Pick one low-risk job, such as sorting website enquiries, and build a four-part scenario in Make: a trigger (Gmail's Watch emails), an AI module from Make AI Toolkit (Categorize text), a router with filters, and output modules for Google Sheets and Slack. Test with Run once, then schedule it. A first build takes one to two hours.
Make is visual but literal. Beginners rarely get stuck on the AI part; they get stuck on mapping (which piece of data goes into which field) and on credits, because every scheduled check of a trigger costs one even when nothing new has arrived. Knowing those two things before you start saves most of the frustration of a first build.
Choose a first job that's safe to get wrong
Your first scenario will have mistakes in it. Choose a job where a mistake costs a few minutes, not a customer. Good first jobs share five traits: the output stays inside the business, it happens at least weekly, the input is text, a person can check the result at a glance, and a wrong result is easy to undo.
| Candidate | Good first build? | Why |
|---|---|---|
| Events company: sort website enquiries by event type and log the details | Yes | Internal output, frequent, easy to check against the email |
| IT support firm: combine overnight monitoring alert emails into one morning Slack post | Yes | Internal, repetitive, obvious when wrong |
| Translation agency: automatically email quotes to new enquiries | Not yet | Customer-facing, involves prices, hard to undo |
The third is a fine second or third project once you've added an approval step. If you're still deciding what to automate at all, choosing what to automate first ranks the usual candidates. The rest of this walk-through builds the first one: an illustrative events company that runs weddings, corporate events and private parties, and receives about 400 website enquiries a month as emails from its contact form.
Make's vocabulary in five minutes
- Scenario: one automation, drawn as a chain of circles on a canvas.
- Module: one circle; a single action in one app, such as "Watch emails" in Gmail or "Add a row" in Google Sheets.
- Trigger: the first module, which checks for new data. It runs on a schedule or, for some apps, instantly.
- Bundle: one item of data passing through, such as one email. Ten new emails make ten bundles.
- Operation and credit: each module run is an operation. Credits are what you pay; for most apps one operation is one credit. A trigger uses one operation per check, whatever it finds, and every later module runs once per bundle.
- Mapping: choosing which output from an earlier module fills a field in a later one.
- Router and filter: a router splits the flow into routes; filters decide which bundles go down each route. Routers use no credits, and a bundle stopped by a filter uses none.
Step 1: sketch the scenario on paper
Ten minutes on paper saves an hour on the canvas. Write down the trigger, what the AI must decide, where each result goes, and what happens to anything unexpected. The events company's sketch:
- Trigger: new email in Gmail with the label Website enquiries (the contact form already applies it).
- AI decision 1: category, one of wedding, corporate, private_party, supplier_or_spam.
- AI decision 2: extract event date, guest count, budget, venue status and phone number.
- Route A: supplier_or_spam: stop, do nothing.
- Route B: every genuine enquiry: add a row to the Enquiries sheet.
- Route C: corporate with 100 or more guests: also post to the #sales Slack channel.
- Unexpected: anything the AI can't read goes to the sheet with category blank, for a person to check.
Step 2: create the scenario and add the trigger
Sign up for Make, click Create a new scenario, then click the large plus in the middle of the canvas and search for Gmail. Choose Watch emails. Make asks you to create a connection to your Google account.
One detour to expect: if your business email is a personal @gmail.com account rather than Google Workspace, Make asks you to set up your own Google Cloud project and OAuth client before it can connect to restricted Google services such as Gmail, Drive and Sheets. Make's help centre walks through the five stages; allow 20 to 30 minutes the first time. Workspace accounts avoid most of this.
In the module settings, choose the Website enquiries label and a sensible maximum number of results per run. When Make asks where to start, choose a recent point rather than your whole history, or the first run will process months of old enquiries and burn credits doing it. Send yourself a test enquiry through the website form, then right-click the module and choose Run this module only, so a real email appears as a bundle you can map from later.
Step 3: add the AI modules
Click the half-circle on the right of the Gmail module to add the next module, search for Make AI Toolkit and choose Categorize text. For the connection, choose Make's AI provider (available on all plans; you pay in credits based on tokens) and the Small model tier to start. On a paid plan you could connect your own OpenAI or Anthropic key instead, which charges one credit per operation while the provider bills tokens.
