AI Secretary for Small Business: Diary, Inbox and Calls Covered

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Secretary for Small Business: Diary, Inbox and Calls Covered.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Secretary for Small Business: Diary, Inbox and Calls Covered.

An AI secretary can sort your diary requests, draft replies to routine email, and answer or screen phone calls, but it should draft rather than decide. Start with inbox drafts you approve yourself, add diary rules next, and put calls last, because a wrong reply on the phone is the hardest to undo.

There is no single product called an AI secretary that does all of this well for a small firm. In practice it is three separate set-ups: the assistant inside your email and calendar, a tool that screens or answers calls, and a set of rules you write about what each may do. Treating them as three small projects, each with its own limit and its own check, is what stops this turning into a lot of missed messages. Each gets its own section below, with the limits stated plainly.

Follow me on Instagram@sagnikteaches

What a human secretary does that you are trying to cover

List the actual duties before choosing tools. For a typical owner-manager they split into three groups.

Connect on LinkedInSagnik Bhattacharya
  • Protecting time. Deciding which meetings to accept, finding slots, moving things when something urgent lands, guarding focus time.
  • Sorting and answering mail. Separating what needs the owner from what does not, replying to routine questions, chasing, filing.
  • Picking up the phone. Taking messages, answering common questions, booking, and knowing when to interrupt the owner.

An AI tool is good at the middle layer of each of these: sorting, drafting, finding slots, taking a message accurately. It is weak at the judgement layer: whether this particular supplier is more important than the one in your diary, whether a short reply from a long-standing customer is a warning sign. The design principle that follows is simple. Give the tool the routine and keep the judgement. The rest of this tutorial is how to draw that line in practice.

Subscribe on YouTube@sagnikteaches

The worked example: a four-person garden centre with one owner

Take an illustrative garden centre with an owner, a shop supervisor and two part-timers. The owner gets about 60 emails a day, 15 of them from suppliers and trade customers, takes six to ten calls a day, and juggles supplier visits, staff cover and delivery slots. In an average week the owner loses roughly six hours to diary juggling, mail sorting and calls that interrupt other work. The target is to hand back about two of those six without losing a single supplier order. We will use this business as the main thread and add other examples as we go.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Part one: the diary

What to hand over

Diary work comes in three flavours: finding a time, deciding whether to accept, and protecting time you have already blocked. AI is reliable at the first, partly reliable at the second and should not own the third.

  1. Finding a time. Use the scheduling features of your calendar tool or a booking link, so a requester picks from slots you have approved. If you want an assistant to propose times in an email reply, give it your rules as text and it will draft something like "I can do Tuesday at 10:00 or Thursday at 14:30."
  2. Deciding whether to accept. Write down five rules the assistant can apply, for example: supplier meetings only on Tuesday and Thursday mornings, no meetings before 09:30, never more than three calls on a day, anything involving staff or money goes to the owner to decide, anything from a new contact gets a short qualifying question first.
  3. Protecting blocks. Block focus time yourself, as busy events, and tell the tool not to book over them. Do not rely on the tool's guess about what counts as protected.

For the product choices, whether AI can manage your calendar and book meetings goes through what the built-in assistants in the main suites do today. Booking tools that let customers pick their own appointment are a separate job from your own diary, so keep them out of these rules.

A prompt for diary triage

If you paste meeting requests into a chat assistant, give it your rules each time, or keep them in a saved Project so they are always present.

You are my diary assistant. Apply these rules to each request below.
RULES:
1. Supplier meetings only Tue or Thu, 09:30 to 12:00.
2. No meetings before 09:30. Maximum 3 meetings per day.
3. Anything about staff, money or contracts: do not accept; mark
   "OWNER DECIDES" and say why.
4. New contacts: draft one question asking the purpose and the
   expected length.
5. Never propose a time I have not listed as free below.

MY FREE SLOTS THIS WEEK: [paste]
REQUESTS: [paste each request]

For each request give: ACCEPT / DECLINE / ASK / OWNER DECIDES,
a one-line reason, and a draft reply of under 60 words.

An illustrative result for a request from a new fertiliser supplier asking for "30 minutes sometime next week": ASK, reason "new contact, purpose unclear", draft "Thanks for getting in touch. Could you tell me what you'd like to discuss and which products are involved? I'm free for a short call on Tuesday between 09:30 and 12:00." The owner reads it, changes nothing, sends it. Saved time, no risk.

Where diary automation goes wrong

Time zones and daylight-saving changes cause the most dramatic errors; if you work with contacts in other zones, add the zone to every proposed time. Recurring events cause the next most: an assistant that moves one occurrence and not the series, or the reverse. Test by moving a harmless recurring item and checking the whole series afterwards. Finally, double-booking after an automatic acceptance happens when two tools both write to the same calendar. Pick one tool as the place that accepts invitations.

