The client who disappears after the second visit rarely disappears because the plan was bad. They disappear because they had a rough week, ate pizza on Saturday, felt embarrassed and did not want to send a message. By the time you notice, 40 days have passed and the follow-up slot on your calendar is empty. A simple weekly check-in prevents a good share of that, but nobody has time to send 35 messages every Monday and read 35 replies.
This post walks through an AI agent for dietitians with two tightly limited jobs: organizing pre-visit intake (the form a patient fills out before the initial assessment) and running the weekly check-in by text or through your patient portal. It collects, organizes and alerts you. It does not counsel, does not change the plan, and does not answer clinical questions. That limit is not just good sense. Under your state license and the CDR Code of Ethics, the nutrition care is yours, and a bot that "does the visit" puts both at risk. For the big picture, see the AI for dietitians guide.
Workflow 1: pre-visit intake by form
The goal is to walk into the first session already knowing the basics and spend the hour on what matters: diet history, relationship with food, the plan you agree on together. The workflow:
- When the patient books, they get a link to the intake form. Use the form inside your practice platform (Healthie, Practice Better and similar) or a form tool that signs a BAA. Plain Google Forms on a personal account is not the place for health data.
- When they submit, the answers land in the platform or in a protected sheet.
- An automation (n8n self-hosted, or the platform's own workflow feature) takes the new submission and sends it to an AI model under a BAA, with the instruction below.
- The AI returns a one-page summary that goes to your secure inbox or the chart, ready to read five minutes before the session.
Questions I would put on the form (short, mostly multiple choice): goal in one sentence, current medications (including GLP-1s and insulin), diagnoses, labs from the last 6 months (upload), pregnancy or breastfeeding, wake and sleep times, who cooks at home, meals eaten out per week, weekly grocery budget (range), foods they will not eat, allergies and intolerances, alcohol, physical activity, episodes of eating out of control (yes, no, prefer to talk in session), diets they have tried. Keep it under 20 questions. Above that, completion rates drop fast.
The AI instruction for the intake summary
You organize pre-visit intake answers for a registered dietitian. Do not diagnose, do not suggest a diet, do not interpret lab values. Return a summary with: 1. FLAGS at the top (only if present): pregnancy or breastfeeding; medications that affect blood sugar, appetite or absorption (including GLP-1 medications and insulin); bariatric surgery; answers that suggest disordered eating (bingeing, purging, severe restriction); allergies. 2. Goal in the patient's own words, in quotes. 3. Routine in 4 lines: schedule, who cooks, meals out, budget. 4. Preferences and foods they avoid. 5. Diet history. 6. Questions left blank or answered unclearly, for the dietitian to explore. Use only what is in the answers. If something is ambiguous, write [ASK IN SESSION]. Answers: [FORM ANSWERS]
Real output
I ran the instruction on answers from a fictional patient (32, wants to lose weight, works second shift, wrote "sometimes I eat a lot at night and then feel awful"):
FLAGS: Takes sertraline 50 mg (reported). Describes episodes of eating a lot at night followed by feeling unwell: [ASK IN SESSION] how often, and whether there is any compensating behavior.
Goal: "I want to lose about 18 pounds and stop snacking at night."
Routine: Works 2 p.m. to 10:30 p.m. at a distribution center, wakes at 9:30 a.m., sleeps around 1 a.m. Cooks for herself, mostly on weekends. Eats out 5 times a week (dinner from the break room vending machines or fast food on the way home). Groceries: US$ 70 to US$ 90 a week, lives alone.
Preferences: Likes rice, chicken, tacos, fruit. Does not eat fish or liver. Drinks 3 cups of coffee a day with sugar and creamer.
History: Keto for 2 months in 2025, "lost 12 pounds and gained it all back." Tried intermittent fasting on her own.
To explore: Did not answer the physical activity question; attached labs are from 2024 [ASK IN SESSION whether there are newer ones].
