AI Training for Dietitians: A Practice Owner's Team Plan

Teach the data rule first: what may never be typed into a general AI tool (client names, contact details, health information, anything from a consultation). Then run a six-week programme of one 45-minute session a week: rules and accounts, client emails, handouts, templates, a privacy drill and a sign-off. Every session ends with a task and a check, not a slide deck.

Most training for dietitians is a course list: a webinar here, a certificate there. What a practice owner actually needs is a plan for the team in the building. People are busy, some are sceptical, and one of them has probably already pasted a client email into a free chatbot. The plan below assumes a small team (an owner, two or three dietitians, an administrator) and treats AI as a drafting assistant for words, never as a source of dietary advice or a keeper of client records.

Follow me on Instagram@sagnikteaches

What to teach first, and why the order matters

Teams usually want to start with the exciting part: "write me a meal plan". Resist that. The order that works is rules, then low-risk tasks, then judgement tasks, then habits.

Connect on LinkedInSagnik Bhattacharya

1. The data rule. It protects clients and the practice, and it is the one thing that cannot be undone. A pasted record cannot be taken back.
2. Low-risk writing. Newsletters, website copy, appointment reminders: no client data, easy to review.
3. Client-facing drafts. Emails and handouts, where accuracy and tone matter and a dietitian must check every claim.
4. Habits. Review steps, a log, and a monthly refresh of the rules.

Leave anything clinical (calorie targets, meal plans for a condition, supplement advice, interpreting lab results) out of the first six weeks. If it comes up, the answer is that these remain the dietitian's professional judgement; see what stays human for dietitians using AI for admin and content for the line. Charting notes are a separate project with their own safeguards, covered in drafting ADIME notes with a review step.

Subscribe on YouTube@sagnikteaches

The data rule, in words a team can remember

A rule people can recall on a busy Thursday has three parts and fits on a card:

• Never in: a client's name, email, phone, address, date of birth, weight, diagnosis, lab values, food diary, photos, or anything from a session. Also nothing from another provider's letter.
• Only with the right account: if a task seems to need client data, stop and ask the owner. The answer may be "use the practice software".
• Always checked: a person reads, corrects and owns every word before it goes to a client.

The agreements matter if you work in a setting covered by health-privacy rules such as HIPAA, where a vendor must sign a business associate agreement (a BAA) before handling protected health information for you. As of November 2026 OpenAI says it does not offer a BAA for ChatGPT Business, Anthropic's BAA covers its Enterprise plans (once the primary owner turns on the health-privacy setting) and its API, and Google lets Workspace admins accept a BAA in the Admin console. Plans and covered features change, so treat this as a prompt to check the vendor page on the day, not as a recommendation. If you have clients under data-protection law such as the GDPR, the same idea applies through a data processing agreement. The practical consequence for training is simple: until you have the right agreement in place, the rule is no client data, and the training exercises use invented clients.

The six-week plan at a glance

A note on time. For a five-person team, 45 minutes of session plus 30 minutes of practice is 75 minutes per person per week, so six weeks is 7.5 hours each and about 37 hours in all. At an illustrative staff cost of $40 an hour, that is roughly $1,500 of time. Trim the between-session tasks for the administrator and the total falls nearer 30 hours. Compare these numbers with the benchmarks in how much time and money AI staff training takes.

Week 1: rules, accounts and honest conversation

Open by asking each person a question: "What have you already used AI for?" Go round the table. You will learn more in ten minutes than from any survey, and you set the tone that disclosure is safe. Then cover three things.

What AI is, in one slide or none

A writing assistant that predicts plausible text. It is fluent and confident, and it can be wrong. It does not know your clients, your protocols or what is true today. Show one live example of an invented fact, such as asking for a reference for a nutrition claim and then checking whether the citation exists. Citations produced by AI are sometimes fabricated, so any reference must be opened and read.

Accounts and settings

Choose one approved tool and account type. ChatGPT Business starts at two users and Claude Team at two seats, and Google Workspace plans now include Gemini in Gmail, so a practice already on Workspace may have an assistant built in. On free and personal plans, chats can be used for model training unless each user switches the setting off; on these business plans, the vendors say business content is not used for training by default. Whatever you choose, write down who owns the account, who can invite people and how a leaver is removed. Turn off any memory or chat-history feature you do not need, since stored context is another place client details can lodge.

The slip-reporting promise

Say aloud: "If you paste something you should not, tell me the same day. You will not be in trouble for telling me. You may be in trouble for hiding it." Put the same sentence in the policy. It is the single most useful thing in the whole programme, because owners only fix what they hear about.

Week 2: client emails

Client emails are the best practice ground because every dietitian writes dozens a week and they are easy to review. The training trick is to use invented clients. Give the team a fictional email and a shared prompt.

Fictional client email: "Hi, I tried the swaps but I'm still starving by
3pm and I've had two takeaways this week. Is it all pointless?"