Part two: the inbox

Three layers, in this order

First, sort. Ask the assistant to label each incoming message as "owner needed", "routine reply", "information only" or "ignore". Second, summarise: for the "owner needed" pile, ask for a one-line summary of what is asked and by when. Third, draft: for the "routine reply" pile, produce a draft in your voice using a saved set of standard answers.

Do not skip the order. Drafting first, before sorting, means the assistant writes replies to messages that should never have been answered by it. The 15 supplier and trade emails a day in the garden-centre example are the obvious place to start, since they follow patterns: delivery date queries, invoice copies, stock availability, price lists.

The standard answers sheet

Create a one-page document with the answers you give most often, in your own words, with the facts filled in: opening hours, delivery days, minimum order, how to request an invoice copy, who to call about a damaged delivery. The assistant must use only this sheet for facts. This is the same discipline as in giving AI your business context: the tool is only as accurate as the page you give it.

Draft replies to the emails below using ONLY the facts in the
STANDARD ANSWERS. If a question needs a fact that is not there,
write "NEEDS OWNER" and say what is missing. Tone: friendly, brief,
first name only, sign off "[your name]". Never agree a price,
discount, refund or delivery date that is not in the sheet.

STANDARD ANSWERS: [paste]
EMAILS: [paste]

An illustrative before and after. Before: a trade customer asks, "Any chance you can get the 20 L pots in by Friday and at the old price?" The assistant, with no rule, might reply that this is "no problem". After the "never agree a price or date" rule, it produces "NEEDS OWNER: asks for a Friday delivery and the old price; neither is in the sheet." The owner decides, and the draft is ready in seconds.

Which tool does this

If you use a major office suite, start with the assistant built into it. Both the main suites offer AI features in their mail clients on certain plans (Google Workspace's Business Starter, Standard and Plus plans list Gemini in Gmail, for instance), with names and prices that change, so check the current plan page before paying. If your plan has no such feature, a general chat assistant with pasted emails does the job, at the cost of copying and pasting. For firms that handle a lot of mail, AI email triage for professional firms goes into sorting rules in more depth.

What to keep manual in the inbox

  • Complaints, however mild. Read them yourself; the tone matters more than the wording.
  • Anything with a price, discount, refund, contract term or legal consequence.
  • Messages from your accountant, bank, insurer or landlord.
  • Replies to anyone upset, grieving or in a hurry about something that matters to them.
  • First contact with a prospective large customer, which sets the tone for the relationship.

Rolling it out safely

For the first two weeks, drafts only: nothing is sent without you. Keep a tally of drafts sent unchanged, drafts edited lightly, drafts rewritten, drafts that should never have been drafted. After 30 drafts in a category, if over 80 per cent are sent unchanged or with light edits, you can consider automatic sending for that category alone, such as "invoice copy requests". Keep complaints and pricing off the list permanently.

Part three: the calls

Three levels of call handling

Calls are where the most care is needed, so think in levels, and start at the lowest.

  1. Voicemail with transcription. Callers leave a message; an AI service transcribes and summarises it into your email or messages. Low risk, since the caller knows they are leaving a message.
  2. Screening. An automated assistant answers, asks who is calling and why, then either forwards, takes a message or books a call-back. Medium risk.
  3. Answering and resolving. An AI voice agent answers common questions and books appointments by itself. Highest risk, since a wrong answer is spoken to a customer in real time.

The garden centre in the example starts at level one and moves to level two for the supplier line only. It stays away from level three until the data show that the questions callers ask are predictable. For comparisons of what these services cost and where they fit, see AI receptionist or answering service for a booking-based business, which compares the two set-ups. Prices vary by provider and plan, so compare current pages against your own number of calls per month.

Setting up call screening

A screening set-up needs four decisions written down before you touch a tool:

  1. Who gets through immediately. A short list of numbers: your top ten suppliers, your accountant, the school, your family. These skip screening.
  2. What the assistant may say. Opening hours, address, a statement that you will call back, and nothing else. No prices, no stock, no promises.
  3. What information it collects. Name, number, business name, reason, how urgent, best time to call back.
  4. What counts as an emergency. Anything that should ring through at once, for example a delivery driver stuck at the gate, with a rule for how the assistant recognises it.

Test it yourself with ten calls from a different phone before real callers reach it: a normal enquiry, a supplier, an angry customer, an emergency, someone who speaks very quickly, someone with a strong accent, a call with background noise, a silent caller, a wrong number, and a caller who asks for a person straight away. Write down what the assistant did with each. The last test matters most: whenever a caller asks for a human, the tool must say it will pass the message to one and take the details, not argue.

What callers must be told

Say at the very start that the caller is speaking to an automated assistant, and offer a way to reach a person or leave a message. Rules on disclosing automation and on recording and transcribing calls differ by place, and they depend on where your callers are, so check what applies to you or ask a qualified adviser. A greeting along the lines of "You've reached the garden centre. I'm an automated assistant and I can take a message or pass it to the team. How can I help?" is honest and short.

