AI Didn't Write My Blog Posts. I Argued With It Until We Were Right.
A working method for AI-assisted analysis — and the moves that separate publishable thinking from plausible-sounding output.
There's a genre of post going around LinkedIn right now that goes: "I used AI to write a thing. Here's how. You can do it too." I read a lot of them. Most of them are bad.
Not because AI can't help with analytical writing. It can — I just published a two-part series on the services-as-software thesis that I genuinely could not have produced as quickly or as sharply without an AI sparring partner. But the genre is bad because it gets the causal arrow backwards. The lesson it teaches is "AI did the thinking and made the output good." That's wrong on the facts. The actual lesson is "AI produced structured first drafts that were directionally right and substantively wrong. The output was good because I pushed back until I was comfortable that the facts supported the analysis."
The stories are different: The first sells software (AI is great!). The second is a working method (AI can be great… when the user questions the results).
What follows is a working model. Specifically: how the Services-as-Software Files actually got built, where my AI collaborator (Claude, since we're being specific) was useful, where it was confidently wrong, and what I did at each fork in the road that the AI wouldn't have done on its own.
The question that started it
Someone sent me Jackie DiMonte's piece on the Own / Operate / Outsource framework — Grid Capital's articulation of why value in software markets is moving to the extremes and the middle is getting squeezed. The question on the table was whether the thesis was isolated to Grid or part of a broader VC shift.
The first draft answer Claude produced covered the framework cleanly, sourced the broader VC consensus (Sequoia, Foundation Capital, Wipro, a16z all converging on related theses), and gave me a clean read.
It also contained a mistake I didn't catch on the first pass. The answer credited DiMonte as the originator of the thesis. She wasn't.
That mistake didn't get fixed until much later in the conversation, when I asked the right follow-up: who was the first VC to start writing about this? The honest answer turned out to be Sarah Tavel at Benchmark in early 2024, followed by Foundation Capital's named essay in April 2024. DiMonte's piece came a year later. Her contribution was the analytical framework, not the underlying thesis.
I would have shipped Part 1 of the series with the wrong attribution if I hadn't asked the provenance question. That's the first lesson of this method: AI doesn't know what it doesn't know, and it confidently presents partial attributions as complete ones. The corrective isn't smarter prompting. The corrective is asking the question the AI didn't think to ask itself.
The first real pushback
A few exchanges in, the draft contained something like "Harvey is shifting from selling software seats to selling legal work through revenue-share deals — building specialized AI agents for narrow legal workflows, then collapsing them into one interface. That's the explicit operate-to-outsource pivot."
I read that and called it. That's not a pivot. That's multi-agent orchestration. Every AI developer is doing that no matter what side of the Own / Operate / Outsource ditch they're on. Greenlite produces the actual construction permit. Harvey's orchestrator sends you to an agent and lets the lawyer figure it out from there. The marketing isn't novel either — every legal tech startup I know is doing this.
The framing the model had used was VC narrative copy it had absorbed and repeated. Compound vertical AI platform and outcomes-based pricing and agents that run workflows end to end are the marketing layer of the operate-tier companies, not evidence of a structural pivot. The orchestrator-plus-agents architecture is now the baseline for every AI product in legal, healthcare, and customer service. Calling it a category pivot is exactly the kind of marketing-layer thinking the post was supposed to dismantle.
That was the moment the analysis tightened. The strict version of the deliverable test — who produces the deliverable the paying customer actually uses — emerged from that pushback. Not from anything in the prior drafts. And once the strict version was on the table, Harvey didn't pass it. Neither did EvenUp. Neither did most of the companies the market was pricing at outsource-tier multiples.
This is the second lesson: AI will absorb and repeat the dominant industry narrative unless something forces it not to. The dominant narrative in venture capital right now is that services-as-software is a real category and that companies marketing themselves into it are mostly legitimate. That narrative is wrong, but the training data is full of it, the press releases are full of it, and the model's first draft will track the narrative unless somebody actively pushes against it. This happens even when AI is analyzing individual documents. If the Executive Summary says: “the company did X,” AI says “the company did X” even if the data doesn’t support that conclusion.
The EvenUp moment
A few exchanges later, with the strict test now established, the model produced a comparative list of companies in each category. EvenUp landed in the outsource column.
I went after it: what is EvenUp producing? It looks like a bunch of agents to me.
