The 90/10 Problem
Why the firms getting real value from AI aren't the ones waiting for it to be perfect
Every other private equity fund partner and operating executive we meet: “we’ll wait until AI is ready”, “we started with co-pilot but Claude’s pretty good”, “we can’t trust it”, “we don’t even know where to start”, “we can just vibe code it with Claude”.
At first glance, this makes total sense: there’s so much hype and the models improve weekly, but AI still makes mistakes. One ill-timed hallucination could put human livelihoods and billions of capital at risk.
From two years of building AI workflows inside live private equity deal teams and portfolio company operations, here’s what we’ve seen: firms are not getting real returns from AI because they discovered a more powerful and accurate AI model. Firms driving immediate value have figured out which parts of their process require 100% accuracy vs which parts benefit from accelerated throughput, where 90%+ accuracy is plenty. These are different problems requiring different solutions.
The distinction is critical because many in high finance treat accuracy as binary, and that framing is costing firms hundreds of wasted monthly labor hours (6 to 7 figure profit impacts). The biggest risk isn’t that AI will hallucinate the wrong answers, but that you try it and give up because of a poor experience. Everything we’re seeing on the ground points to AI driving a transformational shift in how deal teams work and how business gets done. The firms that are winning are treating it like a process to iterate on rather than a product to demo. They’re testing, learning the failure modes, trying newer models as they come out, and refining their own firm-specific workflows week over week. The firms that tried a few tools, decided to wait, or signed a SaaS / Claude Enterprise contract and called it a day may already be falling behind. The gap is widening between firms experimenting with AI and firms redesigning real workflows around it.
Real PE Workflow, Real ROI
Here’s a concrete private equity use case generating real ROI for funds right now. The reality in the lower middle market: small origination teams, hundreds or thousands of inbound opportunities a year, CIMs constantly flying across your desk, each one a 70-page PDF that somebody has to read, assess, and either pass on or escalate. The math doesn’t work; it’s like a factory operating at 30% utilization. With the largest funds building AI SWAT teams while moving down market for deals, competition will intensify, which means you’ll have to review more deals to keep the same pace of capital deployment. You cannot thoughtfully analyze that kind of deal volume with already maxed out headcount, so what actually happens is most deals get skimmed, pattern-matched against gut feel, and passed or pursued based on incomplete reads, because no one wants to share carry with yet another partner.
The firms doing AI well are building systems that read a CIM and produce a first-pass review structured around how that specific fund thinks about investing, in the format their IC loves. What an associate would have built by hand in three hours gets produced in five minutes. With the right guardrails and context, that’s a 90%+ reduction in time-to-first-pass.
The real value unlock happens once you’ve structured data across hundreds or thousands of prior deals. You start seeing your own pipeline differently. You can pull comps against deals you’ve passed on, at a depth and speed you never could before. The institutional memory that used to live in a partner’s head, buried in emails or on the desktop of a former associate who left for a search fund, now becomes intellectual capital the deal team can leverage at scale. AI handles throughput; investment professionals handle taste.
The inverse also holds true. Instead of summarizing what’s in the CIM, you can have AI generate questions and concerns the deal team should be pressure testing before a Monday IC meeting. Some concerns are obvious, most won’t land and will be immediately dismissed. But there could be one unique angle the deal team hadn’t yet thought of, leading to a strong deal that would have never been raised, or killing a deal before burning six figures on diligence.
The question isn’t “is the AI 100% accurate?” It’s “does this AI help us see more high-quality deals, faster, increasing our odds of success”. The former doesn’t matter for this use case, the latter is a strong yes. AI’s job was never to be the decision maker, rather to free up the decision-maker from wasting hours of time sifting through thousands of pages of documents for deals they’ll never invest in.
The accuracy spectrum
Most conversations about AI in finance treat accuracy like a light switch. In practice, every PE workflow sits somewhere on a spectrum. For example:
Thematic alignment: 70%-80% is plenty. Early deal screening, investment theme pressure testing, market mapping. At this stage you want to over index to volume and speed over accuracy. AI surfacing 20 interesting angles, five of which are garbage and 15 of which spark real thinking, is more valuable than a human analyst who carefully vets three, every day of the week. The garbage is easy to spot, and the 15th idea might be the one that changes your fund’s trajectory.
I’ve heard several deal teams describe their approach as essentially “let someone flag the obvious errors, billions vs millions, and move on.” They’re not using AI for LBO models, they’re using it to supercharge workflows around orientation.
