The Signal / AI in practice / October 6, 2026
AI adoption among appraisers Only 2.2% pay. Far more have tried.
A chart in Andreessen Horowitz’s new market report says about 2% of US households pay for AI. For appraisers, the more useful measure is how deep AI goes into the work, and how to take it deeper without stepping outside AO-41.
On September 30, a chart in a 90-page investment deck set off an argument over what AI adoption looks like.
Andreessen Horowitz released State of Markets II on September 30, 2026. Among its 90 pages is a chart showing that about 2.2% of US households paid for an AI service as of April 2026, drawn from PNC Bank’s card-spending data. The share had roughly doubled in a year, and it was still small. a16z, p. 38 ↗ Crypto Briefing ↗ We read the whole report in a companion piece, AI’s next chapter is physical.
The chart traveled fast, and readers split into two camps almost at once. One camp said the public badly underestimates what current tools can do. The other said most people do not need help writing emails or booking flights, which happen to be the tasks shown in most product demos.
“People underestimate it.”
The tools can do far more than most people ask of them.
“People don’t need it.”
The demos show chores most people already handle fine.
Appraisal has its own version of that split. Lenders, AMCs and data vendors are building AI into ordering, review and valuation models. Many fee appraisers still see no reason to pay $20 a month for a chatbot. Both positions can be reasonable. Neither one is settled by the 2.2%, because of what that figure does and does not count.
The denominator
A card panel sees only some of the spending.
PNC’s figure is the share of households in its own customer panel whose personal credit or debit cards show a charge from a generative AI company. That is a clean measurement of one thing. By design it leaves out several kinds of paid and unpaid use:
- AI paid for by an employer or a school
- Subscriptions billed to a business card
- AI bundled into products people already pay for, such as office software or a phone plan
- Free tiers
- 01 / PNC Bank card data, April 2026
2.2% of US households paying
Households in PNC’s own customer panel whose personal credit or debit cards show a charge from a generative AI company.
22 per 1,000The number in the a16z chart. a16z, p. 38 ↗TechDogs ↗
- 02 / Bank of America Institute payments data
~3% of households paying
A second bank’s payment records, measured the same basic way, arriving at a similar number.
30 per 1,000Close to PNC, as expected from a similar method. Bank of America Institute ↗
- 03 / Parks Associates, April 2026
16% of households using a paid AI app
A survey of use, not card charges. Paid access through an employer, a business card or a bundle can count here.
160 per 1,000Survey of use, so it sees what card panels miss. Retail Customer Experience ↗
- 04 / Pew Research Center, February 2026
49% of US adults using AI chatbots
Adults, not households, and use, not payment. Pew also found that 24% of adults use a chatbot every day.
490 per 1,000Different population, different question. Pew report ↗Pew daily use ↗
The survey
Most who answered have already tried it.
The closest appraiser-specific data comes from McKissock, which surveyed its appraiser community in May 2026 about where AI tools had been most useful. 23% said they had not yet applied AI to their appraisal process. The rest picked their top uses. McKissock ↗
of respondents had not applied AI to their appraisal process yet. The other 77% named where it helped most.
Respondents could name more than one task, so the shares do not sum to 77%.
Read it with three discounts
- Self-selected. Respondents chose to answer, and people who use AI are likelier to answer a survey about it.
- No sample size. McKissock did not publish how many appraisers responded.
- An interested party. McKissock sells AI courses for appraisers.
Even with a heavy discount for those biases, the survey puts appraisers who have tried AI far above 2%. A neighboring benchmark points the same way: NAR’s 2025 technology survey found 46% of agents using AI-generated content such as listing descriptions. NAR ↗
Many appraisers have tried a chatbot for a small task, such as cleaning up a neighborhood description, but few have changed how a report gets produced.
Capacity
The plan tier tells you how deep the work can go.
Using AI and paying for real capacity are separate questions. In the general population the gap is wide. Among households that do pay, PNC found the large majority on plans of about $20 a month, with a much smaller share on higher-priced tiers. Yahoo Finance, citing PNC ↗
of US adults use AI chatbots
Pew, February 2026of households pay, on card and payments data
PNC, Bank of Americaaverage monthly spend among paying households
PNC, May 2026No survey breaks out appraisers by plan. We expect, though no survey has measured it, that appraisers follow the general pattern: free and $20 plans are common, higher tiers are rare. The tiers differ in ways that matter for automation.
