specifically." | Dec 2023: 291, Dec 2024: 467, Dec 2025: 702. Derived 355 rush-window signups. | Turned "December is big" into "355 annual lock-ins during a 10-day monopoly window." | ~$5 |
| 06 | January churn spike — treated as advisory fee reaction | "76% are grandfathered at 0%. The fee can't explain January churn. Cross-reference with the promotional calendar." | Mapped December Black Friday promo to January expiration. Churn was promotional hangover, not fee-driven. | Completely reframed the risk narrative. Advisory fee is not causing churn. Promotional dependency is. | ~$5 |
8 major research efforts. Fast and thorough on execution. Consistently missed entire categories of analysis until I pointed them out.
Build a full competitive landscape for the Solo 401k market.
Mapped Fidelity, Schwab, Vanguard, MySolo401k, Ubiquity, ForUsAll. Pricing, features, AUM, setup time.
Mapped Solo OS providers but missed Carry's own competitors and partners like custodians and distribution channels. Collective, Bench, and the Lettuce partnership never surfaced.
Research who would buy Carry if Lettuce doesn't.
Listed Betterment, Human Interest, Guideline as potential acquirers.
Missed the defensive framing. The reframe: 'If Collective buys Carry, they complete their Solo OS and Lettuce loses the retirement piece race.'
Research the December window — exact dates when each competitor stops accepting applications.
Called each provider's processing times, mapped cutoff dates. Built the timeline table.
Didn't connect it to pricing strategy. If Carry is the only option Dec 22–31, can they charge more? That led to the annual-lock-in-only recommendation.
Research SaaS valuation multiples in the current market.
Found the 2026 SaaS crash data, $300B evaporated, multiples compressed from 15–30x to 5–8x.
Didn't separate SaaS valuation from AUM valuation. Applied SaaS multiples to the entire business. The AUM-based revenue is not SaaS. It compounds.
Read the DriveWealth contract and find anything that affects the deal.
Found Section 37.2 (assignment requires consent), ACAT transfer costs ($65/account), early termination penalties.
Didn't flag the interest rate mechanics. The Bank Sweep Deposit Program surfaced only when asked — revealing BD/RIA is 88% interest income.
Research RIA compliance costs for a firm Carry's size.
Compiled a detailed cost breakdown: CCO ($100K–$200K), E&O insurance ($5K–$15K), SEC exam ($70K), total $150K–$400K/yr.
Didn't connect compliance costs to the acquisition structure. Does Lettuce already have an RIA? Can they absorb this under existing compliance? Changed the cost from $400K/yr standalone to potentially $50K/yr incremental.
Analyze the cohort data for retention patterns.
Built survival curves by year, calculated month-by-month churn rates, found the month 12 cliff.
Didn't ask why 2024 cohorts had worse mid-month churn despite better month-0 and month-12 numbers. The leaky middle (months 2–6) is the real problem, not the renewal cliff.
Research the Form ADV and SEC filings for anything material.
Found the grandfathering clause, the fee schedule, the fiduciary duty language.
Didn't cross-reference the grandfathering timeline with the churn data. Grandfathering expires Jan 2027, so any churn before that date can't be fee-driven. That eliminated the biggest false alarm.
| Instruction | What changed | AI credits |
|---|---|---|
| "This reads like a risk report, not a pitch. Lead with why we should buy." | Board Memo restructured. Opened with strategic rationale instead of methodology. | ~$8 |
| "Organize by strategic question, not by data type." | Full page restructure. Merged two overlapping pages into one organized around 5 strategic buckets. | ~$15 |
| "Lead with insights, then data. Not the other way around." | Flipped every section. Insight first, supporting data collapsed underneath. | ~$8 |
| "If the calculation is in the table, don't repeat it in the text." | Cut redundant prose across 15+ sections. | ~$4 |
| "Split: What We Are Buying vs Why We Should Buy." | Created two distinct sections with different purposes. | ~$4 |
| "Be concise. You're writing a board memo, not a research paper." | Cut word count by roughly 40% across the memo. | ~$4 |
| Instruction | What changed | AI credits |
|---|---|---|
| "Match the lettuce.co design system exactly. Study their site." | Audited 6+ pages of lettuce.co. Extracted fonts, colors, spacing, animation patterns. | ~$8 |
| "Too many starbursts. Get rid of them." | Replaced starburst bullet icons with clean card boxes. | ~$4 |
| "Think of yourself as the head designer at lettuce.co." | Shifted from implement what the user said to make a design decision consistent with the brand. | ~$3 |
| "Big numbers should be typography, not inside boxes." | Shifted from data-table-heavy sections to typographic callouts for key numbers. | ~$4 |
| "Scroll through the full memo on mobile to verify card grids stack cleanly." | Mobile audit. All grids confirmed to collapse to single column below 640px. | ~$4 |
Active typing and clicking only. Does not count time spent reviewing, thinking, or waiting for output.
| Category | % | Est. Cost | What It Means |
|---|---|---|---|
| Investigative questions that put meaning behind numbers | 16% | ~$45 | "What's the missing value between members with funded accounts vs. not?" "Why is there a Dec/Jan cliff?" "Is there a grandfather clause?" |
| Teaching the AI how the business works | 15% | ~$55 | Compounding, stickiness, cross-sell mechanics, what's additive vs dependent |
| Catching math and logic errors | 14% | ~$50 | Revenue double counting, wrong denominators, numbers that don't reconcile |
| Deep research and build out requests | 12% | ~$65 | Competitive landscape, valuation model, contract analysis, board memo |
| Teaching the AI how to tell the story | 10% | ~$25 | Section sequencing, insight-first structure, when to use visuals vs words |
| Pointing out it already had the answer | 10% | ~$20 | "You have the data. Do the subtraction." "Go back to the raw AUM data." |
| Teaching the AI how to design | 8% | ~$25 | Studying lettuce.co, killing generic patterns, thinking like the head designer |
| Copy and text edits | 8% | ~$10 | Direct word swaps, title renames, tone corrections |
| Directing deep research and catching blind spots | 7% | ~$40 | "You missed Carry's own competitors." "Research the exact December window." |