Fill in the fields:
- Text to categorize: map the email's subject and text content from the Gmail module.
- Categories: add each with a name and a description that removes ambiguity: wedding ("ceremony, reception or wedding party"), corporate ("company event, conference, awards night, team day"), private_party ("birthday, anniversary, family celebration"), supplier_or_spam ("selling something to us, job applications, newsletters, anything not an event enquiry").
- Assign only one category: Yes.
- Return an explanation for the categorization?: Yes. You'll want to see why during testing, and you can switch it off later.
Add a second AI Toolkit module, Extract information from text, mapped to the same email text. For each item you give a name, a description and a type:
| Name | Description you give | Type |
|---|---|---|
| event_date | The event date exactly as written, or "not given" | Text |
| guest_count | Number of guests if stated as a number, otherwise empty | Number |
| budget | Budget exactly as written, including currency, or "not given" | Text |
| venue_status | "booked", "looking" or "not mentioned" | Text |
| phone | Phone number if given | Text |
Keep dates as text on purpose. An AI model asked for a date type will happily turn "the second Saturday in June" into a specific date, and it has no reliable sense of which year you mean. Text copied as written is less convenient and far safer. Run the two AI modules once on your test email. An illustrative result for an enquiry reading "Looking for a summer party for about 140 staff, second Saturday in June, budget around $25k, we haven't found a venue yet":
Categorize text
category: corporate
explanation: "summer party for about 140 staff" describes a
company event.
Extract information from text
event_date: second Saturday in June
guest_count: 140
budget: around $25k
venue_status: looking
phone: (empty)Step 4: split the flow with a router and filters
Add a Router after the extraction module. It costs nothing to run. Draw three routes from it and click the small wrench on each line to set a filter:
- Route A, "Ignore": category Equal to supplier_or_spam. Attach nothing to it, or a Gmail module that applies a label so you can glance at what was ignored.
- Route B, "Log": category Not equal to supplier_or_spam.
- Route C, "Big corporate": category Equal to corporate, and guest_count Greater than or equal to 100 (use the numeric operator).
Make runs routes in order, and one bundle can travel down several routes if it matches several filters, which is what you want here: a large corporate enquiry is logged and announced.
Step 5: send the results somewhere useful
On Route B, add Google Sheets, Add a row. Create the sheet first with headers in row 1, because Make reads them to show you the columns. Map each column:
| Sheet column | Mapped from |
|---|---|
| Received | Gmail: date |
| From | Gmail: sender email address |
| Category | Categorize text: category |
| Event date | Extract: event_date |
| Guests | Extract: guest_count |
| Budget | Extract: budget |
| Venue | Extract: venue_status |
| Why (AI) | Categorize text: explanation |
On Route C, add Slack, Create a message, choose the #sales channel and write the text with mapped fields, for example: "Large corporate enquiry: [guest_count] guests, [event_date], budget [budget], venue [venue_status]. From [sender]. Logged in the Enquiries sheet." Keep the Slack message short; people should open the email for detail.
Step 6: test with real examples, then schedule
Forward ten real past enquiries (with permission, or with personal details removed) to the Website enquiries label, covering each category plus a spam message and one vague enquiry. Click Run once. After the run, the white bubbles above each module show how many operations it used and, if you toggle credits on, how many credits. Click a bubble to inspect each bundle's input and output.
Check each row in the sheet against its email. In the events company's illustrative test, eight of ten were right first time. One enquiry for "a surprise 40th for my husband's firm" came back as corporate; a better private_party description ("personal celebration, even if colleagues attend") fixed it. One had its guest count left empty because the email said "around a hundred and twenty", which is exactly what the "if stated as a number" rule intends. The team decided that was acceptable: a person reads the email anyway.
Then set the schedule, using the schedule setting on the scenario toolbar (it reads "Every 15 minutes" until you change it). That is the default, and the options include At regular intervals, Once, Daily, Weekdays, Weekly, Monthly, Specified dates and On demand. This choice has a direct cost, which the next section shows. The events company chose Daily with six run times between 08:00 and 18:00, because nobody replies to enquiries at 3am anyway.
