AI Handbook for Notaries: A Written AI Use Policy and Safe Prompts

A notary can safely use prompts that draft or polish business messages from placeholders only: enquiry replies, fee explanations, appointment reminders and follow-ups, with no client names, document contents, ID details or file notes pasted in. Put those limits in a written policy, and keep AI away from the notarial act itself.

A list of prompts goes stale the week a vendor moves a setting, and it says nothing about what happens when an assistant pastes a client's passport number into a free chat app. A written policy answers who may use which tool, what never goes in and who checks what before anything is sent. The prompts then become one appendix.

Follow me on Instagram@sagnikteaches

The policy covers the business side only: the emails, quotes, reminders and follow-ups around an appointment. What you may do in the appointment itself comes from the body that commissions or regulates you, and this policy keeps AI outside it, whichever rules apply to you.

Connect on LinkedInSagnik Bhattacharya

The biggest practical risk is not a clumsy sentence but a tool setting. Individual ChatGPT plans may use your chats to improve models unless you switch that off, and on both ChatGPT and Claude individual plans a thumbs-up or thumbs-down click can make the whole conversation available for training. Claude's business plans treat a feedback click the same way unless the owner switches feedback off. The policy therefore names approved tools and settings, not only wording.

Subscribe on YouTube@sagnikteaches

Where the notarial act ends and admin begins

The first job of the policy is to draw the line in a form an assistant can apply on a busy Friday. Ten tasks cover most of the grey areas.

In some places notaries are barred from advising which document a client should sign, and in others they may be qualified to. The policy should say AI never does it either way, because an assistant has no way to know which rules apply to you, and a confident wrong answer is worse than no answer.

Seven clauses a notary's AI use policy needs

Write these in plain sentences, sign and date the page, and give a copy to anyone who touches client email. The full text is in the block after the list.

1. Scope and the line it draws. Business writing and admin only. Everything in the table above marked not allowed is written into the clause, so nobody has to remember it.
2. Approved tools and accounts. Name each tool, the plan and the account holder. For a practice with an assistant, a business plan beats two personal logins, because the owner controls the settings. Ban free apps and browser extensions that read the page or the inbox.
3. Three data classes. Green is public wording such as your fee schedule and website text. Amber is your own business information with no client identity, such as monthly invoice totals. Red is never entered into any AI tool. Staff only have to know the red list by heart.
4. Allowed tasks. A short list, with a default of no: anything not on it needs the owner's written approval. A default of no keeps the policy to a page.
5. Review rules. A person reads every AI draft before it leaves the practice and checks names, dates, fees and any sentence that sounds like a rule or a promise. Quote totals are worked out by the practice, never by the assistant.
6. Records and disclosure. A simple log of date, task, tool and reviewer, never content, plus one line telling clients how you use AI.
7. Incidents and review. If red data goes into a tool, the person tells the owner the same day. The owner reviews the policy every six months and whenever a tool or its terms change.

AI USE POLICY - [Practice name]
Owner: [name]. Version [1.0], dated [date]. Next review: [date].

1. Scope. AI tools are used for business writing and admin only (emails,
quote wording, reminders, follow-ups, website text). They are never used
for any part of a notarial act: identity checks, reading or judging a
document, certificates or notarial wording, advice on which document to
sign or what it does, or the notarial record.

2. Approved tools. [Tool, plan, account holder]. Nothing else, including
free apps, browser extensions and personal accounts. Model-training and
feedback settings are switched off and the owner checks them every quarter.

3. Data. Never enter: client names or contact details; identity documents
or numbers; document contents or scans; notarial record entries; witness
or signatory details; payment or bank details; passwords. Use
placeholders such as [client] and [fee].

4. Allowed tasks. Only those on the task list attached to this policy.
New tasks need the owner's written approval.

5. Review. A person reads every AI draft before it is sent and checks
names, dates, fees and any sentence that states a rule or promises an
outcome. Totals are worked out by us, not by the tool.