Structured extraction: 90% plus human verification. CIM summarization into firm-specific formats, document classification across a data room with thousands of files, comping vs historical deals and benchmarking vs industry gross margins, revenue quality validation by triangulating across multiple data sets.
This is precisely where most firms get stuck. They see AI misclassify a lender doc as an MD&A or pull 2024 revenue figure in place of 2025 revenue, and conclude the tool isn’t ready yet, but the 90%+ that’s correct just saved someone ~three hours of copy-paste work, and the 5% to 10% “last mile” that needs fixing takes fifteen minutes of human review. The net time saved is still enormous, as long as you design the verification step to be fast and systematic.
We do not want to understate the mission critical nature of the last mile step. Additionally, if it takes longer to verify the AI output than to do the work yourself, a common occurrence for AI beginners, you’ve generated a negative ROI. The key is building a systematic audit trail into the workflows so the reviewer can trace every extracted data point back to its source, see exactly where the AI pulled a number from, and confirm or correct it in context. When you know the failure modes, knowing that AI struggles with unit conversions, or that it sometimes confuses EBITDA adjustments, you can design the review experience around those specific weak spots, rather than asking someone to re-check everything from scratch.
We’ve built document processing workflows that compress hours of work into minutes. Initial accuracy is rarely perfect, sometimes starting around 75%, but it improves rapidly towards 90%+ over time, with the speed of execution the main value driver from day one. Don’t forget, the human reviewer was always going to look at the output either way, there’s too much on the line not to double and triple check your models and decks. AI just changes what you’re reviewing, a mostly-complete first pass, with citations pointing back to the source material, so you can review more and better deals, faster
This same logic applies to benchmarking, but the framing is a little different. Think of it less as AI doing analysis and more as AI acting as a memory aid. When a deal comes in, the system can surface comparable companies and similar deals the fund has seen before, and flag if margins in this industry vary widely vs what’s in the CIM. The AI can help determine if the analysis is worth doing, surface comparable deals, and start you off much further along than if you began from scratch. It’s not telling you how to interpret nuanced shifts in pricing dynamics or cost structure. From there, the deal team does what they’ve always done: they dig into the numbers and company files, talk to their network, formulate their own views on the opportunity. The value is in deploying AI as a safety net to ensure nothing obvious gets missed when your deal team is underwater, not some AI-generated summary document.
Financial precision: 99%+ or don’t bother. The LBO model, sizing the debt package, the returns waterfall, sources and uses, anything that drives the bid or ends up in front of LPs.
This is where the “AI isn’t ready” crowd is correct, and where trying to force it actually wastes time. We’ve seen unit conversion issues, thousands vs millions, come up in AI-processed financial statements. If you’re building a three statement operating and LBO model, an AI that occasionally confuses units isn’t saving you time, it’s creating a liability.
This part is binary: for precise numerical work at the end of the diligence chain, you are often better off doing it yourself. The time you would spend verifying every number exceeds the time you would spend pulling it manually. The art is knowing when to stop asking AI for help.
When hallucination is the feature
This part could make your investment process-trained brain cells uncomfortable. What if sometimes you actually want AI to make things up?
Think about investment theme pressure testing. You can run AI against a thesis and have it generate ten or fifteen potential pressure points, then rank the top five for deeper analysis. Some of those pressure points are nonsense, but one or two could surface a blind spot nobody on the team had considered. The nonsense gets dismissed in seconds, but that one out of the box idea might reshape how you think about an entire sector.
The same logic applies to deal screening. AI generates questions and concerns about a target company and surfaces them to the deal team early in the process, before anyone has sunk weeks or six figures into diligence. Not every question lands, but the questions that do land, save the kind of time and money that make the misses irrelevant and well worth it.
The reason this works is that the cost of a bad idea in a brainstorming context is approximately zero. You read it, dismiss it, move on. The cost of a missed insight is potentially enormous. AI’s tendency to generate creative connections, the same quality that makes it unreliable for spreadsheets, makes it genuinely useful for expanding the surface area of what you consider.
But recognizing that hallucination can be useful is only half the challenge. The other half is designing the experience so that a human can act on it quickly. Dismissing a bad suggestion has to be trivially easy. The investor should not waste time verifying AI outputs line by line; the process should more closely resemble a brief yes/no checklist to quickly screen through. That’s a fundamentally different interaction than checking whether a revenue figure is in thousands or millions. If you design the workflow so the human is doing Quality Assurance (QA) on every cell in a spreadsheet, you’ve set AI up to fail. If you design it so the human is making judgment calls on a curated shortlist, you’ve set it up to succeed.