Step 7: switch on the safety settings
Before you switch the scenario on, open its settings:
- Store incomplete executions: turn it on. It's off by default. When a module fails, Make keeps the unfinished run so you can fix and resume it instead of losing the enquiry.
- Number of consecutive errors: in advanced settings; the default is 3. After that many failed runs in a row, Make stops scheduling the scenario and emails you.
- Notifications: make sure Make's error emails go to an address someone reads daily.
Make also has error handlers you can attach to a single module: Skip, Retry, Resume, Commit and Rollback. Make has renamed some of these, so older guides may call them Ignore and Break. You don't need them on day one, but the moment a module fails for a predictable reason, they're the fix. Stopping Zapier and Make automations breaking silently shows how to use them and how to catch the failures that raise no error at all.
What it costs to run
Here's the events company's monthly credit estimate at about 400 enquiries, of which roughly 80 are spam or suppliers and 50 are large corporate leads. AI module figures are approximate, because Make's AI provider bills tokens on top of the operation; on the Small tier, 18,080 input tokens or 2,260 output tokens make one credit, so a typical enquiry adds a small fraction of a credit. Check the coin counter after a test run for your real number.
| Item | Calculation | Credits |
|---|---|---|
| Trigger checks, every 15 minutes | 96 a day x 30 | about 2,880 |
| Trigger checks, Daily at 6 times (chosen) | 6 a day x 30 | 180 |
| Categorize text, every email | 400 x about 1.1 | about 440 |
| Extract information, every email | 400 x about 1.1 | about 440 |
| Router and filters | No charge | 0 |
| Google Sheets row, genuine enquiries | 320 x 1 | 320 |
| Slack message, big corporate | 50 x 1 | 50 |
| Total with the chosen schedule | about 1,430 |
Two lessons sit in that table. The schedule alone would have cost twice the free allowance if left at 15 minutes. And even with a sensible schedule, 400 enquiries a month needs a paid plan, which starts from about $9 a month. A cheap saving is available: move the extraction module after the router onto Route B, so spam isn't extracted, which saves about 90 credits. The free plan also caps AI Toolkit at 200,000 input tokens a week; two AI modules reading 400 emails a month comes close to that, which is another reason to filter spam before the AI where you can.
The first errors you'll probably hit
- Empty columns in the sheet. Usually a field mapped from the wrong module, or mapping done before the module had ever run, so Make had no sample data. Run each module once, then remap.
- A filter that never passes. Often a trailing space or capital letter in the category. Categorize text returns your category names as defined, which helps; if you used a free-text prompt instead, use a case-insensitive text operator in the filter.
- RuntimeError from the AI module. Make AI Toolkit returns this when token limits are reached, such as the free plan's weekly cap. Filter out long newsletters before the AI step.
- Scenario switched off overnight. An expired or revoked Google connection typically produces an account validation error, and for that error type Make disables scheduling immediately rather than after three tries. Reconnect, then check incomplete executions for anything missed.
- A huge first run. The trigger started too far back. Stop the scenario, delete the extra rows and set the start point again.
The events company after four weeks
Illustratively, the scenario processed 412 emails in its first month. The team checked every row for the first week and then 20 a week. Of the 412, 11 were miscategorised (under 3%), all between wedding and private party; none of the 49 large corporate enquiries was missed. Credit use came to about 1,390. The biggest change wasn't sorting time, which fell from roughly 30 minutes a day to 10, but speed: large corporate leads reached the sales manager's phone within a couple of hours of arriving, instead of whenever someone next worked through the inbox.
The team made one design change after two weeks: it added the Summarize text module for enquiries over 200 words, so long emails appear in the sheet as a two-line summary. Extra modules are cheap to add once the base scenario works. What's worth adding next, and what isn't, depends on how stable the underlying process is; mapping a process before you automate it helps you spot the steps that shouldn't be automated yet.
Where to go from your first scenario
Three natural next builds, in order of risk:
- Draft replies, saved as Gmail drafts. Add a Simple text prompt module that writes a first reply using the extracted details, and a Gmail module that creates a draft rather than sending. A person reviews and sends.
- Enrich the CRM. Replace or supplement the sheet with your CRM's create-contact module, so enquiries stop being retyped.