- Tier 1
Free
$0Occasional chat with daily caps. Data terms that may not fit confidential assignment information.
- Tier 2
Standard
About $20a monthChatGPT Plus, Claude Pro, Google AI Pro
Current models, file uploads, research tools and some agent features. Usage caps that heavy daily work reaches quickly.
- Tier 3
Power
$100 or $200a monthChatGPT Pro, Claude Max, Google AI Ultra
5 to 20 times the standard usage, the top reasoning models, longer documents, and enough agent capacity to run tasks on every report.
- Tier 4
Stacked tools
VariesA power plan plus API access, automation software, data feeds and the AI features in appraisal software, connected into one workflow.
Prices as published in mid-2026 and subject to change. ChatGPT pricing ↗ Claude Pro vs Max ↗ Plan comparison ↗
The $20 plans are capable. Their limits are volume and integration. A plan sized for a few dozen conversations a day can draft a narrative or answer a USPAP question. It runs out of room when asked to reconcile data, build exhibits and run QC on every report a busy appraiser delivers. Levels two and three, described next, typically take a power plan, stacked tools, or both.
Plan choice also affects AO-41 compliance. Business and team plans typically exclude your data from model training by default, which simplifies the confidentiality analysis. Check the current terms of any plan before assignment data goes into it.
Depth
Three levels of AI depth, from trying it to depending on it.
The same a16z report looks at depth inside large companies, using what S&P 500 companies disclosed about AI in the second quarter of 2026. Most mentioned AI; far fewer disclosed a result, and almost none disclosed one they track over time. a16z calls this adoption that is broad and shallow. a16z, p. 27 ↗ The same shape fits an appraisal practice.
S&P 500 AI disclosures, Q2 2026Percent of companies · 0 to 100
Apollo Daily Spark, Sept. 11, 2026, presented by a16z. Separate disclosure categories, not stages a company passes through. Public disclosure can understate what companies measure internally.
- Level 1
Tried it
The appraiser has used AI for a narrative, a client email or a draft ROV response, on and off.
The corporate parallelIn the S&P 500, 74% of companies stated an AI plan or goal, and 69% pointed to a live deployment with usage or adoption statistics.
- Level 2
Measured it
The appraiser can say what it changed: minutes saved per report, revision requests avoided, turn time.
The corporate parallel29% of S&P 500 companies disclosed a quantified result from AI.
- Level 3
Depends on it
If the tool went away on Monday, production would drop. Data extraction into the form, comp screening, or a quality check that runs on every report before delivery.
The corporate parallel2% disclosed a metric tracked over time. None broke one out and tracked it.
Moving target
The tools have changed since appraisers last tested them.
OpenAI released GDPval in September 2025 to test AI against professionals on real work products. It covers 44 occupations, with tasks written by experts who averaged 14 years in their field. Graders compared the AI deliverable with the expert’s without knowing which was which. GDPval paper ↗
The researchers describe the improvement as roughly linear over time, and newer models have kept climbing on independent re-runs of the test. arXiv ↗ 36Kr ↗
Tests like this get outgrown, and researchers build harder ones. For appraisers, the baseline keeps rising: a chatbot that produced a weak market conditions paragraph two years ago produces a usable draft now.
This improvement reaches clients, reviewers and competing appraisers regardless of whether a given appraiser changes anything, so a workflow left unchanged falls further behind each year.
The reader of your report
Your report now lands on a desk with more AI than yours.
a16z’s David George reported that among companies, median AI vendor spending in the top 1% is eight times that of the top 10%. a16z separately put the top 1% of AI spenders at more than 600 times the median company. Some coverage described the 8x gap as a gap between individual users. The figures are about companies. a16z summary ↗ Top-1% figures ↗
In valuation, the top-spending tier is the set of organizations that order, review and model appraisals: lenders, AMCs, the GSEs, and the data and software vendors that serve them. Their AI budgets go toward their own needs, which are faster turn times, automated review flags, fraud detection and lower cost per file.