When the call tool gets it wrong

Illustrative mistakes: a caller says "the Thursday delivery" and the transcript records "the first delivery"; a caller with a surname that is also a common word gets it transcribed as the word; a message summary leaves out the one number that mattered, such as an order reference. Fix the third by telling the tool to always capture reference numbers verbatim. For the first two, check the transcript rather than only the summary on anything important. What to do when an AI receptionist gets a booking wrong goes through the recovery steps if a booking is affected.

Putting the three parts together: a weekly routine

Once the pieces are running, the owner's week looks like this. Morning: read the sorted inbox summary, approve or edit routine drafts, deal with the "owner needed" pile. Midday: check the call summaries, return the calls that need you. End of day: look at tomorrow's diary, as modified by the assistant, to check no rule has been broken. Friday: spend 15 minutes on the tally. How many drafts went unchanged, how many calls needed a human, which messages did the sorting get wrong, and which rule needs changing?

The tally is the thing that keeps this honest. Without it, an AI secretary degrades slowly: a rule goes out of date, a new supplier is not on the pass list, and a message sits in the wrong pile for a week. A 15-minute review once a week catches these early.

A mistake that shows how the pieces interact

An illustrative week shows why the three parts need one set of rules. The inbox assistant drafts a reply to a supplier offering "Thursday at 10:00", because the diary rule says Thursday mornings are open. Meanwhile the call screener, which has no access to the diary, tells a trade customer the owner can ring back "on Thursday morning". Both promises are reasonable on their own, and both land on the same slot. The fix is a single page of shared rules that all three tools follow, plus one calendar as the only source of truth for free time. When you change a rule, change it in that page first and then in each tool. Review the page at the same Friday check-in as the tally, because a rule that lives in three places drifts in three directions.

Four measures that show the set-up is earning its keep

Use four measures after a month, with illustrative targets for the garden centre:

MeasureHow to count itTarget after a month
Time on diary, mail and callsNote start and end times for three weeks before and afterDown by about two hours a week
Missed or wrongly sorted messagesAsk staff and suppliers; scan the "ignore" pile weeklyNone that cost an order
Drafts sent with no or light editsTally on a sheet, per categoryAbove 80 per cent for routine categories
Calls that needed rescueCallers who rang back or complainedFewer than one a week

If the time saved is under an hour a week after a month, the set-up is probably more complicated than your volume deserves. Drop the weakest part, usually call handling, and keep the inbox drafts, which tend to repay fastest.

Other businesses, other limits

The same approach flexes. A pet shop owner might need the diary for grooming appointments and supplier visits, with the inbox set to separate customer enquiries from wholesale offers. A gift shop may find the inbox the main burden around seasonal peaks, with the assistant drafting delivery-cut-off answers from a standard sheet that changes each autumn. A sports club secretary handles fixtures, membership questions and parent emails, where tone matters; drafts stay manual for anything involving a child or a disciplinary matter. For regulated or sensitive work, the limits are tighter again; the sector version for doctors, an AI secretary for doctors, shows how much more care is needed when messages carry personal records.

The data you give an assistant matters in all of them. ChatGPT Business, Claude's Team plan and Google Workspace state that your content is not used to train models by default, but a personal plan such as ChatGPT Plus may use chats for training unless you switch that off in its data controls, so check the plan's data page before pasting customer details. When in doubt, paste less: a first name and a question is usually enough for a draft.

If you would rather work through your own week with someone, that is what my AI implementation consultation covers, and you can reach it through the contact page. The order to remember is the same whatever the business: inbox drafts first, diary rules second, calls last, and a weekly tally throughout.

Questions owners ask about an AI secretary

Is an AI secretary one product or several tools?

For a small business it is usually several, stitched together: the assistant built into your email and calendar for the diary and inbox, and a separate AI phone-answering service for calls. No single tool does all three jobs well, so plan the set-up as three small projects, each with its own check.

Can an AI secretary send emails without my approval?

It can if you allow it, but you should not start there. Begin with drafts that you read and send yourself. Move a category to automatic sending only after you have seen at least 30 drafts of that type and corrected almost none, and never for pricing, complaints or anything legal.

What does an AI secretary cost for a small business?

Inbox and diary help may already be in your suite: Google Workspace's business plans include Gemini in Gmail, for example. A business chat assistant seat costs about $20 to $25 per user per month. Phone-answering services are priced differently, by monthly plan or by minutes, so compare two or three current price pages against your real call volume before choosing.

What should I tell callers about an AI answering the phone?

Say plainly at the start of the call that it is an automated assistant and offer a route to a person. Some places have rules on disclosing automation and on recording calls, so check what applies where your callers are, and keep the greeting short and honest.

Further reads

Sources: Google Workspace pricing page (Gemini in Gmail), ChatGPT Business and Claude pricing pages, OpenAI help article on how data is used (checked 5 Nov 2026).

Want your diary, inbox and phones tidied properly?

On a 1:1 call we'll walk through a normal week of your messages and calls, pick the secretary jobs worth automating first, and check whether the tools you already pay for can do them.

Book a 1:1 call with me