Write the decision first
State the exact decision, deadline, geography, patient segment, and output. ‘Analyze this practice’ is not a usable assignment.
AI field guide
Use AI as a junior analyst with a source trail—not as an oracle.
AI can collect, structure, reconcile, calculate, draft, and challenge. It should not silently decide what is true, what care is appropriate, what may be billed, what a contract means, or whether a practice is worth buying.
For first-time users
Eight steps from “I have files” to a defensible answer.
The strongest prompt cannot repair a disorganized data room or an unsafe upload.
Clear the data gate
Decide whether the material contains PHI, employee data, tax records, credentials, contracts, or deal-confidential information. Use only an approved environment and authorized files.
Create an inventory
List each file, date range, source owner, system of record, export date, known gaps, and whether totals reconcile. Give every file a plain name.
Define the required file structure
Before analysis, have AI propose the rows, columns, identifiers, units, join keys, and validation rules. Approve the structure before extraction.
Extract before interpreting
Run a fact pass that cites file, page, row, cell, or URL. Then run a separate analysis pass. This makes unsupported leaps visible.
Use code for math
Have a controlled spreadsheet or script reproduce every subtotal, bridge, ratio, cohort, and scenario. Keep formulas beside results.
Red-team the answer
Use a fresh conversation or a second strong model to find missing documents, alternative explanations, double counting, leakage, and assumptions that reverse the decision.
Log and approve
Save model/version, prompt, files, source dates, output, corrections, reviewer, and final decision. A qualified human owns every consequential conclusion.
Data gate before model choice
“Not used for training” is not the same as HIPAA approval.
HHS says a covered entity or business associate using a cloud provider to process or store ePHI generally needs a business associate agreement and must still perform risk analysis and apply appropriate safeguards. OpenAI separately states that business-product data is not used to train its models by default. Both can be true; the training commitment does not replace a BAA, access control, retention policy, or organizational approval.
Use for learning, prompt testing, fictional examples, public workforce data, and published guidance. Still verify sources.
Contracts, tax returns, payroll, employee files, credentials, and seller data require authorization, NDA controls, minimum access, and an approved business workspace.
Do not place identifiable patient information in a consumer AI account. Use only an approved covered workflow with the required agreement, risk analysis, technical safeguards, permissions, retention, logging, and incident process.
Model choice · capability based
Choose the workflow first, then the least expensive model that passes your test.
Provider catalogs and model names change faster than a durable diligence guide. Set an accuracy target on a known evaluation set first; reduce cost or latency only after the alternative preserves the required result. Check each provider’s live catalog when the work begins.
TaskRequired capabilitiesEvaluation testHow to use it
Difficult synthesis, red-team review, ambiguous evidenceStrong reasoning, reliable instruction following, and tool usePasses a known set of conflicting-record and decision-reversal testsUse for build-versus-buy judgment, transaction dependencies, and final challenge. Compare current provider catalogs at the time of work.
Large document room, contracts, manuals, mixed PDFsSufficient context or retrieval, stable citations, and completeness controlsFinds every seeded clause and cites the correct file and page without blending versionsContext size is not accuracy. Inventory files, test retrieval, require file/page citations, and reconcile omissions.
Repeated extraction, classification, table cleanupLow-cost structured output after evaluationMeets a pre-set precision and recall threshold on a human-reviewed gold setUse only after a stronger workflow and a human create the test set. Recheck drift when documents or prompts change.
Scans, charts, screenshots, floor plans, EOBsMultimodal extraction with uncertainty and region referencesCorrectly extracts a representative set and flags unreadable or ambiguous regionsAsk for extraction and uncertainty first. Never let visual interpretation silently become a clinical, legal, or financial conclusion.
Formulas, reconciliations, cohort analysis, scenario tablesModel plus an approved calculation toolReproduces control totals and formulas independentlyThe calculation tool should calculate; the model should explain, test, and request missing inputs. Recompute control totals independently.
Current regulations, payer rules, workforce data, local market factsSearch-grounded research with primary-source accessOpens the governing or supporting source, captures effective dates, and distinguishes current law from summariesRestrict domains when possible, record access dates, and open the cited passage before using the result.
The reusable instruction
Paste the evidence guardrail before a task prompt.
This is the part that tells the model how to behave when the records are incomplete or contradictory.
DenQAI evidence guardrail
Paste this first in a new conversation or save it as a project instruction.
You are assisting with dental-practice decision analysis. Work only from the material I provide or current primary sources you can open.
RULES
1. Do not invent missing numbers, identities, locations, codes, contract terms, coverage, clinical facts, or legal conclusions.
