Documents are one of the most common data types in any organization, and one of the hardest to get right in an AI build. In this session, Kevin McGrath, Co-Founder & CEO of Meibel, and Joe Small, Senior AI & Data Science Strategist at York IE, walked through the full path from raw PDFs, scans and drawings to AI agents a business can rely on. Aaron Aguillard, Head of Strategic Growth at Meibel, hosted and ran the live Q&A.
Kevin opened with a pattern he sees in almost every customer: teams jump straight to building the agent, feed it their data, and get results that look almost right. Joe admitted he's done the same, and explained why. Document parsing and data preparation are hard to show in a board meeting. The agent demo is what gets funded.
The trouble shows up when the agent meets real documents. The ones that matter most to a business are usually the messiest: tables that span several pages, footnotes, handwritten notes in the margins. These failures don't throw errors. The model returns an answer that sounds correct, and the problem only comes to light later, through a support ticket or an audit.
Industries like construction, logistics, healthcare and chemical manufacturing run on documents that arrive in every format, from bills of lading to certificates of analysis to patient records. OCR can read the characters, but an agent needs more than that. It needs to know whether a number came from a table, the body text or a page number, and whether another part of the document contradicts it.
Kevin explained that Meibel's parsing and extraction models don't use an LLM at all. LLMs only come in later, when the extracted data needs to be related to other information. Sending whole documents to an LLM for extraction runs into context limits quickly. Even a million-token context window fills up fast with complex engineering or logistics documents.
Ready to build your first AI agent with Meibel? Sign up free, upload your documents and enable your first agentic workflow with confidence scoring & source tracing.
Kevin called the data layer the most important and most overlooked part of the stack. Every element taken from a document should be stored in a form that matches what it is. Tables belong in a relational database, vectors handle semantic search, and graph databases hold relationships instead of serving as a catch-all.
This layer also decides what context the model sees. Less context means lower costs at scale and more consistent answers, since the more context a model gets, the more its answers vary.
Joe said this is where client projects most often stall. Nobody budgets for this layer, and it is never finished: every new document type or vendor format needs someone to decide how its data is stored and found again. Clients who got 80% of the way over a weekend push back on the cost of the last 20%. In Joe's words, that 20% takes a lot longer than a weekend.
Once multiple users or clients share a system, the boundary between the LLM and the data becomes a security issue. Kevin was direct about it: a system prompt telling the model not to show certain data is not a control. If the model can see the data, there's a chance it will surface it. The only safe option is a permission and query layer outside the AI flow that keeps the data from reaching the model in the first place, enforced down to the individual session.
Joe added that prompt-based separation won't pass a contract review, a SOC 2 audit or an auditor's questions. His test before going to production: can you say, before the model runs, exactly what it's allowed to access? Is that rule enforced somewhere the model can't reach? And when a policy is unclear, does the system deny access by default?
The last step before production is measurement. Joe said nearly every client asks for a single score, and a dashboard full of 99s is a vanity metric. A calibrated score that sometimes reads 50 or 60 is more credible, and it shows human reviewers where to look, so they aren't checking everything.
Kevin said teams should track the individual measures instead of one average: completeness, coherence, extraction quality, faithfulness and data grounding. They should also evaluate the context going into the model, not only the output. What "good" means depends on the use case. A construction takeoff needs the exact number every time. A red-teaming exercise needs variance, or it isn't doing its job.
Aaron Aguillard leads Strategic Growth at Meibel, building enterprise partnerships and scaling go-to-market strategy. He brings over 15 years of experience scaling revenue and building strategic alliances in AI, SaaS, and cybersecurity. Prior to Meibel, Aaron served as Founding CRO at Qualifire (2024-2025), an AI security startup where he secured partnerships with TCS and Google Cloud and built the GTM foundation from pre-launch to enterprise traction. Before that, he spent four years as Director of Channel Sales at Namogoo (2020-2024), where he built and led global strategic partnerships with global brands including Infosys, TCS, Deloitte, and BCG.
REQUEST A DEMO
See how Meibel delivers the three Cs for AI systems that need to work at scale.
