Large Language Models, Prompting, and Free AI Tools for Assessors with Will Jarvis

About this Episode

This episode is a solo presentation from Will Jarvis, founder and CEO of ValueBase, delivered at an IAAO conference. Rather than a typical interview, it's a practitioner's crash course on large language models — what they are, how they work, and what they mean for people who spend their careers valuing property. The central argument is straightforward and urgent: assessors who learn to use these tools will thrive; those who don't will struggle. But the nuance underneath that argument is where things get interesting.

Will doesn't sugarcoat the stakes. He draws a direct line from Sam Altman's thesis about AI-driven productivity flowing into real estate values, through the Henry George theorem, and straight into the assessment office. Property tax, he argues, sits at the center of the most important economic transformation in a century. And the people running ratio studies and answering taxpayer appeals are closer to the action than they might realize.

The Smartest Intern You've Ever Had (Who Knows Nothing About Your Office)

Will's most useful framing is this: a large language model is like the smartest intern in the world who just walked through your door with zero context about your jurisdiction, your processes, or your data. That metaphor does real work because it resets expectations in both directions. It tells the skeptic that these tools are genuinely capable. It tells the enthusiast that capability without context produces garbage.

The models operate by predicting patterns in language — that's the core insight from the 2016 "Attention Is All You Need" paper that launched the transformer architecture. They don't reason the way we do. They mimic reasoning. But as Will points out, citing AI models that solved open mathematical problems in months that had stumped human researchers for years, the mimicry has become astonishingly effective.

The practical implication for assessors is that these tools can draft correspondence, summarize appeals, organize field notes, and even enrich parcel data — but only if you tell them exactly what you need and give them the right files to work with.

Prompting Is Management, and Most People Are Bad at Both

The heart of Will's presentation is about prompting — the skill of giving a model clear, specific instructions. He compares it to management, and the analogy is sharp. Most people are mediocre managers. Most people write mediocre prompts. The results track accordingly.

A vague prompt like "explain this assessment increase" will produce a vague, possibly wrong answer. A specific prompt — one that names the attached files, specifies word count, identifies the audience, requests only information from approved sources, and asks the model to flag missing data — produces dramatically better results.

Will even notes that grammar and spelling in your prompts matter. Sloppy input pulls the model toward lower-quality regions of its training data. Encouragement helps. Specificity helps more. And asking the model to ask you clarifying questions before it begins is an underrated technique that mirrors how good employees operate: they push back when the assignment is unclear.

The real skill, Will argues, isn't typing a prompt and walking away. It's the follow-up. Ask what the model assumed. Ask what's missing. Iterate. One-shot prompting — fire and forget — is the amateur move.

Verify Everything, or Pay the Price

Will is blunt about the failure mode he sees most often: people who deploy model outputs without checking them. He calls these "AI slop grenades," borrowing the term from Shopify's CEO. Someone asks a model to build a presentation, never reviews it, and sends it along. The result looks polished. It might be wrong.

The attorney who asked ChatGPT to fabricate case law is the canonical horror story, but Will's point is subtler than that. Modern models hallucinate less frequently, but they can still be confidently wrong, especially when given incomplete context or the wrong source files. In an assessment office, where outputs affect tax bills and survive legal scrutiny, unverified model work is a liability.

This is where human judgment remains irreplaceable. The model isn't in your jurisdiction. It doesn't talk to property owners. It can't walk a neighborhood. Verification — checking facts against source records, confirming calculations, ensuring compliance with your office's approved processes — is the job that remains squarely human. Will frames this not as a limitation but as a career guarantee: the people who can prompt well and verify rigorously will be the most valuable professionals in any office.

Context Is the Bottleneck, Not Intelligence

One of Will's most counterintuitive points is that these models are not bottlenecked by intelligence — they have plenty of that. They're bottlenecked by context. Mass appraisal involves tens of thousands of parcels, hundreds of characteristics, multiple tax years, jurisdiction-specific exemptions, and local market conditions. No model arrives knowing any of that.

Will describes his own workflow: training a model on his voice and communication style, connecting it to meeting recordings and internal messages, then iteratively correcting its email drafts until the error rate dropped from 60% to 2%. That process took months. The lesson for assessors is that these tools reward sustained investment. The more context you feed them — your state's property tax code, your office's FAQ documents, your ratio study standards — the more useful they become.

Key Takeaway

The assessors who will define the next decade of this profession aren't the ones with the most technical sophistication — they're the ones who learn to manage AI the way good supervisors manage people: with clear instructions, relevant context, and rigorous follow-up. The models are already smart enough. The question is whether we're disciplined enough to use them well.