2. Separate every material statement into: DIRECT RECORD, OFFICIAL GUIDANCE, INDEPENDENT ANALYSIS, ATTRIBUTABLE ACCOUNT, INFERENCE, or ILLUSTRATIVE EXAMPLE.
3. Cite each extracted fact to file name + page/row/cell, or URL + issuing organization + publication/effective date + access date.
4. If a source cannot be opened, say so. Do not cite a search snippet as final evidence.
5. Keep facts, calculations, assumptions, and recommendations in separate columns.
6. Show formulas and units. Reconcile control totals before interpreting them.
7. Assign a verification state (verified / probable / unverified / exclude). Keep source kind, support relationship, recency, commercial interest, contradictions, and missing evidence in separate fields.
8. Do not provide a diagnosis, choose treatment, invent documentation, select a higher-paying code, auto-submit a claim, or replace legal/accounting/clinical review.
9. Stop and ask questions when a missing input could materially reverse the decision.
OUTPUT
A. Executive answer in 5 bullets or fewer
B. Source-backed evidence table
C. Calculations with formulas
D. Contradictions and missing evidence
E. Downside case and decision-reversal conditions
F. Next diligence request with owner and deadline
G. Human reviewers requiredCopy-ready prompt library
Strong assignments for the work DenQAI actually requires.
Replace bracketed fields, attach only approved data, and tell the model which files are authoritative. These prompts create candidates and evidence requests; qualified people still make the decision.
Build a source-controlled market census
Turning board, NPI, payer, map, website, call, property, and group-practice results into people, sites, and control—not a dentist-name count.
Apply the DenQAI evidence guardrail already provided.
I am evaluating [GEOGRAPHY] for [PATIENT SEGMENT / PAYER PRODUCT / SERVICE SCOPE]. Create three linked tables: PERSON, SITE, and CONTROL.
For every person-site claim, capture source, source date, observed clinical days, specialty/scope, accepting-new-patient status, exact payer product, verification state, contradiction, and next recheck. Reconcile duplicate people across sites. Do not treat a license, NPI, biography, brand, or directory listing as proof of active clinical capacity.
Calculate confirmed physical FTE and segment-effective FTE as ranges. Show every factor: observed clinical days ÷ local full-time days × scope × new-patient access × payer access. Add probable capacity and credible planned entrants separately.
Return: CSV-ready tables; contradictions; call-verification list; control/ownership questions; confirmed/probable/adverse capacity cases; and the facts most likely to reverse the entry decision.Create a local research plan before browsing
You do not know which sources to collect for a county, town, commute shed, or payer segment.
Apply the DenQAI evidence guardrail already provided.
Design a primary-source research plan for [LOCATION] and [DECISION]. Separate: population and economics; provider universe; site-level capacity; ownership/control; payer access; patient utilization; wait times/leakage; hygienist and staff recruiting; real estate/construction; Medicaid; and credible entrants.
For each field, give: exact variable; preferred issuing source; backup source; geography; date range; refresh frequency; known limitation; and how the result changes build, buy, wait, or walk away. Do not collect data yet. First produce the data dictionary, source hierarchy, file-naming convention, and a 10-call field-verification script.Normalize pre-sale performance
Reconciling recent collections to what the buyer can actually repeat after close.
Apply the DenQAI evidence guardrail already provided.
Using the attached five-year and monthly records, build three separate schedules:
1. a steady-state bridge from SELLER REPORTED COLLECTIONS to BUYER-REPEATABLE COLLECTIONS using exactly five non-overlapping dollar lines: seller-only/nonrecurring activity; buyer clinical-model exclusions; patient/referral loss; buyer-specific payer/fee effect; and collection loss;
2. first-year cash timing that separates prior-service receipts, current-service deferral, and the bank-clear period; and
3. cash available for debt after recurring operating cost, market clinical compensation, replacement management, recurring capital, and a stated free-cash buffer.
Keep prepayments, credit balances, working capital, practice reserves, and assumed obligations in a separate closing sources-and-uses schedule. Test provider arrivals/departures; seller days and hours; production per day; procedures the buyer will refer; fee and payer changes; patient and hygiene continuity; new patients; unscheduled treatment; lab/material cost; staffing below market; recurring capital; and deferred replacement.
Do not use one blunt haircut or multiply stacked retention percentages. Tie every dollar adjustment to quantity × price × provider × patient cohort × time period, show overlap controls, and reconcile the underlying bank, practice-management, and tax periods. Produce reported/base/downside cases and a list of adjustments that lack sufficient evidence.Compare a payer offer line by line
A carrier has supplied a fee schedule and the office needs to know what it actually means for this practice.
Apply the DenQAI evidence guardrail already provided.
