The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham

1. Key Themes

Documents Are the Interface Between Synthetic Intelligence and Human Knowledge

Abraham frames documents not as files but as the artifacts of white-collar work, which means every agent doing real work must read from and write to them. This is the origin of the pivot from memory to documents, and the foundation of Reducto's ambition. As he put it: "agents are going to do all sorts of human work and it happens to be the case that documents to us are almost these abstractions and artifacts of human work, right? So much white collar work across healthcare, and legal, and finance, and insurance is in things like PDFs and spreadsheets. You read from them, you write to them, and they catalog that knowledge." [00:02:35]

Errors Compound in Multi-Step Pipelines, So "90% Accurate" Isn't Good Enough

Accuracy on a single page is a misleading metric because real workflows chain parsing, splitting, extraction, and editing. Abraham argues that this compounding is why every incremental point of accuracy unlocks more automation. "If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right?" [00:09:37]

Frontier Models Fail Silently on Recall, Not Precision

A non-obvious benchmark finding: general-purpose frontier models are quite precise when they extract something, but they quietly omit information on long documents. This is the worst failure mode because it is nearly invisible in production. Abraham on the Macro1 benchmark: "we found actually that frontier models were better than a lot of specialized document parsers at precision... But the failure mode was actually recall for the frontier models. Like they would silently drop information. That's the kind of thing that it's really, really hard to catch in production." [00:22:08]

Harnesses and Self-Correcting Agents Can Beat Human Annotators

Reducto's DeepExtract wraps models in tooling for self-review and validation, and Abraham claims it surpasses humans on structured extraction. Human-labeled ground truth itself turns out to be noisy: "It's the first time where I find that agents can be better at structured extraction than humans on many tasks. Like we find situations where human-annotated ground truth has a lot of errors that they wouldn't have otherwise known." [00:00:00] The mechanism: "We create a lot of tools for the model to be able to review its own outputs, to create validation rules. Like if you have a massive invoice, how do you have a parent agent make sure that the items summed up to the total?" [00:22:58]

Specialization Beats Generality on Cost, Latency, and Quality

Reducto trained its own post-trained model (R1) for the narrow domain of documents. Because the domain is bounded, it can reach frontier-grade accuracy at far lower cost. "Whereas the vision encoder for a model powering ChatGPT, it needs to be able to take a selfie and, you know, turn it into Studio Ghibli image. We know the domain of information that we're working around. We need our models to be exceptional on this specific set of tasks." [00:20:24] He adds that "we are able to be orders of magnitude more aggressive on price than a frontier model at a comparable accuracy profile." [00:20:52]

The Document Problem Is Unbounded, a Layered Long Tail

Each layer of complexity reveals another: document-level, page-level, region-level. Charts, nested tables, multi-page tables, blueprints, nested PDFs, and giant spreadsheets each need dedicated approaches. "First-order problem, how do you know if a document is complex or not? And then second-order problem, how do you know if a given page is hard? But then within pages, you have regions that are harder than others." [00:28:13] On charts specifically, "we trained like a dedicated chart agent. I feel really good about saying Reducto's the state of the art there." [00:28:42]

Volume and Price Curves Signal an Infrastructure Inflection

Reducto crossed seven billion pages the morning of the recording, now processes billions of pages every month, and lists at one cent per page. The stated goal is making Reducto the default path for every page a model reads: "the thing that we want to get to is that every single page that needs to be read by a model is read through Reducto first. And that needs to be frictionless." [00:11:40] On growth pace: "We keep breaking our record for pages processed in a week." [00:12:36]

Benchmarks Have Become Marketing; Test on Your Own Data

Abraham gives customers a pointed warning about the state of AI evals. "benchmarks have become marketing over time, and the like downstream side effect is it's really hard to know what is good in your context when every... everything is like state-of-the-art all the time." [00:24:38] His recommendation: use LLM-as-judge with a thoughtful rubric on your own data distribution, and avoid overly strict matching rules since "you could represent the same information faithfully two different ways and you don't want to penalize that." [00:24:20]

