The slowest part of answering a client tax question is rarely writing the email. It is turning a loose sentence into the right issue, finding the rule that governs it, and deciding which missing fact could reverse the answer.
AI can compress that work, but only if the workflow is designed around the real bottleneck. Asking a model, “Can my client deduct this?” produces a fast paragraph and a slow review. You still have to reconstruct its assumptions, check its citations, and work out whether it answered the question the client actually asked.
A better workflow produces three things together: a plain-English answer, an openable primary-source citation, and a short file note showing how you got there. For a routine question, that can often be done in about 15 minutes. The goal is not 15 minutes at any cost; complexity still gets the time it deserves. The goal is to stop spending 15 minutes drafting before the research has begun.
Start with a decision, not a prompt
A client message usually arrives as a topic:
I bought a new SUV this year. Can I write off the interest?
That is not yet a research question. It omits the facts that determine the result: when the debt was incurred, whether the vehicle was new to the taxpayer, where final assembly occurred, whether the loan is secured by the vehicle, how the vehicle is used, the amount of interest, modified adjusted gross income, and filing status.
The first useful job for AI is not answering. It is fact extraction. Give the model a de-identified version of the message and ask:
Convert this message into (1) the tax issue, (2) facts stated by the client, (3) facts that are missing, and (4) search terms for primary authority. Do not answer the tax question or supply citations.
That instruction keeps the model in a job it does well. It can organize the story and expose assumptions without pretending to know the governing subsection.
If the tool is not approved for client data, remove names, identification numbers, account numbers, addresses, and other unnecessary details before the message goes in. The model needs the operative facts, not the identity behind them.
The 15-minute workflow
For a bounded federal question with no factual dispute, the sequence looks like this:
| Time | Job | Output |
|---|---|---|
| 0–3 minutes | Frame the issue | Stated facts, missing facts, tax year, entity, jurisdiction |
| 3–7 minutes | Retrieve authority | A small set of primary sources with exact identifiers and links |
| 7–11 minutes | Read and test | Operative text, exceptions, effective dates, adverse rule |
| 11–14 minutes | Draft | A conditional client answer and follow-up request |
| 14–15 minutes | Preserve | Source links, excerpts, assumptions, reviewer conclusion |
This is a triage workflow, not a promise that every question belongs in a quarter hour. A multistate position, a disputed classification, or a transaction with meaningful exposure should leave the fast lane as soon as its complexity becomes visible. A good 15-minute process is valuable partly because it tells you when not to give a 15-minute answer.
Step 1: Build a fact gate
Before searching, turn the missing facts into a gate: the minimum set you need before the answer can become final.
For the vehicle-interest question, the gate might be:
- Was the debt incurred after December 31, 2024?
- Was the vehicle purchased new, with its original use beginning with the client?
- Did final assembly occur in the United States?
- Is the debt secured by a first lien on the vehicle?
- Was the vehicle purchased for personal use?
- How much qualifying interest was paid during the year?
- What are the client’s filing status and modified adjusted gross income?
- Does the client have the VIN and the lender’s interest statement?
The list does two useful things. It prevents a premature yes, and it gives the client a short, specific request instead of a vague “send more information.”
Facts are where speed is usually won. If the client says the SUV was used, the research can stop early. If final assembly was outside the United States, the federal personal-interest deduction is not available. You do not need three pages of analysis to identify a failed threshold condition.
Step 2: Retrieve a rule card, not an essay
The research request should ask for a compact rule card:
Find the current federal authorities governing qualified passenger vehicle loan interest for tax year 2026. Return the exact section and subsection, the relevant text, effective dates, limitations, and a primary-source URL. Separate enacted law from proposed guidance. Do not draft a client answer.
For this question, the source set begins with IRC §163(h)(4). The statute creates a temporary exception to the personal-interest disallowance for qualifying vehicle-loan interest in tax years 2025 through 2028. It also contains the $10,000 annual limit, the income-based reduction, and the VIN requirement.
IRS Topic No. 505 supplies a current administrative summary: the debt must have been incurred after December 31, 2024, the vehicle’s original use must begin with the taxpayer, final assembly must occur in the United States, and the debt must be secured by a first lien. The Treasury Department and IRS have also published proposed regulations under REG-113515-25 addressing definitions and operating rules.
