12 AI Automations for Your Business: Where to Start Concretely

Auto-replies to incoming emails

A freelance consultant spends part of her morning answering the same messages: "did you get it?", "do you have the PDF?", "still on for Tuesday?". Say 1h30 a day (a starting assumption: replace it with your real figure after a week of tracking). Over 21 working days that is 31.5 hours, about four and a half working days (of 7 hours) a month for one person.

Concrete tool: Gmail (or Outlook), Make.com and a model provider's API. Make has a free plan, and its limits change: check Make's pricing page.

Setup (estimate: 1 h):

1. Create a Make.com account and connect Gmail through OAuth.
2. Build a "New email matching filter" scenario with a simple filter, for instance a "Reply needed" label that you apply by hand at first.
3. Add a generation step with the prompt below, then a "Create draft Gmail" step.

You are [your first name], [your job]. Reply to the email below in 3 lines maximum, direct tone, sign [your first name]. If the email asks something you cannot answer, write "to confirm" instead of making it up.

Email: [email text]

Expected outcome: in the first week you read and approve drafts instead of writing everything. Once the drafts are stable for one type of message, a meeting confirmation for example, you can consider sending that one directly. For subtler replies, prepare a library of templates and send them yourself.

Automatic categorization of support tickets

An online shop that is growing gets its messages through support: refund, pre-sale question, bug, complaint, feature request. Every ticket gets a label, sometimes two, and whoever picks it up spends, say, 30 seconds to 2 minutes just working out where it goes (my assumption, time your own). Across 300 tickets a day, that is between 2.5 and 10 hours of triage a day.

Concrete tool: Zendesk, Intercom or Freshdesk (whichever you already have), Make and a small model as a classifier. The cost depends on the model and the volume: check your provider's pricing page before promising anything.

Setup (estimate: 2 h):

1. Pull your real tag list (the actual one, not the wiki version: do a CSV export).
2. Build a Make "New ticket" scenario that sends the ticket to the model with the prompt below, which forces a pick from a closed list.
3. Pipe the result back to your support tool's API to apply the tag.

Reply only with a tag from this list: [billing, bug, question, complaint, other]. Ticket: [ticket text]

Expected outcome: triage done before anyone opens the ticket. It will get some wrong: you fix those by hand, and you improve the list and the prompt by reading the errors. If the triage holds up, the team can win back most of those 2.5 to 10 daily hours of triage. Measure the accuracy on your own tickets before you rely on it.

Meeting recap generation

In a small web agency, nobody wants to write up the client call, and nobody reads the recap if it lands four days later. Between current practice (nothing, or three lines in a chat) and a structured recap in Notion right after the call, the difference shows immediately.

Concrete tool: a transcription tool (Otter.ai, Fireflies, or the built-in transcription in Teams or Google Meet if your plan includes it: check your plan's documentation), Make for the structuring, and Notion, Teams or Slack for storage. Transcription tool pricing changes: check their official pages.

Setup (estimate: 1 h):

1. Turn on automatic transcription in your video tool, or invite the transcription bot to the meeting.
2. Create a Make webhook that fires when the transcript is ready.
3. Use the prompt below, then send the result to Notion.

Summarize this meeting in 3 sections: Decisions, Actions (with owner and deadline), Open questions. Markdown format. If an owner or a deadline was not stated, write "not specified". Transcript: [transcript]

Expected outcome: a recap that exists where nobody was writing one, with actions that have an owner and a date. You read it before it goes to the client: an action assigned to the wrong person gets expensive. Also tell the participants before you record and transcribe a meeting.

Personalized quotes from a client brief

A freelance developer gets a brief by email, rephrases it, walks through the usual method, plugs in the rates and formats the whole thing. A good part of that work is copying, which is exactly what AI does well. How long does a quote take you today? Time the next three: that is your point of comparison.

Concrete tool: an intake form (Typeform or Tally), Make, a model to draft, and a PDF generator (PDFMonkey, DocRaptor, or a Google Docs template with the Docs API). Check each tool's pricing page for the subscription.

Setup (estimate: 3 h):

1. Build a form with the key questions: sector, size, need, estimated budget, urgency, constraints.
2. A Make webhook passes the answers into the prompt below.
3. Send the output into a PDF template or a Google Doc, which you review before it goes to the client.