More parties at the table now have help.
Appraisers have usually known more than the other parties to a transaction about how a value opinion is built. Borrowers, agents and many loan officers could not evaluate a comparable selection or an adjustment on their own. Each of them now has help.
Pastes the report into a chatbot and gets a list of pointed questions about every comparable.
Produces a polished ROV with alternative sales in a few minutes.
Flags an adjustment that falls outside the market’s paired-sales range before a person opens the file.
People who sell AI automation services describe the same pattern in their own field. As clients learn what the tools can do, the premium for knowing how to operate them shrinks. Web design went through this earlier. Site builders turned a basic website into a weekend project, and the designers who still earn large fees are exceptional at one specific part of the job, usually strategy or sales.
Losing value as a skill
- Data entry
- Formatting
- Boilerplate
- Transcribing public records
Keeping its value
- Knowing a submarket
- Choosing and defending comparables
- Supporting adjustments with market evidence
- Reconciling to a credible conclusion
- Explaining the reasoning to a reviewer in plain language
Only the appraiser can sign for comparable selection, adjustments and the final opinion. Time freed from the first list should go toward that work.
The split
The line falls between moving information and deciding value.
The best targets are tasks that move or format information without deciding anything about value. A typical residential assignment has a long list of them.
Order intake
What AI can handle: Pull intended use, intended users, due date and special instructions from the engagement. Flag conflicts with client guidelines.What stays with the appraiser: Accepting the assignment, competency, scope of work.Scheduling and status
What AI can handle: Draft appointment confirmations and progress notes.What stays with the appraiser: Any communication touching value, or that could read as influence.Data reconciliation
What AI can handle: Compare GLA, room counts, lot size and year built across MLS, public records and plans. List the discrepancies.What stays with the appraiser: Deciding which source is reliable, and recording why.Photos
What AI can handle: Sort, label and match photos to rooms and form fields.What stays with the appraiser: Condition and quality ratings.Market data
What AI can handle: Turn your exported data into charts and a first-draft narrative.What stays with the appraiser: Market conclusions and the adjustments drawn from them.Exhibits
What AI can handle: Build location maps, trend charts and annotated photo pages.What stays with the appraiser: Confirming each exhibit is accurate and not misleading.Pre-delivery QC
What AI can handle: Check internal consistency, missing fields, math and UAD data errors.What stays with the appraiser: Fixing the issues and signing the certification.Revisions and ROVs
What AI can handle: Sort requests and draft responses that cite workfile data.What stays with the appraiser: Whether new information changes the opinion.Back office
What AI can handle: Invoices, CE tracking, license renewal dates.What stays with the appraiser: Final review of anything sent to a client or a board.
Comparable selection, adjustments, highest and best use, reconciliation and the final opinion of value stay with the appraiser. A tool can suggest comparables or run a regression, and AO-41 allows tools in research and analysis. The appraiser still has to understand the output, test it against the data, and support the conclusion in the workfile.
Anything AI supplies from memory, such as a sale price or a zoning rule, needs a source you can open before it reaches a report. We covered why in AI hallucinations: what has improved, and what has not.
One overlooked use: a clear market-trend chart or an annotated photo page often explains a conclusion faster than a paragraph does, and AI tools now make those exhibits cheap to produce.
The standard
AO-41 treats the tool as an aid and leaves the decision with the appraiser.
The Appraisal Standards Board adopted Advisory Opinion 41, Use of Technology in an Appraisal or Appraisal Review Assignment, on April 23, 2026, effective immediately. BVWire ↗ AO-41 text ↗ Like every advisory opinion, it creates no new requirements. It explains how the existing competency, ethics, scope of work and record keeping rules apply when an appraiser uses spreadsheets, regression software, AVMs, machine learning or generative AI. Board members said the intent was to encourage responsible use of technology.
The appraiser stays in control of the assignment. AO-41 says tools cannot comply with USPAP, and it treats tool output as input to the appraiser’s analysis. The appraiser decides whether that output is accurate, relevant and appropriate for the assignment. ISA ↗ ISA on the USPAP rules ↗
- 1
Evaluate the tool
Its purpose, assumptions and limitations.