Prompt: Draft a reply for a dietitian, 110 words, warm and practical.
Acknowledge the difficulty without judgement; give two general ideas for
an afternoon snack (no calorie numbers, no diagnosis, no supplement
advice); invite them to bring it to our next appointment on [date].
Sign off as [dietitian first name]. British English.

A typical draft is kind and sensible, but may drift into advice the practice would not give by email, such as suggesting a specific protein target. Teach the team to run a three-point check on every draft: Is there clinical advice I did not choose to give? Is there a claim that needs a source? Does it sound like me? Then each person edits the draft until it does.

Before and after

Illustration. Before: "It is not pointless at all! Many people experience hunger. Try high-protein snacks such as 30g of cottage cheese." After the dietitian's edit: "It is not pointless, and what you describe is very common. Let's look at the gap between lunch and dinner together on [date], and in the meantime keep a note of when the hunger hits." The edit removed a specific number nobody had assessed and moved the substance into the appointment, where it belongs.

Real clients will not appear in these sessions. In practice, when a team member replies to a real client, the safest pattern is to write from a stored template and personalise by hand, and to paste nothing from the client's email. A way to build that template library from invented examples is shown in the consent and cancellation prompt for nutritionists.

Week 3: handouts and the source-checking habit

Handouts are where AI saves most time and where it does most harm if unchecked, because a handout is read, kept and shared. The session has one job: install the habit of tracing every factual claim to a source.

1. Each person asks the assistant for a one-page handout on a general topic, for example "eating well on a budget", with a reading level of about age 12.
2. They highlight every factual claim: a food said to lower something, a quantity, a link between a nutrient and an outcome.
3. For each highlight, they find a source they trust (a guideline, a national dietary standard, a peer-reviewed review) and note it in the margin.
4. Any claim without a source is cut or reworded to a general statement.

Count the result. When teams run this exercise, a first draft will often contain claims that are either too strong or have no source you would be happy to cite. Record the count on a flipchart. The number matters less than the lesson: this is a draft, not a reference. Add a footer to every handout: "Reviewed by [dietitian name] on [date]. General information, not personal advice."

A useful extra is the reading-level pass. Ask the assistant to rewrite the checked handout at a simpler level, then check that nothing was changed in meaning: "a little" sometimes becomes "none", and a hedge such as "may help" becomes "helps".

Week 4: templates the whole practice shares

Once the team can draft and check, build the shared library. The aim is a folder of templates that everyone uses, so the practice sounds like one practice. Good first six:

• New-client welcome and what to bring to the first appointment.
• Appointment reminder and cancellation wording.
• Follow-up after a first consultation (generic, with placeholders).
• Missed-appointment message.
• Referral-out message ("this is outside our scope").
• A monthly newsletter skeleton.

Assign one template per person. They draft with the assistant, edit to the practice voice, and the owner approves. Save each with a date, the approver and the next review date. When all six exist, a new starter can be productive in a day, which is the hidden benefit of the whole programme. For the wider list of jobs that suit AI in a dietitian's practice, see AI tools for dietitians: admin, content and follow-up by job.

Also add a standing prompt, saved in a shared Project or document, so everyone starts from the same instructions: en-GB spelling, no clinical advice, no calorie numbers, placeholders for names and dates, never invent a source. Whether you keep it in ChatGPT Projects, Claude Projects or a plain document matters less than that it lives in one place and has one owner.

Week 5: the privacy drill

Drills beat lectures. Print ten scenarios and ask each person to mark "fine" or "not fine", then discuss. Illustrative set:

1. Ask the assistant for a newsletter on winter vegetables. (Fine.)
2. Paste a client's food diary and ask for a summary. (Not fine.)
3. Paste your own appointment reminder wording to improve it. (Fine, if it has placeholders and no real names.)
4. Upload a photo of a lab report to ask what it means. (Not fine.)
5. Type "a client of mine, a 42-year-old with type 2 diabetes who lives near the park". (Not fine; this can identify a person in a small community.)
6. Ask for ten Instagram caption ideas about meal prep. (Fine.)
7. Forward a client's email to a personal Gmail so you can paste it at home. (Not fine, twice over.)
8. Ask the assistant to reword a polite late-payment reminder with [amount] and [name] as placeholders. (Fine.)
9. Use a free browser extension that summarises emails. (Not fine until checked: it may read the whole inbox.)
10. Use a transcription app to record a client session. (Not fine without a signed agreement, client consent and a decision by the owner.)

Then walk through the slip procedure. If something goes in by mistake: stop, delete the conversation if the tool allows, tell the owner the same day, note what was pasted (without repeating it), and decide whether the client or a regulator must be told. Deadlines can be short. Under HIPAA, affected individuals must be told without unreasonable delay and no later than 60 days after a breach is discovered; under the GDPR, a controller must usually notify the supervisory authority within 72 hours of becoming aware of a breach. Find out which rules apply to your practice before you need them. Whether an incident counts as a breach is a judgement for your compliance adviser, not for the training session.