Write a quote from this client brief: [form answers]. Sections: Context, Approach, Deliverables, Timeline, Price excl. VAT. Use only these rates: [your rates and their unit]. If information is missing, write "to confirm" instead of making it up.

Expected outcome: a structured first draft in a few minutes, which you correct instead of writing from scratch. What you gain is measured on your own quotes, not on an average.

Weekly Slack and Discord recaps

In a six-person agency, someone comes back from vacation to hundreds of unread Slack messages. Three or four important decisions are in there, buried between the #random memes and the #dev debugging threads. They will read a few, give up, and ask a colleague.

Concrete tool: the Slack or Discord API, Make with a weekly trigger, and a small model to summarize. The marginal cost depends on message volume and the model: check your provider's pricing page.

Setup (estimate: 1 h):

1. Connect your Slack workspace to Make with the official app.
2. Build a scenario triggered every Friday late afternoon that fetches the last seven days of the channels you pick.
3. Use the prompt below.

Summarize this week in 5 points maximum per channel: decisions made, blockers identified, topics to follow. Only cite what appears in the messages. Messages: [channel messages]

Expected outcome: a "Weekly recap" email in your inbox every Friday. Important decisions surface, as long as someone wrote them down in the channel: a summary cannot find what was never said in writing.

Smart CV screening for hiring

This section touches on law and ethics, so I treat it with more caution. An agency hiring an apprentice quickly gets a few dozen CVs. AI can help you sort them. It can also reproduce, at scale, the very biases you are trying to avoid.

Concrete tool: a Google Drive folder as input, Make, and a model for structured extraction. You extract facts, you do not have it score: the final judgment stays yours.

Setup (estimate: 3 h):

1. Create a "Incoming CVs" folder on Drive that Make watches.
2. On every new PDF, trigger an extraction with the prompt below.
3. Send the result to Airtable or Notion, with a "Worth a look / Maybe / Not for this role" column that you fill in by hand.

From this CV, extract: years of experience, mastered technologies, current role level, announced availability. Reply in strict JSON. Do not use the name, the photo, the age or the school. CV: [CV text]

Expected outcome: you read a short sheet instead of the full PDF, then you decide. Never ask the model to score on criteria linked to discrimination (name, photo, school, age). The GDPR restricts decisions made solely by automated processing, in its Article 22: keep a human in the loop and, if in doubt, get legal advice.

E-commerce product description generation

An e-commerce seller with hundreds of references has hundreds of descriptions that read the same ("quality product, fast delivery"). They give no reason to buy. And Google's spam policies target content produced at scale with no value for the reader, whatever tool wrote it: the goal is useful listings, not numerous ones.

Concrete tool: Shopify or WooCommerce, Make, a model, and a prompt that enforces length and tone. The cost per listing depends on the model: test on ten products and look at your bill before you run the whole catalog.

Setup (estimate: 2 h):

1. Export a CSV of your products with title, category and key attributes (material, size, use).
2. Build a Make batch scenario: for each row, the prompt below.
3. Send the result back to your shop through the API (the description_html field for Shopify).

Write an 80-word description in 2 paragraphs, [direct, warm or technical, depending on your brand] tone. Include 1 customer benefit and 1 technical detail taken only from the attributes provided. Attributes: [product attributes]

Expected outcome: descriptions that differ from one product to the next and contain only what your attributes allow you to claim. Review a sample: an invented technical detail on a product page becomes a promise you have to keep.

Automatic replies to qualified contact forms

On a Saturday at 11pm, someone fills in a freelance graphic designer's contact form. She answers on Monday morning, and the prospect has asked for other quotes in the meantime. That delay is not inevitable: a quick acknowledgement with a proposed slot changes the first contact.

Concrete tool: a form (Typeform, Tally or your site's own), Make, a model, and your CRM if you have one.

Setup (estimate: 1 h):

1. Wire the form webhook into Make.
2. Use the prompt below.
3. Send the draft through Gmail or your emailing tool, and create the prospect record in your CRM (Pipedrive, HubSpot or a plain spreadsheet) in parallel.

