The way to do this is to write the answer bank first and the machine second. You collect the same twenty questions that land in your inbox, your web form, your phone and your social messages every week, and you write one plain answer to each, in your own words, from facts you can answer for. Then you put that bank into a project or as standing context in the assistant, on a plan tier that does not train on what you paste, and the assistant drafts the reply to each incoming question from the bank. A person reads the draft and sends it, which is the safe default. Answering once is the whole method. The assistant only types it back out.
What the job is
Every service business runs on the same small set of questions. What do you charge? When are you open? How long does a job take? Do you do X? What should I bring? Can I change the date? The list differs by trade, but the shape does not, and the questions arrive four or five different ways: by email, through the web form, by phone, and through social messages. Count the replies you sent last week and see how many said the same thing.
Typing the words is cheap. The cost is in pulling them out again, checking the price against the rate card, and reworking the sentence so it sounds like you. Once the answer exists somewhere, that cost disappears for the life of the price.
Where the plain templates stand
Before you bring a machine in, note what you already have. Gmail has had templates all along: you turn them on under Settings, then save any draft as a template and insert it from the compose window, though Google notes the mobile app does not support them. Outlook splits the job in two: Mail Templates saves a whole message including recipients, subject and attachments, and the My Templates add-in stores short snippets like standard replies, both of which need a qualifying Microsoft 365 subscription. Quick Parts, Outlook's older snippet drawer, does the same thing at the phrase level.
Templates solve half of this job without any AI at all, and if you have not turned them on, that is where you start. But a template and a Quick Part both do one thing perfectly: they paste the same words again. Neither reads the email that arrived. The assistant can read the incoming message, find which of the twenty it is, and answer in that specific context. That is the half worth a machine.
Collect the real questions for two weeks
Set a two week window. Put the real questions into one list: from the inbox, the web form, the phone, the social messages. Group the lookalikes. What you want at the end is the list of things you actually say again, in the order you say them. If a question appears once in two weeks, it is probably not on the list. The list should be short.
Then write the bank. One plain answer per question, in your own words: the price, the hours, the lead time, what the customer should bring, what you can and cannot do. One or two sentences each. This is the step that matters, and no assistant can do it for you. It cannot know which of those you will vouch for, and a draft from it is a draft you would have to rewrite. The bank is also the only part of this that you own outright. The tool you point it at will change, and the bank outlives it. We have written separately about which parts of your work go into a cloud assistant, in which parts of your work can go in ChatGPT.
The loop
Exhibit 1
The loop of answering repeat questions runs in five steps and never stops
The loop runs in five steps, and it does not end at step four.
- Collect the real questions for two weeks, from the inbox and the front desk.
- Write the answer bank, one plain answer per question, in your own words.
- Put the bank into a project or as standing context in the assistant, with instructions that say to answer only from it.
- The assistant drafts the reply to each incoming message from the bank, in your tone, and a person reads it and sends it.
- When a price or a policy changes, update the bank, and everything after that changes with it.
Step five is the one people forget, and it fails quietly. The bank goes stale the day the price changes. If the bank says one price and the rate card says another, the assistant will answer with the old one in full confidence.
Put the bank where the assistant can see it
Both of the big assistants have a place for this. Anthropic describes a Claude Project as a self contained workspace with its own chat history and knowledge base, where you upload the documents Claude should work from. The same page says you can define project instructions for each project, which is where the house rules for your replies live. OpenAI describes a ChatGPT Project the same way: a workspace where you group chats, upload reference files and add custom instructions so it remembers what matters and stays on topic. Upload the bank as a reference file, and write instructions that say: answer only from this document, in this tone, and say you do not know if the answer is not in it.
The plan tier matters more than the brand. On the consumer tiers the defaults run opposite ways. Anthropic uses consumer chats to improve the model only if you choose to allow it, under a control called Model Improvement in the privacy settings, and OpenAI's consumer plans train on your conversations unless you turn the switch off. On the business tiers both stop training on your inputs by default. And one detail on the OpenAI side: even with the switch off, clicking the thumbs up or thumbs down sends the whole conversation associated with that feedback into training, so on a consumer plan, do not click the thumbs.
The free tiers of both projects are usable for this, but small: Claude Projects are capped at five on the free plan, and the larger retrieval backed knowledge base is on the paid plans; ChatGPT Projects are on every plan with the free tier capped at five files. The paid tier of either assistant costs little next to an afternoon of retyping the same answer. And to be precise: the answer bank is your pricing and your policies, not customer data, so the training question is one of caution. We have compared the two assistants for back office work in Claude versus ChatGPT, and the business tier of either is the right home for this.