The Professional-Services Market Hasn't Had Its "Cursor Moment" Yet

Abraham draws an analogy to code generation: in 2023 code tools were a fraction of today's size, and now many are past a billion in ARR, but paper-heavy professional services haven't followed. His diagnosis is that the blocker is second-order, not value. "Healthcare has so much work that is paper heavy, and insurance, and legal, and finance. I think a lot of it is like there are these second-order problems that are not just intelligence, and we want to help our customers solve this." [00:15:35]

Agents Moving From Reading to Editing and Acting

The next frontier is write-side tooling: agents that can edit Office documents, apply translations, and reconcile contradictory versions. "it's actually surprisingly hard for agents to work with the Office Suite, which is weird because that's where a lot of human work happens." [00:17:38] Reducto is launching a ChatGPT plug-in described as "a headless version of Microsoft Word that agents can use." [00:18:21]

2. Contrarian Perspectives

The Labs Won't Close This Gap, and Document Work Is Not "Just Intelligence"

Most investors assume model labs will absorb adjacent infrastructure. Abraham argues the opposite, based on the sheer depth of document corner cases and the labs' focus elsewhere: "the labs could maybe go into this deep dive of... trying to build all these like individual tools, but I personally believe that they're on an arms race for creating, you know, super intelligence for the world. And the work that we're doing is more so like infra around that that helps make that more useful." [00:00:39] He also points to concrete long-tail examples: "folders of PDFs that contain other PDFs" and no good open source tool for faithfully rendering Word files, which is "why libraries like LibreOffice often fall short." [00:19:25]

Agents Can Be More Accurate Than Humans, and Ground Truth Is the Weak Link

The prevailing view is that humans are the gold standard and AI must approach their accuracy. Abraham inverts this: "Like humans make all sorts of mistakes, right? Like people miss contents, they might drop rows, maybe they get lazy and only fill in part of the information. But if you have a harness where the agent is not only extracting the information, but correcting itself, reviewing its outputs, you kind of get to unbind that in the favor of perfect accuracy." [00:10:25] Implication: evals built on human labels, or on cheap LLM labels (he mentions "maybe they used Sonnet instead of Opus"), may be penalizing the better system. [00:26:06]

Bigger Frontier Models Aren't the Answer for PDF Parsing at Scale

It is tempting to assume the strongest general model should simply handle everything. Abraham counters that economics rule this out: "You can't throw Astra at all of your PDF parsing tasks, but we want Astra-level quality for the work that we're doing." [00:20:52] Pipelines that decompose documents and use big models only in the last mile beat one-shot calls on price and latency.

The Real Barrier to Adoption in Healthcare and Law Is Consequence, Not Culture

When asked whether slow adoption is cultural, Abraham rejects the easy answer. In code, you can roll back a bad PR; in healthcare you cannot. "I don't think we can pin it on just like willingness to adopt. I think there's a real question of like how do you have sufficient guardrails, like how do you give the human confidence in the decisions that they're making?" [00:16:42] This reframes the opportunity as a trust-and-citation infrastructure problem rather than a change-management one.

Reducto Rejects the "Intelligent Document Processing" Category Label

Rather than accepting the market-map slot of IDP, Abraham distances the company from it: "I generally don't think of the intelligent document processing market as our category because the ambition that existed in IDP is very much not what we are hoping to do. We want to be at a place where agents can help augment human work." [00:30:59]

3. Companies Identified

Reducto

Document-processing infrastructure for AI agents (parsing, splitting, extraction, editing) built on its own post-trained VLM (R1). Founded by Adit Abraham and Raunak; a16z-backed per the host framing. Mentioned as the subject. "Today, literally this morning, we crossed our seven billionth page. At this point we process billions of pages every single month." [00:11:40] List price is "One cent per page." [00:12:06]