Write a reply of 5 lines maximum: acknowledge the message below, offer a call slot within 48 hours with this link: [your booking link], sign [your first name]. If the message mentions urgency, adjust the tone. Message: [form message]

Expected outcome: the prospect gets a reply within minutes, even at night. I suggest leaving the replies as drafts for a week or two first, long enough to see what the model writes when the message is vague or off topic.

Automated competitive intelligence

A marketing consultant has a "to read" tab that keeps growing. Meanwhile a competitor shipped a feature, dropped a price or hired someone. Doing this watch by hand is easy to put off.

Concrete tool: Google Alerts, Make, a model for the digest, and a Slack channel or a recap email. Google Alerts sends you the results by email, which Make can read from your Gmail inbox.

Setup (estimate: 1 h):

1. Create a few Google Alerts on your main competitors and two or three keywords from your sector.
2. Have Make read the emails from these alerts, with a Gmail trigger filtered on the sender.
3. Set up a daily or weekly trigger with the prompt below.

Summarize these articles in 3 sections: Competitor moves, Sector trends, To investigate this week. Cite the link of each article. Articles: [articles for the period]

Expected outcome: a short, regular read instead of a tab that piles up. The value shows over several months, when you reread your digests and spot what kept coming back before it became noise.

Automated client onboarding

You sign a new client in a small web agency. Now you have to create the Slack channel, invite them to Notion, schedule the kickoff, send the brief template, create the shared Drive folder and set the client to "active" in the CRM. When you do all that by hand, you regularly forget a step or two.

Concrete tool: your CRM as the trigger (a "signed" status, for instance), Make, and the APIs of Slack, Notion, Google Drive and your booking tool.

Setup (estimate: half a day, because each connection is tested separately):

1. Define a clear status in your CRM ("Contract signed") that fires the scenario.
2. In Make, chain: create the Slack channel named after the client, add the internal members and the external guest, copy the Notion project template, create the Drive folder with the right permissions.
3. Send a welcome email with the links and a booking link for the kickoff.

Expected outcome: no forgotten step, and a first message to the client right after signature. Here the language model matters little: most of the work is plumbing between tools. Test with a fake client before the first real one, especially for the shared folder's permissions.

Dynamic pricing for recurring services

A freelancer on a monthly retainer often has the same rate as months ago. Inflation may have eaten part of the margin, costs may have gone up, and they have not raised the price because it is uncomfortable. The title says "dynamic", but this is a price review: AI will not negotiate for you, it helps you prepare the conversation and pick who to talk to first. If your rates move, the neighboring topic is unpaid invoices: what you can delegate when chasing them.

Concrete tool: your billing tool (Pennylane, Stripe Billing or an Excel export), Make, a model, and Notion or Airtable for tracking.

Setup (estimate: 2 h):

1. Export your list of recurring clients: current rate, date of the last review, tenure, estimated margin.
2. Use the prompt below.
3. Generate a note per client: justification for the new rate, talking points, an alternative if they refuse.

For each client in the list below, compute the gap between their current rate and [the target rate you want to reach]. Rank the 3 clients to contact first (largest amount at stake). Show your calculation. List: [clients, current rates, dates of last review]

Expected outcome: a ranked list and a preparation note per client, which you read and adapt. Check the arithmetic yourself on one client: the model can get math wrong, and you are the one announcing the rate. It is not the most glamorous automation, but it is the one that touches your income directly.

NPS and text feedback analysis

A coach sends an NPS survey to a group programme and gets 200 responses. The score is 42. Now what? The score alone tells you nothing actionable: what matters is in the comments. Reading and sorting them by hand takes several hours (my estimate), and it is the kind of work people put off.

Concrete tool: your survey tool (Typeform, Delighted or Google Forms), Make, a model as a classifier, and a Notion or Airtable summary table.

Setup (estimate: 2 h):

1. Create a Make webhook on every new response.
2. Use the prompt below.
3. Group everything in a table, with a "main negative themes over the last 30 days" view.

Classify this comment in 1 sentiment (positive, neutral or negative) and 2 themes maximum among [product, support, price, onboarding, other]. Reply in strict JSON. Comment: [comment text]

Expected outcome: you see why your detractors are detractors without rereading everything. Still read twenty raw comments before trusting the summary: what you are looking for is a theme you did not expect, and to spot it you need to have read some raw comments. To put the figures into shape every month, monthly reporting with ChatGPT in Excel and PowerPoint follows on well from this table.