6. Records. We log date, task, tool and reviewer, never content. Clients
are told: "We use AI tools to help draft routine messages. A person
reviews every one, and your documents and identity details are never
entered into them."

7. Incidents. If restricted data is entered by mistake, tell the owner the
same day and do not click thumbs-up or thumbs-down on that conversation.
The owner follows the incident steps attached to this policy and
decides who must be told.

Signed: [owner]     Date: [date]

Attach two pages to it: the ten-task table above, which sets out what is allowed and what is not, and the incident steps near the end of this page. Those are the parts staff actually look at.

What never goes into an AI tool, and the stand-in to use

Staff follow a red list when it comes with a replacement for each item. Without one, they paste the real thing because the task has to get done.

A quick test before pasting: could someone who found this text identify a person, a document or a transaction? If yes, it is red. How to anonymise client data before you paste it into AI goes through the edge cases, such as dates and unusual job titles that identify a person without a name.

Tool settings the policy should name

Vendors rename settings often, so the policy says "the training switch" and the owner records the date each quarter when they last looked at the screen. What the vendors' own pages said in November 2026:

The feedback rows matter most and are the easiest to miss. Someone sees a good draft, clicks the thumbs-up, and on an individual plan the entire conversation, including whatever was pasted, may go to the vendor for training. The policy line is blunt: no thumbs, ever, on a work conversation. Does ChatGPT train on client data? Business versus free plans goes deeper on the plan differences.

The prompt set, by task

Every prompt takes placeholders only. You fill the real details in after the draft comes back, and the review rule applies to each one. Five cover most of a practice's week.

Replying to a new enquiry

Write a reply to a new enquiry for my notary practice. Courteous British
English, under 140 words.
Facts you may use: the service is [service type, e.g. "witnessing and
notarising a signature on a declaration"]; appointments are
available [in the office / at the client's address / by video, as I
offer]; the fee is [fee from my schedule] plus [travel fee]; the next
free slot is [date and time].
Do not say what documents the client needs, whether a document will be
accepted, or what it means legally. If the enquiry asks for any of
those, say I will cover it by phone and offer a call.
End with one question only: which date suits them.

An illustrative first draft that comes back:

Hello [first name],
Thank you for your enquiry. I can notarise your signature on a declaration. Appointments are available at my office or at your address, and the fee is $[fee] plus $[travel] for a visit. My next free slot is [date and time].
Please bring two forms of photo ID.
I would rather go through what you need, and any questions about the document, by phone so I can answer properly. Would you like me to call you, or shall I hold the slot? Which date suits you best?
Kind regards, [your name]

Two things to fix before sending. The ID line is invented: the prompt supplied no ID rule, the right list depends on whoever regulates you, and a wrong list wastes a client's trip. Delete it and attach your own approved wording. And the draft asks three questions where the prompt asked for one, so keep "Which date suits you best?" and cut the rest.

Turning a fee schedule into a quote

Write a quote email using only the figures below. Do not calculate,
round, add or convert anything, and do not mention tax or the document.
- Notarial fee per signature: [amount from my schedule] x [number]
- Travel: [amount]
- Evening or weekend surcharge: [amount, or none]
- Total: [my total, worked out by me]
Explain in one short sentence what each line covers. State that the
quote is valid for [14] days. Under 120 words.

Say the schedule gives three signatures at $15.00, travel of $25.00 and an evening surcharge of $20.00. You work out the total yourself, 3 × $15.00 = $45.00, then $45.00 + $25.00 + $20.00 = $90.00, and give the assistant that total. The draft comes back as a tidy list: notarial fee $45.00, travel $25.00, evening appointment $20.00, total $90.00, valid for 14 days. Check it against your schedule line by line. The usual slip is an extra sentence such as "All prices include tax", which states your tax position without your having supplied it. Delete or replace it with your own wording.

