Using Free AI Tools to Strengthen Your Cost Approach with Will Jarvis

About this Episode

Will Jarvis, CEO of ValueBase, recently walked through something most assessment offices know they need but rarely have time for: stress-testing the three judgment calls that underpin every cost approach value. In a hands-on session, Will demonstrated how free AI tools — Claude, ChatGPT, Gemini, and ValueBase's own ValPal — can be pointed at your depreciation tables, effective ages, and condition ratings to surface problems that have been hiding in plain sight. The central argument is not that AI should set your values. It's that the knowledge to improve your cost approach already exists in most offices. What's missing is time — and these tools compress weeks of analyst work into minutes.

The session was practical by design: five paste-ready prompts, real outputs from a Florida county with 100,000 parcels, and a weekly routine anyone can start this Friday.

Your Cost Approach Is Only as Good as Your Weakest Judgment Call

Will frames the cost approach around three judgment calls that determine defensibility: depreciation tables, effective ages, and condition ratings. Most practitioners know this. The problem, as Will puts it, is "usually time, not knowledge" that keeps these three from getting the attention they deserve. Appeals season, data quality work, deed processing — the list of urgent tasks crowds out the important ones.

The depreciation table is arguably the most consequential of the three. Small miscalibrations compound across thousands of parcels. Will walked through a single prompt that takes a CSV of improved sales (sales price, land value, replacement cost new, actual age, effective age) and extracts market depreciation by five-year age bands, fits a smooth age-life curve, and flags where your current table diverges from what the market is actually doing. In a test against 214 sales, the model identified that the existing table was running slightly heavy on new builds and flagged a thin band of older properties where the sample size warranted skepticism. Critically, the model told you where not to trust it — which is exactly the kind of output you want from an analyst.

Effective Age: Where Uniformity Goes to Die

Will didn't mince words here: effective age is "one of those things where uniformity goes to die. It lives in heads and oftentimes not in writing." Two appraisers look at the same house and land years apart. Without a written rubric anchored to observable criteria, drift is inevitable.

The approach Will demonstrated is two-layered. First, use AI to draft a one-page rubric that ties effective age adjustments to specific, exterior-verifiable criteria — roof condition, siding, window quality, mechanicals. The rubric should be written so a first-year appraiser applies it the same way a twenty-year veteran would. Second, use AI vision models to cross-check field photos against that rubric. Upload a front elevation photo, and the model will list observable condition indicators, rate them, suggest an effective age adjustment range, and — importantly — state what it cannot see and what an interior visit would need to confirm.

This last point matters. A year ago, Will notes, the image models were too inconsistent to trust. They've improved significantly, but they still work best as a tiebreaker or second opinion, not as the primary rater. The recommended workflow: photo on the phone, prompt saved as a shortcut, human makes the final call.

Condition Ratings: Let the Data Tattle on Itself

Condition rating drift is insidious because each record looks plausible in isolation. Will's approach is to load an entire neighborhood into the model and let it flag parcels where the condition rating is inconsistent with the age cohort and sale evidence. In testing, roughly ten percent of parcels get flagged — not as findings, but as questions worth a day of field checking.

The follow-up prompt generates a condition rating field guide for your staff, with two-sentence definitions, photo examples (ChatGPT can even generate illustrative images on the fly), and quarterly training cadence. Draft it with AI, mark it up with your veterans, and keep it alive. This is the unsexy process work that compounds over time.

The Ground Rules That Make This Work

Will was emphatic about safeguards, and this section deserves attention from anyone tempted to skip straight to the prompts. Strip identifiers before uploading — free-tier tools will store and potentially use your data for training. Only use publicly available information. Verify every number: "AI is a brilliant analyst and a confident liar." Will shared a telling example where a model returned an incorrect COD because it interpreted the metric using median deviation instead of average deviation. The fix was simple — specify "use IAAO formulas" in the prompt — but the error would have been invisible to someone not checking the math.

The prompt pattern Will follows is deliberate: set the role (mass appraisal, not fee appraisal or marketing), describe the data and columns explicitly, give one specific task, and demand proof. Each layer of complexity multiplies the error rate, so build up iteratively rather than one-shotting complex instructions. Document your prompts. Check your IT policy. Treat the model like a junior employee who needs supervision but can process information at extraordinary speed.

Close the Loop: Ratio Studies as the Final Check

After running cleaned-up depreciation tables, effective ages, and condition ratings back through the cost model, Will ran a ratio study against 20,000 sales. The median ratio came back a touch high. The COD was outside IAAO standards. The PRD indicated regressivity. None of this was hidden — the model surfaced it, hypothesized causes, and asked investigative questions rather than jumping to conclusions. That's the right posture for a tool in this role.

Key Takeaway

The barrier to a better cost approach was never expertise — it was the calendar. Free AI tools have collapsed the time it takes to extract market depreciation, draft rubrics, scan for outliers, and run ratio checks from weeks to minutes. But the real discipline isn't in the prompting. It's in the verification. Every office that books a weekly half-hour to run these prompts and check the outputs will compound improvements that show up in their uniformity metrics within a quarter. The offices that skip the checking step will compound something else entirely.