Lesson objective
Draw a new-patient intake orchestration as numbered steps and assign each step the narrowest set of records and fields it needs to do its job.
Plain-language explanation
Why this matters now
Intake is the busiest paperwork process a practice runs, and the one most often sold as an automation opportunity. A new patient books online, answers a form, uploads an insurance card, sends a referral from another office, and arrives with a stack of outside records. Somewhere in there a practice has to capture demographics, verify coverage, scan an identity document, chase a missing element, and get the right pieces in front of the right person before the appointment.
Automating that well means thinking about it as a sequence of steps with different data appetites, not as one process that reads everything.
Regulation (United States, federal; binds covered entities and business associates): 45 CFR 164.502(b) and 164.514(d). Under 164.502(b)(1), when using or disclosing protected health information, or requesting it from another covered entity or business associate, the practice "must make reasonable efforts to limit protected health information to the minimum necessary to accomplish the intended purpose". Under 164.514(d)(2) a covered entity must identify the persons or classes of persons in its workforce who need access to PHI, the categories of PHI to which access is needed and any conditions of access, and make reasonable efforts to limit access accordingly. Paragraph (d)(3) lets routine and recurring disclosures be limited by policies or standard protocols, while other disclosures need criteria and individual review. Paragraph (d)(4) applies the same discipline to requests made to other covered entities.
Read (d)(2) again with an automation in mind. It is written about people in the workforce, and this programme applies it by analogy: AI Coalition Network recommendation — treat each automation as though it were a workforce role, and write down the categories of PHI it needs and the conditions of its access, exactly as the practice does for a registration clerk. No recorded rule says an automation is a workforce member for this purpose, so present the analogy as this programme's method, not as a requirement.
A question to sit with
If someone asked your intake automation to print everything it read about one new patient last Tuesday, how much of it would the practice have been able to justify as necessary for the step it was taking?
What goes wrong
One credential for the whole flow. The automation logs in once with access to everything and carries that access through every step, so the step that checks a policy number can also read the psychiatric history in the uploaded records.
Merge on a name. Two patients with the same name and similar dates of birth are merged or cross-filed because the matching step was allowed to decide rather than propose.
Uploads land in the chart unreviewed. A patient-supplied document is filed straight into the record, so the legal record now contains something no clinician has seen.
The form grows. Because the automation can process anything, intake starts asking for more: full medication history, household income, immigration status. Each added question has to be justified by a step that needs it.
Everything is "routine". The practice writes one standard protocol covering every intake disclosure, which is exactly what 164.514(d)(3) does not contemplate for non-routine ones.
The practical answer
Number the steps and write the data class beside each. A practical new-patient orchestration for a fictional practice looks like this. Nine steps, nine different appetites.
| # | Step | Records it may read | Fields it may write | Level |
|---|---|---|---|---|
| 1 | Receive the online booking | The booking record only | Staging table | 2 |
| 2 | Create or match a patient record | Name, date of birth, contact details of candidate matches | Nothing; proposes a match | 1 |
| 3 | Send the intake form link | The matched record's contact details and channel consent | Task status | 3 |
| 4 | Ingest the completed form | The submitted form only | Staging table | 2 |
| 5 | Classify the identity and insurance uploads | The uploaded images only | Document type, staging | 2 |
| 6 | Check eligibility | Subscriber name, date of birth, member identifier, plan | Staging result | 2 |
| 7 | Flag missing elements | The staging record only | A task for the registration clerk | 2 |
| 8 | Route anything not administrative | The triggering answer and its context | A task for clinical staff | 2 |
| 9 | Post confirmed data to the record | Staged values a clerk has approved | Demographic and coverage fields | 3 |
Two things in that table do the work. Step 2 proposes rather than decides, because a wrong match is the single most damaging error in intake and it is very hard to unpick. Step 9 is level 3 because it writes to systems other steps and people will rely on.
Give each step its own scope, not the flow's. A step that checks eligibility needs five fields; it does not need the uploaded records. AI Coalition Network recommendation: where the platform allows separate service accounts or scoped tokens per step, use them, and where it does not, record that limitation in the authority register so the practice knows what it accepted. Module 08 builds the identities.
Distinguish routine from non-routine at design time. Sending the same intake form to every new patient is routine and recurring; a standard protocol fits it. Sending a specific set of records to a specific outside office because a patient asked is not routine, and 164.514(d)(3) expects criteria and individual review. An automation should never quietly promote the second kind into the first by handling it with a template.
Ask what each question on the form is for. Write a one-line purpose against every intake field, naming the step that consumes it. Any field without a consuming step comes off the form. This is the cheapest privacy control a practice has, and an automation project is the moment it is easiest to do.
Keep patient-supplied documents outside the record until a person reviews them. A staging area is not a filing cabinet; it is a holding pen with a named reviewer. Nothing a patient or another office uploads reaches the legal medical record without someone deciding it belongs there.
Watch the practice's own habits, not only the tool. Fluent, tidy output invites staff to skim. The ECRI Top 10 Health Technology Hazards for 2026, published 21 January 2026, placed misuse of AI chatbots in healthcare first, and listed at ninth health technology implementations that prompt unsafe clinical workflows. ECRI is an independent nonprofit; its list is voluntary guidance and binds nobody. It is a useful reminder that the risk is often the workflow around the tool rather than the tool.