Following up the day after

Write a short follow-up email for the day after an appointment.
Facts: the appointment was completed; the client's next step is [e.g.
"send the notarised document to the receiving party"]; invoice [number]
for [amount] is attached and due in [number] days; review link [link].
Thank them, mention the invoice and the review link once each, and keep
it under 90 words. Mention nothing about the document's contents, its
effect or any client detail beyond the first name, which I will add.

The line to watch for in the output is anything about the document's effect, such as "your document is now legally binding". The prompt forbade it and an assistant will sometimes offer it anyway as a friendly flourish. Strike it. Also check that the review link asks for feedback on your service and nothing about what the appointment was for, since public review pages are searchable by anyone.

Chasing an unpaid invoice in three tones

Draft three versions of an overdue-invoice reminder, to send 7, 21 and 35
days after the due date: friendly, firm, final. Facts: invoice [number],
amount [amount], originally due [date], payment methods [list I supply].
Each under 70 words. No threats, no late fees, no mention of legal action
and no mention of what the appointment was for. In the final version say
I will phone before taking any further step.

The "no late fees" instruction exists because assistants like to add them. A first draft of the firm version might read: "Invoice [number] for $[amount] was due on [date] and is now 21 days overdue. Please arrange payment using the details below, or reply if there is a problem so we can sort it out." That is usable. If it adds "a late fee of 5% will apply", you have promised a charge your terms may not support, so delete it. Escalation beyond the phone call is the owner's decision, never the tool's.

Declining to advise, politely

A prospective client has asked which document they should sign and what
it will do. I cannot advise on that. Write a reply of under 100 words
that thanks them; says I confirm signatures and identity but cannot
advise on which document to use or its effect; suggests they ask the
organisation that requested it or a qualified adviser; and offers an
appointment once they have the document. Do not cite any law or rule.

A typical output ends: "Once you have the document, I would be glad to arrange an appointment." Check the middle for a sentence like "it is against the law for me to advise you". Whether that is true depends on where you practise and who commissions you, so replace it with "that is outside what I do as a notary" unless your regulator publishes wording you can quote. This reply is the one to save as a template, because the question arrives weekly and a rushed human answer is where practices drift over the line.

Other tasks on the allowed list

For a wider list of safe jobs, six admin jobs a notary can hand to AI covers the day-to-day tasks, and this policy is what lets you say yes to each of them with confidence.

A two-person practice, before and after the policy

Take a notary and one part-time assistant, about 60 appointments a month. They answer 25 enquiries a week, send 12 quotes, follow up 15 appointments, send 15 reminders and chase 4 late invoices. Before the policy, each message is typed from scratch or copied from an old email, with fees looked up each time. After it, the assistant uses the five prompts above, and the "after" minutes include a one-minute read-through of every draft. All figures are illustrations.

Over a month that is about 15 hours (3.5 hours × 4.3 weeks), worth roughly $530 at an internal value of $35 an hour. Two ChatGPT Business Standard seats cost $40 a month billed annually or $50 monthly, so the tool is the small part of the bill. The real cost is the owner's setup time: about three hours to write and sign the policy, one hour to check settings and 30 minutes each quarter after that. The saving only holds if the review step stays in, which is why it sits inside the "after" figures and not on top of them.

Review rules that catch the usual failures

Print this list and keep it beside the screen where drafts are read.

1. Every figure against your schedule. Fees, travel charges and totals come from your documents, never from the draft. Recompute totals by hand.
2. Invented requirements. ID lists, deadlines, retention periods, what a document needs. If the prompt did not supply it, it is a guess.
3. Statements about legal effect. Words such as binding, valid, enforceable, accepted or recognised.
4. Promises. Availability, turnaround and outcomes you have not agreed to.
5. Leftovers. Unfilled [brackets], the wrong first name or an old date copied from a template.
6. Voice. Spelling, tone and sign-off in the style your clients already know.