Cursor

AI code editor. Cited as the analogy for how fast a vertical can scale once tools cross a capability threshold. "I admire Cursor a lot as a company, and I remember one of my friends came to dinner with a shaved head. They'd all shaved their heads for a million in ARR." [00:14:56] Abraham notes "now, like you look at the market and many of them are past a billion in ARR." [00:15:35]

Harvey

Legal AI company; an early Reducto customer and a model of AI-native vertical apps. "we started working with a lot of the leading AI-native companies, like the Harveys of the world, the Rogo's, and so on." [00:12:56] Abraham wants Reducto to serve "both Harvey's and the largest enterprise in the world." [00:30:31]

Rogo

Financial-services AI company. Named as an early AI-native customer. [00:12:56]

Basis

Accounting AI company. Abraham singles it out as a strong application-layer builder attacking manual data entry: "in accounting, we're seeing Basis do incredible work, for example." [00:36:33]

Macro1

Ran an independent long-extraction benchmark in partnership with Reducto, which surfaced the recall failure mode of frontier models. "There's actually a really interesting benchmark that Macro1 put out. It was independently run and we got to partner with them on it." [00:22:08]

Anthropic (Claude)

Referenced for "Claude for Excel," as evidence that Office-suite work is an emerging agent surface, and for Sonnet/Opus model tiers in the eval discussion. "I don't know if you've spent much time with like Claude for Excel and like the different plugins that are out there." [00:17:38]

OpenAI (ChatGPT)

Reducto is shipping a plug-in on ChatGPT; ChatGPT is also used as the example of a generalist vision encoder. "We are starting with a plug-in on ChatGPT where you'll be able to edit whatever you want." [00:18:21]

Unnamed Healthcare Search Provider (Reducto Customer)

A large provider of a search tool for doctors working with paper packets "inches in height." Abraham uses them to illustrate stakes. "we have this customer that works in healthcare. They're a really large provider of what's effectively a search tool for doctors to be able to capture information." [00:32:17]

Unnamed Large Accounting Services Company and BPO Provider (Reducto Customers)

A major accounting services firm handling Customs and Border Patrol forms, and a BPO doing back-office work for large companies, both cited as customers moving past manual data entry. "We work with one of the largest accounting services companies where they do like Customs and Border Patrol forms. There's a BPO company that does the back-office work for a lot of these larger companies as well." [00:36:33]

Hedge Funds, Private Equity, and Tech Companies (Customer Segments)

"we've started getting adopted by the largest hedge funds, the largest private equity firms, tech companies." [00:12:56] Abraham's framing: "if you're a hedge fund, your pipelines for aggregating information are your competitive edge in the market... if you can be faster and more robust in gathering all the signals that are available, that's alpha." [00:13:23]

Y Combinator

The accelerator Reducto went through; the pivot from memory to documents was announced the day they got in. "I remember the day we got into YC, we told Diana that we were going to pivot." [00:02:20]

LibreOffice

Open source office suite cited as falling short at faithfully rendering Word files. "there aren't good open source tools to be able to represent that faithfully. And that's why libraries like LibreOffice often fall short." [00:19:25]

Microsoft (Word/Excel)

Referenced as the dominant human-work surface agents struggle with; Reducto aims to be "a headless version of Microsoft Word that agents can use." [00:18:21]

USCIS

Cited as an example of a paper-bottlenecked government process with a huge backlog. "if you look at the backlog that currently exists for USCIS for immigration, so many of these are just like paper proxies that are bottlenecked by needing high-enough accuracy that warrants a human today." [00:34:41]

4. People Identified

Adit Abraham

Co-founder and CEO of Reducto; computer vision background. Notable for steering Reducto from memory to document infrastructure and building a specialized VLM, R1, and the DeepExtract harness. "Our background was in computer vision, and we started talking to a lot of these early AI-native companies and surprisingly kept finding this recurring trend of people were struggling with their document processing infra." [00:03:30]