Remember this
- 164.502(b)(1) requires reasonable efforts to limit PHI to the minimum necessary for the purpose, and 164.514(d)(2) requires identifying who needs which categories of PHI and under what conditions.
- Treating each automation as a workforce role with a written data scope is this programme's method, not a recorded requirement.
- Number the steps and assign records and fields per step; the flow's access is not the step's access.
- Matching a patient is proposed, never decided, and nothing a patient uploads enters the legal record without a named reviewer.
- Routine and recurring disclosures may run on a standard protocol; other disclosures need criteria and individual review, and a template must not quietly convert one into the other.
Medical practice example
Brightfield Pediatrics, a fictional four-clinician practice, is replacing a paper-and-phone intake with an orchestration built on its patient-engagement platform. The first design has one service account with full chart access, "so the workflow does not break". The privacy official, Renata Chow, asks the vendor to walk through every step with her and, for each, to say what the automation actually reads. Three answers change the design. The eligibility check turns out to pull the whole demographic record including guarantor and employer, when the payer transaction needs subscriber name, date of birth, member identifier and plan; the practice narrows the query. The document classifier is configured to read prior documents in the chart "for context", which would let an intake step see a child's entire scanned history to decide whether an image is an insurance card; Renata removes that access and accepts a slightly higher rate of unclassified items, which a clerk clears in the morning. And the matching step is set to auto-merge above a confidence threshold, which the practice disables outright after Renata asks what happens when twin siblings are registered on the same day. She then works through the intake form itself with the front-desk lead, Ana Villalobos, writing a purpose line against each of its forty-one questions. Eleven have no consuming step: they were inherited from an old paper form and nobody had read them in years, including a question about household income that the practice has never used. Those come off. Two more, about previous practices and preferred pharmacy, are kept but moved after the appointment confirmation, because asking them earlier delayed bookings without benefit. The finished map runs to nine numbered steps on one page, each with its records, its fields and its authority level, and it is the document the practice hands to a locum manager.
Responsible-use warning
Do not let an intake automation decide a patient match. A wrong merge puts one person's information into another person's record, is discovered late, and is painful to unpick across the chart, the claim history and any outside office that has since received records. The matching step proposes, a named person confirms, and the practice keeps evidence of who confirmed. The same rule protects the practice from the quieter version of the error: two records for the same patient, each holding half the story.
Human-review requirement
The privacy official approves the data scope of each step against the practice's minimum necessary policy and reviews it when the flow changes. The practice manager or named administrative owner owns the intake outputs, confirms the match and the staged demographics before anything posts, and owns the escalation routes. Any clinical content that surfaces at intake belongs to a licensed clinician, who decides what enters the record. Reviews are dated and recorded so that a later question about a particular patient can be answered from the trail.
Practical exercise
Use a fictional practice, or your own with no patient data.
- Write your new-patient intake as numbered steps, from the first booking event to the moment the record is ready for the visit. Aim for between seven and twelve steps.
- Beside each step, list the records it must read and the fields it must write. Be specific: not "the chart", but the fields.
- Mark any step whose read scope is wider than its job and write the narrowed scope you would ask the vendor for.
- Assign each step an authority level from Module 01, and mark the two steps that would hurt most if they were wrong.
- Identify every step that proposes rather than decides, and name the role that confirms each proposal.
- Take your intake form and write a one-line purpose against every field, naming the step that consumes it. List the fields with no consuming step.
- Separate your intake disclosures into routine and recurring ones that a standard protocol can cover, and others that need criteria and individual review.
- Describe where patient-supplied uploads sit before a person reviews them, who that person is, and how long items normally wait.
- Write the rule for what happens when the classification step cannot tell what a document is, and time how long a clerk takes to resolve one.
- Give the finished map to someone from another department and ask them which step they would least like to be wrong. Compare their answer with yours.
Knowledge check
-
What does 45 CFR 164.502(b)(1) require, and who does it bind?
Answer guidance: It requires a covered entity or business associate to make reasonable efforts to limit protected health information to the minimum necessary to accomplish the intended purpose when using or disclosing it, or when requesting it from another covered entity or business associate. It binds covered entities and business associates.
-
Why is the patient matching step allowed only to propose?
Answer guidance: Because a wrong merge puts one patient's information into another's record, is usually found late, and is hard to unpick across the chart, the claim history and any office that has since received records. A named person confirms the match and the practice keeps evidence of who confirmed it.
-
How does this programme apply 164.514(d)(2) to an automation, and what is the caveat?
Answer guidance: It recommends treating each automation as though it were a workforce role and writing down the categories of PHI it needs and the conditions of its access, as the practice does for staff. The caveat is that no recorded rule treats an automation as a workforce member, so this is the programme's method rather than a requirement.
-
What is the difference between a routine disclosure and one needing individual review under 164.514(d)(3)?
Answer guidance: Routine and recurring disclosures may be limited by policies or standard protocols, such as sending every new patient the same intake form. Other disclosures require the entity to develop criteria and to review each request individually, and an automation must not handle the second kind with a template.
Sources and further reading
- U.S. Department of Health and Human Services (Electronic Code of Federal Regulations) — 45 CFR 164.502(b) Standard: Minimum necessary; 45 CFR 164.514(d) Implementation specifications. Accessed 2026-09-15.
- ECRI — Misuse of AI chatbots tops annual list of health technology hazards (2026-01-21). Accessed 2026-09-15.
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for LLM Applications 2026 (2026-08-04). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.