- Automatic sending, with a human check. Only once drafts are consistently good, and only with an approval step in front of the send. Adding human approval steps to AI automations shows the patterns that work in Make, where a native approval module isn't available on most plans.
If you find yourself building a lot, it's worth checking you're on the right platform for your volume and your team's skills; comparing Zapier, Make and n8n for AI automation covers pricing models, learning curves and where each one breaks down.
Before you build in Make
Do I need an OpenAI or Anthropic account to use AI in Make?
No. Make AI Toolkit can run on Make's own AI provider on every plan, including Free, and bills the usage in Make credits based on tokens. On paid plans you can instead connect your own OpenAI or Anthropic key, in which case the module uses one credit per operation and the provider bills you separately for tokens. Start with Make's provider; switch only if volume makes your own key cheaper.
Can I build this on Make's free plan?
You can build and test it, and a low-volume version may run on the free 1,000 credits a month. Two limits bite first: every scheduled trigger check uses a credit even when nothing has arrived, and AI Toolkit on the free plan has a weekly cap of 200,000 input tokens. For a few hundred emails a month, expect to move to a paid plan, which starts from about $9 a month.
What's the difference between AI Toolkit modules and Make AI Agents?
Toolkit modules each do one defined job you configure, such as categorise this text or extract these fields, in the order you place them. An AI agent is given a goal and a set of tools and decides for itself which to use. For a first automation, toolkit modules are easier to test and predict, because every run follows the same path you drew.
Further reads
- How to Add AI Steps to Zapier: Classify, Summarise, and Draft — The same three AI jobs built in Zapier, with task maths.
- How to Tell If a Process Is Ready to Automate With AI — Check the job is stable enough before you build.
- How to Stop Retyping Data Between Apps With AI Automation — Next step: stop retyping enquiry details into other apps.
- Automation Audit: Find the Zaps and Scenarios Nobody Owns — Keep an owner and notes for every scenario you add.
- AI vs Rule-Based Automation: Which Does Your Task Need? — When a plain filter beats an AI module.
- What Is n8n, and Should a Small Business Use It? — The self-hosted alternative, if you outgrow Make.
- Is It Worth Automating a Task You Only Do Once a Week? — The payback sum for weekly tasks, the factors beyond time that change the answer, and three worked cases: one yes, one partly, one no.
- How Driving Schools Use AI to Keep Instructor Diaries Full — Fill cancelled slots in an hour, replace pupils before they pass, cut dead travel between lessons and spot quiet weeks early, with the data each step needs.
- How to Automate Purchase Orders and Supplier Emails With AI — Turn rough purchase requests into approved POs, send them automatically, chase silent suppliers and let AI sort replies, without auto-accepting price changes.
- Power Automate for Small Businesses: When It Beats Zapier — A decision guide for Microsoft 365 businesses: what your licence covers, the premium-connector line, how licences really work and three costed scenarios.
- Fleet Admin for Small Businesses: Automate Services, Checks and Fuel — Run a small fleet's services, daily checks and fuel records from one register, with reminders, phone forms and AI that reads receipts and flags odd fuel use.
- 8 Worked AI Automations for a Service Business, With Costs — One illustrative testing laboratory builds eight automations, from enquiry logging to a Monday summary, with task counts, AI costs, setup time and what broke.
- What Does the ChatGPT API Cost for a Business Automation? — Token prices for OpenAI's current models, one email automation costed at 500, 5,000 and 50,000 runs a month, and how to cap the bill.
- How to Implement AI in a Small Business With No Tech Team — Running AI with no IT staff: who covers the technical jobs, which tools need no code, how to secure logins, and what to do when something breaks.
- No-Code AI Automation for Small Businesses: Where to Start — Pick a first job, choose between Zapier, Make and Power Automate, build a nursery's enquiry flow, and count tasks and credits before you pay.
- Can AI Sort and File Incoming Documents for Your Business? — When AI can sort and file your incoming paperwork, what it costs in the software you run, and the review pile that stops misfiles.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Make help pages on credits, operations, scheduling, error handling and custom Google OAuth clients; Make apps documentation for Make AI Toolkit. Checked September 2026.