- 2
Evaluate the data
The data the tool relies on, and whether it suits the assignment.
- 3
Perform the analysis
Independently. The output is one input to your analysis, and your independent judgment still decides the outcome.
- 4
Evaluate the output
Whether it is credible, accurate and appropriate for this assignment.
Three parts of AO-41 shape how the hand-off list gets automated.
Confidential information that can’t be protected stays out of the tool.
If an appraiser can’t take reasonable steps to protect confidential information and assignment results from unauthorized access, that information must not go into the tool. AO-41 adds that failing to learn how a tool handles confidential data may be careless or negligent. Before automating intake or QC, read the tool’s retention and training terms, use business accounts with training turned off where offered, and redact borrower details a task doesn’t need.
Output that supports an opinion is kept.
Tool output that supports an opinion belongs in the workfile, or the workfile must point to where it is stored. Depending on the assignment, that can mean saved chat logs, exported analysis, screenshots, or a note on how the AI-assisted output was checked.
Required only if silence would mislead.
AO-41 states that USPAP has no requirement to disclose a tool simply because it was used. Disclosure is required when leaving out the appraiser’s steps would mislead intended users. The American Society of Appraisers’ January 2026 statement recommends standard AI language in reports, so check what your designation body and your clients expect. ISA ↗ ASA ↗
The climb
One task at a time, measured.
Getting to level three means building AI into one recurring task until the task depends on it.
- 01
Pick one task
Choose one task from the hand-off table that you do on every report.
- 02
Time it first
Time it on 10 reports before you change anything.
- 03
Run the AO-41 steps
What the tool does with your data, whether its output can be checked, and how you will check it.
- 04
Build it in
Add the AI step, write down how you verify its output, and time the next 10 reports.
- 05
Keep or drop
Keep the step if the time savings hold and your error rate does not rise. Drop it if either fails.
- 06
Add the next task
Then do it again with the next one.
Start with tasks where an error is easy to spot, such as data reconciliation or invoicing. Move to tasks where an error could reach a report unnoticed, such as narrative drafting, only after your verification habit is solid.
Each task automated this way saves time without moving a value decision out of the appraiser’s hands. A practice that adds one task a month has a different cost structure within a year, and more hours for the analysis that clients and reviewers examine most closely.
Firms and lenders
Your exposure is set by your least careful user.
AI automation consultants report demand moving from building systems for clients toward workshops that teach client teams to use the tools themselves. Firms with several appraisers or trainees, and lender valuation teams, face the same question. Uneven adoption produces uneven work product. If one appraiser runs a verified QC process and another pastes borrower data into a free chatbot, the firm’s exposure is set by the second appraiser.
- Policy
Write one AI policy covering approved tools and account types, what data may be entered, how output is verified, what goes in the workfile, and when to disclose.
- Shared prompts
Build prompts and templates once, for tasks like data reconciliation and pre-delivery QC, and give everyone the same versions.
- Practice
Set a short recurring practice session so skill levels stay close.
- Review
Review trainee work that used AI the same way as any other trainee work, with the verification record included.
Choosing which tasks to automate deserves most of the thought, because the tool that handles any given task will likely change within a year.
How far the profession has to go
Appraisers have tried AI in large numbers, usually one task at a time and probably on free or $20 plans. A practice with AI at its core, where data checks, exhibits, QC and revision handling run on every report under the appraiser’s review and within AO-41, is still uncommon.
Already at level two or three?Tell us what you automated, how you verify it under AO-41, and what it changed in your turn times.→Sources checked October 5, 2026. The household figures come from different organizations using different methods and populations, and should not be compared as one series. The S&P 500 figures are public disclosures, not internal measurement. Our level-three estimate is a hypothesis, not a finding. This article is general information, not legal or compliance advice.
The Signal / Monthly briefing
Track AI in appraisal, with sources.
One sourced briefing in the first week of each month.
Unsubscribe in one click. We don't sell your email or share it with sponsors. Privacy Policy