The status labels matter. The Code is enacted law. An IRS topic is practical guidance. A notice of proposed rulemaking is not a final regulation. A fluent summary that blends all three together makes review harder, not faster.
This separation is also why a connected AI tax research assistant is different from a general chatbot. The useful output is not a paragraph that sounds researched. It is the authority the paragraph can be checked against.
Step 3: Run four checks before drafting
Open the sources and check four things:
- Existence: Does the cited provision actually exist at the stated address?
- Fit: Does it govern this taxpayer, tax year, transaction, and jurisdiction?
- Completeness: Did the result include the definitions, limits, exceptions, and phaseouts that can change the conclusion?
- Status: Is the source enacted law, a final regulation, proposed guidance, an IRS explanation, or a court opinion?
For the car-loan question, “up to $10,000” is not a conclusion. Under §163(h)(4), the otherwise allowable amount is reduced when modified adjusted gross income exceeds $100,000, or $200,000 on a joint return. The return must include the VIN. The original use and final-assembly requirements can end the analysis before the dollar limit matters.
This is the point at which a fast answer becomes a defensible one. You are not rereading everything the model knows about automobiles. You are testing a short list of controlling conditions against the client’s facts.
Step 4: Draft two answers from the same research
Once the authorities are verified, ask the model for two outputs: a client-facing response and a file note. Give it only the approved facts and sources, and tell it not to introduce new citations.
The client version might begin:
Potentially. For tax years 2025 through 2028, interest on certain loans for new, U.S.-assembled vehicles can be deductible even if you do not itemize. The loan and vehicle have to meet several conditions, the deduction is capped and can be reduced based on income, and the VIN must be reported. Before I confirm the amount, please send the purchase date, VIN, lender’s year-end interest statement, and confirmation that the vehicle was purchased new.
That answer is short because the research behind it is organized. It leads with the conclusion’s status—potentially—then gives the rule at client altitude and asks for the facts needed to finish.
The companion file note can remain technical:
- Client question and tax year
- Facts received and facts still assumed
§163(h)(4)as the governing statutory provision- Relevant IRS guidance and its status
- Eligibility conditions tested
- Income limitation and VIN requirement
- Source URLs and date accessed
- Reviewer conclusion and next step
The two outputs serve different readers, but they should never contain different reasoning. The client gets clarity; the file gets the trail.
Step 5: Make uncertainty visible
AI drafts tend to smooth over unresolved facts because a complete paragraph sounds better than an incomplete one. Tax work needs the opposite instinct. A useful response makes the boundary visible:
- Known: the client purchased a vehicle in 2026 and financed it.
- Not yet known: whether the vehicle and debt satisfy each statutory condition.
- Current answer: potentially deductible, subject to verification.
- Next action: collect the VIN, purchase and loan documents, interest statement, and income information.
“I need two more facts before I can answer” is often the fastest correct response. The model should help you identify those facts, not write around their absence.
What not to automate
Keep three decisions with the practitioner.
Do not let the model decide which facts are true. It can identify a missing fact; it cannot infer that a client’s vehicle was assembled in the United States because the manufacturer is American.
Do not let it upgrade a source’s authority. A publication or IRS webpage can be useful without becoming the statute. Proposed regulations should remain labeled proposed.
Do not let it turn a conditional result into a final conclusion. The model is rewarded for completing the answer. Your job is to stop where the evidence stops.
The same division of labor applies when the deliverable is longer. In an AI-assisted tax research memo, the facts and conclusion still belong to the practitioner, the authorities still come from sources that can be opened, and the model still earns its keep by drafting from verified inputs.
Fast means a shorter path to evidence
The wrong measure of speed is how quickly text appears on screen. The right measure is how quickly you can reach an answer that another reviewer can reproduce.
Frame the issue before asking for a conclusion. Retrieve a rule card before requesting prose. Open the authority before relying on the summary. Draft the client answer and file note from the same verified set.
That is how AI makes a client response faster without making the review shallower. It does not remove the research trail. It shortens the path to one.
This article is general information, not tax advice. The vehicle-interest example is illustrative; verify current authority and the client’s facts before applying it to an engagement.