Compare the OFFERED fee schedule with OFFICE FEES, CURRENT ALLOWED AMOUNTS, and actual trailing [12/24/36]-month units. Use only code identifiers and descriptions from data the practice is licensed and authorized to supply.
For every row calculate: offered change; weighted annual revenue change; collection probability; lab/material cost; chair minutes; contribution per clinical hour; patient volume required to offset the discount; and share of total variance. Flag missing or zero fees, bundling/downcoding/LEAT questions, non-billable restrictions, leased-network exposure, amendment rights, credentialing/portability, and termination/runout terms.
Return the 20 codes driving the decision, a negotiation request by code with evidence, contract questions for counsel, and the conditions under which participation is economically rational. Do not decide coding from coverage or recommend unsupported billing.Audit treatment-mix transfer and values alignment
The target is crown-heavy, perio-heavy, implant-heavy, or otherwise dependent on a philosophy the buyer may not reproduce.
Apply the DenQAI evidence guardrail already provided.
Analyze procedure mix by provider, year, patient cohort, active patients, clinical days, units, dollars, payer, new/established status, and referral pattern. Compare crowns/onlays with direct restorations; perio diagnoses and services with complete periodontal records; hygiene with visits and recall; endodontic, surgical, implant, prosthetic, and adjunctive work with buyer capability.
A high or low rate is a SIGNAL, not a quality judgment. Do not diagnose or infer overtreatment from a percentage. Identify the blinded chart sample and tooth-/patient-level evidence an independent dentist must review: assessment, diagnosis, images/findings, alternatives, consent, documentation, follow-up, longevity, remake/refund/complaint, and open-case status.
Return buyer-retainable production by service family, philosophy-gap risk, ethical hard stops, transition communication needs, and the evidence that would explain or invalidate each outlier.Find documented work that may have been missed
Reviewing completed work, appointment events, denials, or ledger omissions without turning AI into an autonomous coder.
Apply the DenQAI evidence guardrail already provided.
Review the supplied de-identified record of what actually occurred. Extract each documented clinical or administrative event, who performed it, date/location, materials, time, images/orders/messages, existing ledger entry, claim status, and missing evidence.
For each event, create a CANDIDATE pathway only: payer claim; patient administrative charge; patient noncovered service; included/bundled; appeal/correction; write-off review; or not billable. Then list the current licensed CDT source, payer policy, plan/contract term, notice, state/federal rule, documentation, provider qualification, and human approval needed before any action.
Do not supply a final code when licensed/current code content is absent. Do not invent work, alter notes, upcode, auto-post, or auto-submit. End with a biller queue and treating-dentist review queue.Assess a role with a fair work sample
Testing billing, front desk, assisting, hygiene, management, or clinical leadership during diligence or transition.
Apply the DenQAI evidence guardrail already provided.
Build a role-specific assessment for [ROLE] in a [PRACTICE TYPE] using only job-related future-state requirements. Exclude age, health, disability, family status, personality stereotypes, retirement assumptions, and protected information.
For each competency, create: observable work sample; allowed materials/tools; time box; scoring anchors 1/3/5; evidence captured; safety or escalation expectation; backup/teaching test; environment dependency; and lawful accommodation/process note for counsel.
Score knowledge, demonstrated execution, judgment, digital/AI fluency, learning speed, documentation, and ability to teach/back up separately. Distinguish employee skill gaps from missing authority, missing software, poor training, or an owner who never delegated. Convert results into a 30/60/90-day plan with owner, access, training, backup, due date, and recheck.Red-team the investment memo
Before LOI, final lending, closing, construction commitment, or relocation.
You are the independent red-team reviewer. Apply the DenQAI evidence guardrail already provided. Do not improve the sales narrative.
Read the investment memo, source index, models, and unresolved-items log. Identify: unsupported facts; stale sources; circular assumptions; double counting; selection bias; denominator errors; seller dependence; clinical/philosophy transfer gaps; payer non-portability; staffing and wage realism; capex deferral; authority/signature dependencies; privacy/security risk; family burden; and cases where a composite score masks a hard stop.
Recompute the decision under at least five shocks: lower retention, staff exit, delayed credentialing/cash, credible entrant, and immediate capex. Add one shock specific to this market. Return the strongest case AGAINST the preferred decision, the evidence that would rebut it, and a signed-off list of conditions precedent.Keep the full prompt pack with the deal file.
Download prompt packVersion it with the source inventory, model used, output, corrections, and reviewer sign-off.
The best use of a second model
Do not ask it to agree. Ask it to break the answer.
Give the second pass the source index, formulas, and preferred decision. Require the strongest opposing case and the exact evidence that would resolve it.
Return to tools & research