Exhibit 2
The answer bank holds one plain, checked answer per repeat question, numbered in the order they arrive
A bank for a small service firm usually holds the six questions above, in the order they arrive: the rate card you are willing to publish with a note on what varies, the hours and their exceptions, the lead time as the range you will actually hit, a yes, no or referred for each thing people ask whether you do, the short list of what to bring, and the rule on changing a date. The answers have to be facts you will stand behind in a meeting. If a price in the bank is one you would not say to the customer's face, it does not go in the bank.
Draft from the bank, send like a human
This is the daily part, and it is where most of the work is. In the morning, or whenever, you take the incoming messages, paste each one into the project, and the assistant drafts a reply from the bank in your tone. Then you read it and send it. For anything with a price, a date or a promise in it, that reading is not optional. The assistant's job is to save you typing, not to save you deciding.
This is the same pattern as the rest of the back office: the machine does the first draft and you do the judgment. We have described it for triaging the inbox and drafting replies, and the same logic applies to drafting a quote from your past quotes: the machine pulls the numbers, the human owns the number that matters most.
Before anything goes out, check four things:
- The answer matches the bank. If it does not, the draft was invented.
- No invented policy. If the bank does not have a policy on it, the reply says you will check. It does not make one up.
- No price the bank does not have. A price you have never written down does not get typed out by a machine.
- No promise on a date you have not actually committed to.
A draft that fails any of those four goes back to the human, and the fix usually goes back into the bank.
The bank is also your FAQ page
The bank doubles as a website page. Publish the same plain answers, in the same order, on a page called Frequently Asked Questions. Write it for readers and for the AI answers that will quote your site when someone asks their assistant about your business. That is the subject of how a small business shows up in AI answers. Do not write it for a Google rich result, because that is over: Google's changelog records the FAQ rich result as no longer appearing in Google Search starting May 7, 2026, and it removed the FAQPage documentation from Search Central on 15 June 2026. If you already have FAQ markup on the site, Google's position on unused structured data is that it does not cause problems and has no visible effects, so there is nothing to clean up in a hurry.
When the answer changes, you update one sentence in the bank, and the page, the drafts and the phone script follow it.
The live bot is a later decision
A chatbot on the website is a later decision. The hosted products bill in ways that reward reading the pricing pages closely. Their pages, checked in September 2026, print the following. Intercom's Fin, its AI agent, is $0.99 per outcome, and you can run Fin on a help desk you already have at that rate with no seat cost, though Intercom notes a minimum monthly commitment of, for example, 50 outcomes. Help Scout starts at $25 per user per month and sells its AI chatbot as an add-on at $0.75 per resolution, and you can put a monthly ceiling on it, at which point the chatbot switches itself off. Zendesk starts at $19 per agent per month paid yearly for its Support plan and $55 for the Suite plan that includes AI agents, and it bills AI work in automated resolutions without printing what one costs. Tidio is the cheap end: a free tier, a Starter plan printed at $24.17 a month for 100 billable conversations, and its Lyro AI agent from $32.50 a month for 50 AI conversations, with a billable conversation counted only when a human of yours replies.
The cost is metered, and the meter is defined by the vendor: an outcome, a resolution, an AI conversation, each with its own fine print about what counts, and none of the pages prints a total monthly number for a small firm, so get a written estimate for your volume before you sign. If a bot answers on your site, say so plainly. The FTC's position from its 2024 Operation AI Comply sweep is that using AI tools to trick, mislead or defraud people is illegal and that there is no AI exemption from the laws on the books. California's bot law gives you the standard to write to: a person using a bot is not liable under the section if the person discloses that it is a bot, and the disclosure has to be clear, conspicuous, and reasonably designed to inform the other person that it is a bot. The law is California's and it is narrow, aimed at bots used to mislead someone in a commercial transaction. Treat it as the standard worth meeting.
What the assistant is not for
The method has a wall, and it is worth writing down before you hit it. The assistant cannot handle questions that are really negotiations, complaints where the customer is buying an apology, or anything where the honest answer is it depends, let me look. Those three never come from the bank, because the right answer to them is a judgment plus a relationship, and the bank is a list of facts. When one of these lands, it goes to a person, and the loop stays clean.
Exhibit 3
The questions the assistant may draft and the ones that always go to a person split cleanly by whether the answer lives in the bank
- Goes in: price questions from the rate card, hours, lead times, what to bring, date changes, whether you do X.
- Stays out: negotiations, complaints, anything with it depends in the answer, and any message where the customer is angry.
If you want a firm version of this set up, we do it in our AI enablement work: the answer bank, the project, the review step and the FAQ page.