How We Built K1Manager with AI — Dynamic Automation

Building a tax technology product is an unusual intersection of domain knowledge and software engineering. Most tax professionals don’t write code. Most software engineers don’t understand the nuances in compliance and accounting. K1Manager started with the mission of bridging that gap.

This is the story of how — and more importantly, why.

The problem that couldn’t be ignored

In the fund-of-funds tax compliance world, manual K-1 processing is one of those things everyone deals with but few question. Every season, the same pattern: hundreds of K-1 PDFs arrive, each one needs manual data entry into a template, and the work consumes the team for weeks.

The tools have evolved over the years — from paper and typewriter to early Excel, then macro-based solutions, and eventually proprietary internal software. These solutions helped run calculations faster and enforce consistency, but they added layers on top of the fundamental data rather than changing how it moved.

The core of the process — extracting values from a stack of dynamic PDFs, applying basic tax judgment, and manipulating amounts into a workpaper — stayed the same. Whether you were doing it on paper or in Excel, the mental work was the same. That layer, the part where a staff or senior is reading, interpreting, and moving data, is what never really got automated.

The forms are digital. But the workflow is still manual. Open PDF, read value, type into Excel, move to the next field. Repeat thousands of times.

It seemed like there had to be a better way. K1Manager was a bet that there was.

Why build something new?

The existing landscape has options, but each comes with trade-offs:

ML-based extractors use machine learning to read K-1 PDFs and return extracted values. They work — and for edge cases and unusual formats, they’re the right tool. But they need tuning, they carry token costs (think blank K-1s, K-1s with fewer than five line items), and there are data security considerations when sensitive financial data is involved. The thing is, there’s a lot of consistency in any given K-1 population. When you can leverage that consistency with deterministic parse logic on the vast majority of documents, you reserve the ML route for where it’s actually needed. It’s not one or the other — it’s using each where it’s strongest.
Generic OCR solutions — like built-in Adobe PDF converters — extract raw values, but the output is inconsistent. You can’t build a reliable workflow on top of an extraction layer that doesn’t give you structured, predictable results.
Offshore teams reduce cost, but they don’t reduce error rates, and they introduce communication overhead and turnaround delays.

Building consistent, reliable footnote detection in a deterministic way — that’s where understanding the consistencies in how K-1s are prepared makes a real difference.

Building with AI as an accelerator

K1Manager wasn’t built by AI — it was built with AI.

AI served as a development accelerator, but not in the way you might think. This wasn’t a no-code exercise where someone described a problem and AI produced a product. It required breaking the process down step by step, thinking through every path the data needs to travel, and piecing that together into coherent logic at the platform level.

That means understanding the process from multiple angles:

Process decomposition: Breaking down each step of K-1 processing — not just what happens, but what could happen at each stage. A platform has to account for all the possible paths a piece of data might take, and architecting that logic is the real work.
Architecture decisions: Exploring different approaches to state management, data flow, and component design. AI provides rapid feedback on trade-offs, letting you iterate on structural decisions that would otherwise take much longer to evaluate.
Domain-code bridging: The hardest part of building tax software isn’t the code — it’s encoding how people actually work at every level. How a staff prepares the data. How a senior reviews it. How a manager assesses quality. What KPIs and data points matter to a senior manager when making engagement-level decisions. Having domain knowledge across those levels is what lets you build a solution that actually delivers — and AI helps translate that understanding into data structures, algorithms, and workflows.
Code comprehension and debugging: AI generates code, but you can’t take the output at face value. You need a fundamental understanding of how code works — enough to read through it, evaluate whether the logic matches your intent, and catch when it doesn’t. That means thinking through different approaches yourself, not just accepting the first suggestion. It means debugging when things break, understanding why they broke, and being able to guide the next iteration. There’s a real development skillset involved, even if you’re not writing every line from scratch.

The result is a product shaped by someone who understands the workflow, accelerated by AI tools that made the engineering more accessible — but not automatic.

What we learned

Understanding the flow of data is what matters. Every process has data moving through it — inputs, transformations, outputs. The starting point is understanding what that data is, where it lives, and how current solutions handle it. From there, it’s about asking whether the way it moves today is the way it has to move.

Leverage consistency where it exists. K-1s and K-1 preparation follow patterns. Footnotes have structures. Workpapers have conventions. When you recognize those consistencies, you can build deterministic parse logic that handles the bulk of the work reliably, and use ML for the genuine edge cases. You don’t have to choose one approach — you use each where it fits.

Build for the workflow, not just the extraction. A data extraction tool is useful. But a platform that integrates your tracker, your workpaper, your SharePoint K-1 population — and writes the results back to your data template accurately — that’s what actually changes the day-to-day. K1Manager is designed around the end-to-end process: extract, track, modify, review, aggregate, and output.

What’s next

K1Manager is the first product, but the vision is broader. As the industry shifts with AI, there’s a real opportunity to evaluate the intensive manual workflows across tax and accounting — and to apply domain expertise into focused automation that fits the way professionals actually work.

We’re identifying those processes one at a time, and building tools that let you lean into AI’s strengths while you focus on what you do best: making judgment calls.

Interested in learning more? Request a demo to see K1Manager in action.