Raunak

Reducto co-founder, mentioned alongside Abraham in the origin of the company. "when you and Raunak first started working on Reducto, I think working on some memory related problems." [00:02:07]

Diana (YC)

The Y Combinator partner who was told about Reducto's pivot the day they were accepted; described as "surprisingly accommodating about" the change. [00:02:20]

Jennifer Li

a16z General Partner and interviewer, who framed the data and compute inputs to LLMs and probed the lab-competition and accuracy questions. "there are two things important inputting into the models. One is data, which you know, Reducto is turning PDFs and documents into that data that fed into LLMs. The other one is just compute." [00:31:29]

5. Operating Insights

Build Parsing Outputs That Carry Citations Down to the Word

Reducto invests in citing not just which document, but which individual word justifies an answer. This enables agents to reconcile contradictory versions of a file (e.g., multiple redlined term sheets) and enables humans to trust outputs. "we put so much time into citing like not only, 'Hey, this is the answer and this is the document,' but like, 'Here's the individual word for why we know X was the answer.'" [00:14:23] Operators building agent products in regulated domains should treat traceability as a core feature, not an add-on.

Turn Production Errors Into a Training Flywheel

The R1 model is described as a culmination rather than a breakthrough, driven by systematically catching errors from the hardest customer distributions and feeding them back. "For a while we spent a lot of time on catching errors in production, like finding out how we can create feedback loops of, 'Hey, we messed up this table, how do we create a better tabular representation?'" [00:06:59] Pair this with a heavy investment in eval infrastructure and "a really good RL loop around post-training and mid-training the model." [00:07:27]

Design the Harness to Be Model-Agnostic and Self-Verifying

DeepExtract is deliberately agnostic to the underlying model and supplies tools for validation rules (such as invoice line items summing to totals) and a parent agent that decides when to re-check. "In fact, the harness is kind of agnostic to the model." [00:22:38] This lets the product ride model improvements without rework while compensating for model weaknesses.

Preprocess Messy Inputs Into Model-Friendly Representations

Don't hand raw files to a model. For Excel, Abraham notes "there's a lot you can do of creating a representation of the Excel file that is meant for models to consume as opposed to just giving the raw file directly." [00:27:01] Route complexity adaptively: determine document, page, and region difficulty, then apply heavier pipelines only where needed. [00:28:13]

Help Customers Validate Their Evals Before the Head-to-Head

Golden datasets are often flawed (human omissions, cheaper LLM labelers). Reducto proactively partners with prospects to validate the eval itself first. "we try to partner with teams a lot to make sure that they feel good about how they're going to test before they even get into any sort of head-to-head comparison." [00:26:29] This de-risks sales cycles and avoids losing to a bad benchmark.

6. Overlooked Insights

Handwritten and Physical-Packet Medical Records Show "Sub-50%" Accuracy in Incumbent Tools, a Large Unaddressed Frontier

Mentioned almost as an aside, but it signals how much value sits beyond the "high 90s" clean-document regime. Of one healthcare customer's hardest cases, Abraham says "there are cases where it's like sub-50%. Like the best you can do is have some approximation of where information sits, and then a human has to look at it. But like a given patient's packet could be tens of thousands of pages." [00:33:44] Paired with the stat that "in 80% of appeals, the original decision was overturned" for Medicare prior-auth claims [00:34:39], this points to a vast, quantifiable inefficiency in adjudication that better document fidelity could address.

Compounding Volume Growth Plus a Penny-Per-Page Price Implies Document Reading Is Becoming a Utility

Abraham casually notes the time to each new billion pages keeps shrinking and that he expects to pass eight billion within weeks of passing seven (targeting "end of October" for the announcement). [00:12:31] Combined with the stated goal of negotiating pricing for "tens of billions of pages" [00:12:06] and a stated aim of being the default reader for every model-read page, this reads less like a SaaS tool and more like a metered infrastructure primitive, akin to payments or cloud storage, whose value accrues to whoever becomes the standard path.