When I built a firm, my edge was never technology; it was process and client service. After I sold it, I assumed I understood how a well-run practice worked. Then AI came into the picture, and something felt unsolved and worth figuring out. That curiosity became Project Junto, an experiment with four students, a tax pipeline and an outcome that surprised me.

Where Most of Us Start and Why That’s Fine

If you’re using AI in your practice, you’re likely prompting. You ask a question, AI answers. You draft a client email faster, summarize documents or look something up. AI is bolted onto your existing processes, which is a legitimate place to start. But there’s a lot more runway.

Rather than ask where AI can help in existing processes, we asked what does a process look like if you design it around what AI can do, and then figure out where humans belong in it?

The Team

Cal Poly San Luis Obispo freshmen Kaya Paz (information systems) and Gabrielle Georgieva (economics, with an accounting concentration), and sophomores Merrick Wiessner (economics, with an accounting concentration) and John Olsen (finance) joined this project through my financial accounting class. None of them had prepared a tax return. Some of them had never written a line of code or used AI.

Describing her mindset going in, Georgieva said, “Before starting, I had minimal experience with AI and prompt engineering, and I had never worked with platforms like Intuit ProConnect before. At first, the process felt intimidating.”

The project name, by the way, came from Paz, who discovered that “junto” means a small group working together on an intelligent idea. We went with it immediately.

Our goal was specific: Take a document from client drop-off to a tax-ready return with no human intervention, inside a standard firm environment, with structured review levels built in at each stage. We used Google Workspace, though the same architecture works in Microsoft 365. The pattern is what transfers, not the specific tools.

The Tools

Google Drive handled document storage and routing. Google Document AI classified incoming files and extracted field data. Gemini ran the extraction and verification logic, though Claude, GPT or other models work equally well. We primarily used Claude to write code, running everything within a standard Google Workspace business account. Google Sheets served as the structured workpaper where every extracted value got checked. And Intuit ProConnect offers a sandbox environment where we handled entry testing using dummy client data throughout.

Real tax forms and tax returns were used, but no real client information was included. The sandbox allowed us to run fully realistic end-to-end workflows in a completely isolated environment.

The Work That Happens Before You Touch Any AI

Before writing a single line of code, we wrote out every step a human takes when processing a tax document. Not a summary. Every micro-decision.

When you do this, the complexity is surprising even to experienced preparers. This highlighted the judgment calls that are layered into what feels like routine handling: Recognizing document type from a quick scan; knowing which fields matter and which are noise; knowing when a number looks wrong in context. The complexity is surprising even to experienced preparers, and none of it is obvious until you try to make it explicit.

That’s why you can’t simply tell AI to “review the return” for the same reason you can’t say that to someone on their first day. The instruction assumes knowledge that has to be written down before it can be transferred.

Doing that work turned out to be valuable independent of anything that followed, as the firm had its own process fully documented. Once it was written down, we had a spec, a training document and an audit trail all at once.

Three Pods, One Pipeline

We organized the team into three pods named (naturally, given the audience) FIFO, Weighted Average and LIFO. Each represented a distinct stage of the pipeline: the eyes, the brain and the hands.

Paz led FIFO: document intake, classification and extraction. His job was to take whatever a client dropped off (clean PDFs, low-quality scans or multi-page Gusto exports with three copies of the same W-2) and produce standardized, structured output with a confidence rating attached. Not perfect output, but standardized output. That distinction matters.

Olsen led Weighted Average: the verification and review layer. His job was to take Paz’s extraction and ask, “Is it right?” He built logic gates for the easy checks (FICA calculations, Medicare withholding math) and structured the AI review notes to flag anomalies in plain language a human could act on. During one of our sessions, his verifier caught that a wage figure was calculating 12.4 percent higher on the Social Security line than it should have, which correctly identified a wrongly transposed number before it ever reached entry. The system flagged it, showed its work and waited for the human.

Wiessner and Georgieva led LIFO: getting verified data from the workpaper into ProConnect. Wiessner’s framing for what they were trying to build became the team’s north star: “I want this to be as easy as email. I see it, I do it.” That’s a deceptively hard target. It requires thinking about whether the data gets entered correctly and the human workflow around it. How does someone know what needs to be done, in what order and with what confidence before they start?

“When we set smart deadlines, we were able to work closer as a team because we had a better idea of what each member was working on,” Wiessner said. “Having a clear direction and endpoint allowed me to know where I want to be by the end of the week and challenge myself to go a bit further.”

The teams also created agreements defining exactly what each pod would receive and what it would hand off. That allowed the teams to work simultaneously and nobody was waiting for anyone else to finish before they could start designing their piece. And because the handoff contracts were clean, you could swap out tooling in one pod without breaking the others.

What the Pipeline Actually Does

A document arrives in our client’s upload folder. It gets classified and split. Fields are extracted and written into a structured Google Sheet workpaper. Verification checks run automatically: arithmetic gates, confidence thresholds and cross-references against other documents in the file. Items that pass move toward entry staging in ProConnect. Items that don’t surface for human review with full context, such as what was extracted, what triggered the flag and what needs a second look.

The workpaper is the trust layer: A structured checkpoint at every stage. A reviewer can see exactly what the system saw, what it extracted and where it had confidence versus where it didn’t.

What This Requires

The hardest part of this project wasn’t technical.

Georgieva progressed from having minimal AI experience and never touching ProConnect to updating JSON files, reviewing them for correctness and iterating on prompts to improve extraction outputs.

Olsen rebuilt his entire verification script midway through after realizing it was incompatible with the rest of the pipeline.

He described it this way: “When working on an individual task from a large project, it’s easy to focus entirely on the task at hand and forget that each one is a component of the final machine. What this taught me is to be careful not to get tunnel vision by always remaining cognizant of the overall purpose of the project.”

They figured it out.

“Instead of needing to be together to brainstorm, we shared ideas over text when we were available, completed our individual tasks and came back together the next week aligned and ready to move forward,” Paz reflected. “Everyone took accountability for themselves.”

What they needed wasn’t expertise. It was protected hours to think carefully about the process without competing demands. If you’re a practitioner trying to build something like this while running a full client load, that tension doesn’t resolve on its own. Find someone who can focus on it.

The Question Worth Flipping

We proved in a few weeks what we set out to: you can go from client document drop-off to a tax-ready return with no human intervention, using tools already available in a standard firm environment, with review levels built in at every stage.

Almost.

The final step into ProConnect did require a copy-paste handoff before the AI could complete the entry. The limitation, though, wasn’t the pipeline. It was the tax software itself. That’s a barrier that will get removed. And in the meantime, it’s a good example of what the whole project felt like: a series of small hurdles you solve one at a time.

Our experiment brought together accounting expertise and workflow design thinking in a way most of us weren’t trained to do. These students walked in without assumptions about how tax work was supposed to be structured. They asked why at every step, which turned out to be exactly the right starting point.

And while we focused on tax returns because the complexity made it a worthwhile proving ground, none of this is specific to tax. Anywhere in your firm where the same process runs on repeat is a candidate for the same approach.

The tools are ready. The architecture is proven. The first step is deciding to think about your own process differently: Rather than inserting AI into your processes, what would your processes look like if you designed them around what AI does well—and then figure out where you need a human in the loop.