Notice that the AI did not call the night episodes binge eating and did not suggest anything. It just put them at the top so you can raise them. That is the right behavior, and it depends on the first line of the instruction.
Workflow 2: the weekly check-in by text
Every Monday at 10 a.m., each active patient gets a short message with three questions. The AI summarizes the replies and you get an ordered list: first the people who need you, then everyone else. The one who answers the patient is you, in your own voice.
The message I would use:
Hi {first_name}, this is the assistant for {dietitian_name}, RDN, with your weekly check-in. Reply however you like, even a voice memo in the portal:
1. From 0 to 10, how closely were you able to follow the plan?
2. How was your hunger, especially afternoons and evenings?
3. Did anything get in the way (work, a party, digestion, sleep)?
{dietitian_name} reads every reply and gets back to you by Wednesday. Reply STOP to stop these check-ins.
Notice what is not there: weight. Asking for weight every week through an automation can harm patients with a history of disordered eating and creates anxiety for a lot of people. If you want to track weight, agree on it in session with the patients for whom it makes sense.
About the channel. Texting is what patients actually read in the US, but plain SMS is not encrypted. The safest setup is to send the check-in as a portal message with a text notification ("You have a new message from your dietitian"). If a patient prefers plain SMS, document their consent to unencrypted messages, keep the questions free of diagnoses, and register your number for A2P 10DLC so carriers do not filter it. Get written opt-in before the first automated text; the TCPA covers automated texts, and the STOP line is not optional.
The AI instruction that summarizes the check-ins
You summarize weekly check-in replies from patients for a registered dietitian. Do not reply to the patient and do not give advice. For each patient, one line: name | adherence score | summary in 15 words or fewer | signal (RED, YELLOW or GREEN). RED: score of 4 or lower, mention of intense guilt, vomiting, skipping meals entirely, a new physical symptom, or any question about medication. YELLOW: score of 5 to 7, or one concrete obstacle (travel, party, night shift). GREEN: score of 8 or higher, no obstacle. Sort: REDs first. Transcribed voice notes are marked [VOICE]. Replies: [REPLIES]
Sample output with five test replies:
Jessica | 3 | "awful week, ate everything wrong, can't even look at the plan" | RED (intense guilt)
Robert | 7 | very hungry around 5 p.m., gets to dinner starving | YELLOW
Amanda | 6 | work trip Wednesday to Friday, ate at restaurants | YELLOW
Paula | 9 | followed it well, digestion better with the oatmeal | GREEN
Mark | 8 | fine, asks if he can swap rice for pasta at dinner | GREEN
With that list you spend 20 minutes on Monday: you call or send a voice memo to Jessica, give Robert a fix for his afternoon snack, and leave short replies for Paula and Mark. Without the list, you would read 35 threads in the order they arrived. And notice Mark: the pasta question looks harmless, but it is still a nutrition question, so it waits for you.
What the agent never does
- It does not answer questions about food, supplements or medication. The standard reply is "I'll pass this to {dietitian_name}, who will get back to you."
- It does not change the meal plan or send a new one.
- It does not comment on weight, plate photos or bodies.
- It does not pretend to be you. It introduces itself as an assistant, which is also the honest way to handle disclosure rules some states now apply to AI-written health messages.
- It does not text outside the hours you set, and never after someone replies STOP.
Tools and monthly cost
| Piece | Option I would use | Approximate cost (check the site) |
|---|---|---|
| Form, portal and messaging | Your practice platform (Healthie, Practice Better and similar), which signs a BAA | Usually already in your plan |
| Automation | n8n self-hosted on a server you control, or the platform's built-in workflows | Free to about US$ 24/month, plus hosting |
| AI model | OpenAI or Anthropic API, small model, under a BAA (both offer one to eligible API customers on request) | A few dollars a month for 50 patients |
| Text notifications | Twilio or the platform's SMS, with A2P 10DLC registration | Around one cent per message plus carrier fees, and a small monthly registration fee |
| No-build alternative | The check-in and reminder features already inside your platform | Varies by plan |
For 50 active patients, the Monday texts cost under US$ 5 a month. The real cost is the automation and the afternoon it takes to set it up. If you do not want to build anything, there is a middle path: send the check-in as a saved message template from your portal, then paste the week's replies, with names replaced by initials, into the AI with the summary instruction. That is 15 minutes of manual work and costs nothing extra.