A realistic miss: an assistant sends a reminder whose draft said "your notarised document will be recognised wherever you need it". Nobody read the line closely, and the client later asked why a receiving body had refused it. The sentence was a guess made by a tool that did not know the receiving body, and rule 3 would have caught it. The fix to the policy was one line: any sentence about acceptance or recognition is deleted, not softened.

When something restricted was pasted anyway

It will happen once, usually in a hurry. A policy that treats it as a sacking offence guarantees it is hidden next time, so write steps instead.

1. Tell the owner the same day. No blame, and no clicking thumbs-up or thumbs-down on that conversation.
2. Record the type of data (never the data itself), the time, the tool, the account and whether training and feedback were off.
3. Delete the conversation and any attachment in the tool, and check the history for related chats.
4. Ask the vendor through its privacy channel what deletion does for data it already holds. OpenAI offers a Privacy Portal for individual accounts, and Anthropic's privacy site explains how to contact it on privacy matters.
5. Ask your regulator or insurer whether the client, or anyone else, must be told. Do not guess in either direction.
6. Add what you learned to the policy, usually a new line in the red list or a changed setting.

Telling clients in one sentence

Put a single line in your terms and in your confirmation email footer: "We use AI tools to help draft routine messages. A person reviews every one, and your documents and identity details are never entered into them." It is accurate only if the policy is followed, so write it after the policy exists and not before. Ask your regulator whether it expects more, and keep the line next to your privacy wording where a client will look for it. For the broader version of this document in other professions, the AI acceptable use policy for a small professional firm shows how the same clauses carry over.

Questions notaries ask about using AI

Can a notary use ChatGPT at all?

For business messages, yes, under a written policy: draft emails, quote wording and reminders from placeholders, then review before sending. Keep it out of the notarial act and away from client identity, document contents and the notarial record. Check your regulator's guidance too, since some bodies publish their own.

Do I have to tell clients I use AI?

Rules differ by regulator and by what you put into the tool, so ask yours. A short line in your terms or confirmation email is cheap and honest: AI helps draft routine messages, a person reviews every one, and client documents and identity details are never entered.

Is a free AI account good enough for a notary practice?

Only with the training switch off, no client data and no thumbs-up or thumbs-down clicks on conversations. A business plan removes the training risk by default and lets the owner switch feedback off for everyone. Two ChatGPT Business Standard seats list at $40 a month billed annually.

What if my assistant pastes a client document into AI?

Treat it as an incident, not a telling-off: note what went in, when and into which tool, delete the conversation, check the training and feedback settings, and tell the owner the same day. Whether the client or your regulator must be told depends on your rules, so ask your regulator or insurer.

Can AI check whether a document is correctly signed?

No. Reading a document, judging whether it is complete and checking identity belong to the notarial work, and putting any of it through an AI tool would also break the never-paste rule. Use a paper checklist you control and keep the assistant for business writing only.

Further reads

• Mobile Notary Software With AI: Scheduling and Invoices Compared — Scheduling and invoicing software, with the AI features compared.
• Mobile Notary Marketing With AI: Profile, Posts and Review Replies — Profile text, posts and review replies without client detail.
• How to Roll Out an AI Policy So Staff Actually Follow It — Getting an assistant to follow the policy once it exists.
• Is Gemini Safe for Confidential Business Data in Workspace? — The Workspace settings behind the Gemini line in your policy.
• Does Microsoft 365 Copilot Keep Your Business Data Private? — What Microsoft's data promises cover and what they leave out.
• What to Check in an AI Tool's Privacy Policy and Terms — A way to read any vendor's privacy terms yourself.
• Process Server Software: Dispatch, Proof of Service and Where AI Fits — Nine jobs process server software must do, four tools checked against them, a 140-job cost sum and where AI can help with intake and drafts.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help article on how data is used to improve models and its Business pricing page; Anthropic privacy centre pages for consumer and commercial products; Microsoft enterprise data protection page; Google Workspace generative AI privacy hub, all read in November 2026.