EvenUp markets aggressively as services-as-software — "Pre-Litigation as a Service," "$10 billion in settlements" — and the model has let the marketing language pull the categorization toward outsource even though the strict test puts EvenUp clearly in operate-tier. The personal injury attorney still signs the demand letter. The attorney carries the malpractice exposure. The insurance carrier negotiates with the law firm, not with EvenUp. EvenUp makes the lawyer faster at producing the work product. EvenUp does not produce the work product.
The categorization got corrected. More importantly, the model had to articulate why it had made the mistake in the first place — and that articulation turned out to be the most useful thing in the entire series:
“Regulated professional services — law, medicine, accounting, much of finance — are structurally constrained from producing true outsource companies unless the AI company itself is the licensed provider. Bar rules, UPL restrictions, medical licensure, and CPA requirements all force a human professional into the producer-of-record role. The AI company can compress the cost of producing the work product but can’t legally be the producer.”
That paragraph, which became the structural backbone of the entire series, did not exist in any draft before the EvenUp correction forced it. The AI didn't think to articulate it. I didn't think to ask for it directly. It emerged from the friction of refusing to accept a sloppy categorization.
Lesson three: the corrections are where the real analysis lives. First drafts get you the obvious moves. Corrections force the AI to articulate principles it would otherwise leave implicit.
Pressure-Test Your Stack Before They Do
The same structural pressure-testing this method produced for the Services-as-Software Files is the work I do every week with founders inside the Investor Readiness Vault™ — running your business through the diagnostic before a diligence partner does it for you. If you read this far, you're already doing the harder half of the work. The Vault is the rest.
The framing that wasn't the model's
By the time the analysis got to which companies were most exposed, I articulated something the AI had been dancing around without naming:
“What we’re finding is that companies are marketing to the outsource category even if they’re operators. That’s hype. VCs are pushing a thesis that doesn’t actually exist in real life.”
That was the framing the entire series was built around. Not "the services-as-software thesis is real but stretched." Not "some companies are operate-tier in disguise." But: VCs are funding a thesis whose real-world footprint is much smaller than the funding implies, and the marketing layer is doing the work of the missing fundamentals.
The blog posts crystallized around that frame. The model had been close to it in pieces — the analysis of margin economics, the discussion of the arbitrage VCs were running, the deliverable test — but the unified articulation hadn't surfaced until I named it.
Lesson four: the AI is good at producing the parts. You're the one who produces the unifying frame. Don't expect the model to give you the headline thesis. Expect it to give you the components, and trust your own judgment on how those components compose into something worth saying.
The working method, generalized
Five things I'd take from this if I were teaching it:
Use AI as a sparring partner, not a writer. The output that ends up in print is not what the AI produced. It's what survived your pushback. If you don't push, you publish marketing copy with your name on it.
Refuse to accept sloppy categorization. This is the single most useful move. When the AI puts something in a bucket, ask whether the bucket holds up under a strict test. If the test is loose enough that everything passes, the test isn't doing any work and the bucket isn't either.
Ask provenance questions. Who said this first? Where did this number come from? What's the actual citation? The model is confident about origins it doesn't know. The questions that catch attribution mistakes are the ones you'd otherwise be embarrassed by.
Watch for absorbed narrative. If the AI's answer tracks the dominant industry narrative on a contested topic, it probably is the dominant industry narrative — which means it's the answer you'd get from any reasonably-informed observer, not the answer that's worth publishing. Push until the answer departs from consensus or until you've confirmed consensus is actually right.
Bring the unifying frame yourself. The AI assembles components. You see how they fit together into something with a thesis. If you outsource the frame, you publish a competent summary. If you bring the frame yourself, you publish something that says something.
What this isn't
This isn't a post about how AI is going to replace senior analysts or lawyers or strategists. It's a post about what it actually takes to use AI well for analytical work — which is, in every dimension that matters, more domain expertise and more disciplined pushback than the genre admits.
The Services-as-Software Files turned out the way they did because I argued with the model for three days. The thinking is mine. The drafting was collaborative. The corrections were the work.
If you came here hoping for a how-to-prompt guide, that's not what this is. The lesson isn't a prompt. The lesson is: bring something to the conversation worth defending, and then defend it.
Helping founders think through their category positioning is the work I do every week inside the Investor Readiness Vault™. The same kind of structural pressure-testing this method produced for the Services-as-Software Files, applied to your specific business — before a diligence partner does it for you. Book a 20-minute call →