HIPAA: what changes when the agent touches health data
- Every vendor that sees PHI (form, automation, AI model, texting service) needs a signed BAA. No BAA, no PHI. This is the rule that rules out most "connect anything to anything" automation apps on their standard plans.
- Use the API under a business agreement, never your personal ChatGPT account. Ask the vendor about data retention and turn on the strictest option available to you.
- Limit access to the summaries to you and anyone you supervise, and delete data from patients who ended care after the period stated in your records policy.
- Ask for consent on the form, saying that the answers will be organized with the help of software that includes AI.
When I would NOT build this agent
- Fewer than 15 active patients. Send the Monday message by hand from the portal. The automation does not pay back the setup time.
- Patients in eating disorder treatment. An automated check-in can feel cold or like being policed. Here the contact is personal and coordinated with the treatment team.
- A premium practice. If your selling point is "you talk to me directly," an assistant can feel like a downgrade. Use only the intake summary, which the patient never sees.
Common mistakes
- Letting the agent "help" with a simple question. "Can I swap rice for pasta?" sounds harmless, but it is counseling. It goes to you.
- A 60-question form. Above 20 questions, completion drops. Save the rest for the session.
- Promising a reply and not delivering. If the message says "by Wednesday," reply by Wednesday. Otherwise the check-in hurts the relationship.
- Forgetting to pause. A patient who ended care keeps getting Monday texts. Keep an "active" field and filter on it.
- Skipping 10DLC registration. Then wondering why half the texts never arrive.
Ready-made replies for each signal (red, yellow, green) and the meal plan prompts are in the ChatGPT prompts for dietitians.

Want to skip this part?
Ready-made AI agents for dietitians: nothing to build, no prompts to write, they already work for you.
- Turns intake forms into a one-page summary with the red flags on top
- Sends the Monday check-in and shows you first who needs attention
- Collects and organizes, and never answers a nutrition question in your place
Step by step
- Choose where PHI lives (your practice platform under a BAA) and list every vendor that will touch it
- Build an intake form with under 20 questions and a consent line about AI-assisted organization
- Connect the form to an AI model under a BAA with the intake summary instruction
- Set up the Monday check-in with written opt-in, a STOP option and an active-patient filter
- Run the summary instruction on replies and block 20 minutes every Monday to answer, reds first
Frequently asked questions
Can a dietitian use a chatbot with patients?
Yes, for administrative tasks and collecting information, introduced as an assistant. It should not counsel, answer nutrition questions or replace you in the conversation with the patient.
How much does a check-in agent cost?
For 50 patients, the texts cost under US$ 5 a month; adding automation and the AI model, the total runs about US$ 30 to US$ 80. The manual version with portal templates costs nothing extra.
What questions should go on the intake form?
Goal, medications, diagnoses, pregnancy, daily schedule, who cooks, meals out, budget, foods avoided, allergies, alcohol, activity and past diets. Keep it under 20 questions.
Should the weekly check-in ask for weight?
Not through an automation. It can cause anxiety and harm patients with a history of disordered eating. Agree on weight tracking in session, patient by patient.
Is it HIPAA compliant to use AI with patient data?
Only if every vendor that sees PHI has signed a BAA and you use business or API accounts, never a personal chatbot account. Ask for consent on the form and limit access to the summaries.
Can I send check-ins by regular text message?
Yes, with written opt-in for automated texts under the TCPA, a STOP option, A2P 10DLC registration and the patient's consent to unencrypted messages. Keep diagnoses out of the text.