Finding the right use case and designing the right interface around it is the name of the game.
Designed in, not bolted on
It’s not a coincidence the firms getting real value from AI share things in common. Everyone quickly learns that clean data is table stakes: garbage in, garbage out. The firms winning designed the human review layer into the workflow from the start, rather than bolting it on after the AI disappointed them.
Bolted on looks like this: AI generates a complete analysis, human checks every line for errors, finds enough errors to lose trust, tool gets abandoned. The human is positioned as a safety net, which means they’re doing QA work that feels tedious and is adversarial to the very reason they adopted the tool in the first place.
Designed in looks like this: AI generates a first pass, human makes the judgment call that was always theirs to make. AI output is the starting material and not the final product. The human is positioned as the decision-maker, which means they’re doing the work they were hired to do, just with better inputs.
The right question to ask when building these workflows isn’t “can AI write our IC memos” but “what does our IC actually need to see in the first 30 seconds, and how do we get that in front of them faster?” When you start from there, the AI output gets designed around the partner’s existing decision-making process. The format feels familiar. The thesis filters reflect how the fund actually thinks. When something’s wrong, it’s obvious, because the partner knows what right looks like in their own context.
The other piece that doesn’t get talked about enough is iteration. Getting AI to produce useful output is not a one-shot process. Prompts, context, the structure of the input data are all extremely important. The firms that lean in and invest in refining these attributes over weeks and months end up with workflows that are dramatically better than what they started with. The firms that try it once, see mediocre results, and walk away are quitting at the worst possible time, “buying high and selling low”. The first version is always the worst version, just like building a financial model or PowerPoint deck, or any other tech tool that has to be tuned to your specific process.
The uncomfortable math
Sure, 100% accuracy is ideal. But requiring it 100% of the time is also an excuse, and a fundamental misunderstanding of your own workflows.
A fund that reviews 25 deals a week with a four-hour manual process per deal is spending 100 associate-hours per week on first-pass review. An AI that handles 90% of that work with 90% accuracy, requiring ten hours of human review to catch and fix errors, saves 80+ hours per week. That’s one to two full-time associates’ (depending on the fund and deal stage) worth of bandwidth redirected from summarization to actual analysis, judgement and execution.
The fund that waits for 100% accuracy ships nothing. The fund that designs around 90% accuracy ships a workflow that makes their team measurably better, and is compounding that value week over week.
This is not about lowering standards. It’s about understanding that different parts of the process have different standards, and that a human who reviews AI output is making better decisions than a human who’s too buried in grunt work to think clearly.
Food for Thought
When Bloomberg terminals first showed up on trading desks, they didn’t replace traders. They raised the floor. The work that used to eat up hours of an analyst’s day, pulling quotes, compiling data, tracking market movements, became a commodity overnight. The people who had been spending their time on data gathering suddenly had to justify their value through judgment, insight, and speed of decision-making. The terminal didn’t make anyone smarter, but it made the baseline level of preparation dramatically higher.
AI is doing the same thing for private equity, except it’s uneven. Some firms are already operating at a fundamentally different level of throughput and coverage, and others are still waiting for the technology to be “ready”.
There’s a broader frame here too. Public markets have always had better infrastructure: standardized data in public SEC filings, real-time market data feeds, systematic tools for analysis. Private markets have always required more manual work and processing, more ad hoc analysis on messy, inconsistent data to get the same approximate point of due diligence, albeit with more detailed disclosures and different diligence style. What AI is starting to enable is bringing that public-market discipline to the private side, not by making private markets look like public ones, but by automating the grunt work that has always made private equity analysis labor-intensive. When you raise the floor on data processing and deal screening, you free your team to do more of the human work that requires judgment, relationship building, building conviction, understanding the nuances of what makes a business tick that don’t show up during a management presentation.
Instead of asking whether AI is “accurate enough” for private equity the question everyone should be asking is, which part of the current processes are bottlenecked by accuracy, and which by throughput? For most funds, the bottleneck isn’t that they can’t analyze deals carefully enough. It’s that they can’t look at enough good deals carefully. They’re drowning in volume and starving for thinking time.
AI at 90% accuracy, deployed in the right parts of the workflow, with human judgment designed in from the start, solves the actual problem. Waiting for 100% accuracy solves a theoretical one and just allows your competitors more time to capitalize on what works now.