Week 6: review, log and sign-off

The last session turns habits into process. Three pieces:

• A review step. Every AI-assisted client document gets a second look from a named person for the first month, then a random 1-in-5 spot check. Keep the check to the three questions from week 2 (advice I did not choose to give, a claim without a source, a voice that is not mine), so it takes about a minute per document.
• A log. One shared sheet with date, task, who, what went wrong or nearly did, and what changed. Review it monthly in the team meeting.
• A one-page policy. Approved tools and account, the data rule, the review step, the slip procedure, and a signature line. The longer how-to is in how to roll out an AI policy so staff follow it.

Have everyone sign, date and file it. Set a calendar reminder for 90 days: add what you learned, remove what no one uses.

A role-play to run in the last ten minutes

Pair people up. One plays a client who has just received an email that quoted another client's first name by mistake (the AI draft had carried a placeholder through). The other plays the dietitian who must respond within two minutes: apologise plainly, say what happened, say what you are doing about it, and ask the client what they would like. Swap roles. The point is to practise the human response, because the first reaction to a slip is usually the worst: defensiveness, or silence. Afterwards, write the three sentences everyone agreed worked best into the policy as the standard wording.

Keeping the habit alive after week 6

Training fades without a rhythm. Add one agenda line to your monthly team meeting: "AI: log, one win, one near-miss, one change to the rules." It takes ten minutes. Once a quarter, refresh the data rule with one new scenario from the log, rerun the privacy drill with a different set of ten, and check whether vendors have changed their terms, since a plan that did not train on your data in spring may have different defaults by autumn. New starters get Sessions 1 and 5 in their first fortnight and sign the policy before they touch an account. When the team asks for something new, such as an AI note-taker or a chatbot on the website, do not bolt it on: put it through the same questions (what data, which agreement, who checks) and, if it touches client information, treat it as a separate project.

How you know the training worked

Measure after the programme, not during it. After a month, check:

• Rule recall: ask each person to state the three-part data rule. Everyone should manage it.
• Log activity: a log with no entries after a month usually means people are not recording, not that nothing is happening.
• Spot-check results: read five AI-assisted client documents at random. Count unsourced claims, stray names and clinical advice you would not have given. Aim for none.
• Time saved: ask each person for a rough weekly figure and compare it with the list of repetitive tasks they wrote down in week 1. There is no standard number to hit; a modest saving with clean spot checks is a better result than a large one with errors.

If rule recall is weak, repeat Session 1 and Session 5. If spot checks find errors, tighten the shared prompt and review step, not the staff. If you want an outside view on which tasks to teach first in your practice, that is a good use of my AI implementation consultation.

Questions practice owners ask before training the team

Should every team member get AI training, including admin staff?

Yes, at least the first two sessions. Admin staff handle the inbox, the booking system and client messages, which is where most accidental data sharing happens. Dietitians need the full programme. A short version for part-time or student staff keeps the same data rule and drops the template-building sessions.

How long does the programme take?

Six weekly sessions of about 45 minutes plus roughly 30 minutes of practice between sessions. For a five-person team that is about 37 hours of paid time over the six weeks. Run it as a standing agenda item in an existing team meeting if you cannot find extra hours.

Do I need a paid AI plan to run the training?

Not for the first two sessions, which only need an account and fictional examples. For ongoing use pick one work account type, because business plans differ on training use and agreements. Most mainstream vendors list the terms on a public page; read it before you put any client information near the tool.

What if a team member already uses AI on their own account?

Ask, without blame, in session one. Many will. Agree a date by which work tasks move to the approved account, and treat any earlier slip as a learning case, not a disciplinary one. Staff who fear punishment stop telling you what they do, which is the worst outcome.

Further reads

• ChatGPT or Claude for Freelance Dietitians? Pick by Job and Data — Pick a chat assistant by job and data for a small practice.
• AI Tools for Nutritionist Practices: Client Check-Ins and Handouts — Client check-ins and handouts, tool by tool.
• Nutrition Newsletter Ideas: A Month of Emails Drafted With AI — A month of newsletter emails drafted with AI.
• Does a Five-Person Business Really Need an AI Policy? — Whether a small team really needs a written policy.
• How to Train Staff to Spot AI-Written Phishing Emails — A fifteen-minute add-on session on AI-written scams.
• A Simple AI Risk Register for Small Businesses (With Template) — Keep a running log of what training uncovered.
• Dietitian Software Compared: Which Practice Platforms Have Useful AI? — Practice Better, Nutrium, NutriAdmin and Kalix compared on the AI each really has, what it costs per practitioner, and where it saves time.
• AI Tools for Learning and Development: A Small-Company Shortlist — A job-by-job shortlist for whoever runs training in a small firm: skills matrix, training plan, onboarding and checks, using tools you may already have.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center, Anthropic privacy centre and Google Workspace Admin Help on business agreements for health data, checked November 2026; session plan and figures are the author's illustrations.