← Back to Blog

Make.com AI Agents: How to Build One That Does a Real Job, With Credit Costs and a Quote-Request Example

A Make scenario follows one path. The Run an agent module lets it choose. Here is when that is worth paying for, a quote-request agent a service business can build this week, and what each run costs in credits as of November 2026.

Tools13 min read•

A scenario follows a path. An agent picks one.

Say a landscaping company gets 40 quote requests a week through a web form and a shared inbox. A Make scenario can drop every one of them into a Google Sheet and ping the office. What it cannot do is notice that the third request is an existing customer asking about a job already on the calendar, the fifth is a vendor pitch, and the ninth wants a service you do not offer. A person reads each one and decides. That reading-and-deciding step is what Make's agent module is for.

Make calls the feature Make AI Agent (New), and the module you add to a scenario is called Run an agent. It has been in open beta since February 2, 2026, and Make's help center says both the features and the pricing may change. It is available on every plan, including Free, when you use Make's AI Provider, and paid plans can connect their own OpenAI or Anthropic key. This guide covers when the agent beats a normal scenario, how to build one, and what it costs in credits, with every figure taken from Make's own pages in November 2026.

Plain scenario or Run an agent: which one the job needs Question Plain scenario Run an agent Same input every time? Yes: form fields, new rows No: emails, free text, files Who picks the next step? You, with routers and filters The model, from your tools Cost per run 1 credit per module action 1 credit per operation + tokens Predictable? Fully. Same path every run No. Tool calls vary by input Best use Moving data between apps One judgment step in that flow

How Make shows up in the builds on this site

Nineteen of the 117 documented systems on this site list Make among their tools, and almost all of them use it as the plumbing between a voice platform and a calendar or CRM, not as the brain. A Plumbing AI Receptionist That Books Emergency Calls and Logs Everything to a CRM runs on Vapi for the call, Make for the booking logic and Airtable for the log. A $3,000 AI Voice Receptionist for Salons and Barbershops Built With Vapi and Make sells the same pattern as a package.

Two cheaper builds are closer to a pure Make job. A $497 Voice AI Agent for Local Businesses Built Entirely on Free Tools runs on Make's free plan and Claude's free tier. $4,200 in the First Month Using AI While Working a Full Time Job describes a gym owner automation in Make and ChatGPT that took four hours to build and saves the owner more than ten hours a week. All of those figures are self-reported by the builders.

The lesson from the whole set is worth stating plainly: a fixed scenario does most of the work. The agent module earns its cost on one step, the step where a person would otherwise have to read something and choose.

When the agent beats a plain scenario, and when it does not

Use Run an agent when the input varies, when the next action depends on what the input says, or when the job needs a lookup in one place before acting in another. Reading a free-text request, deciding whether it is a lead, a customer, or junk, and drafting a reply is a good fit. So is reading an invoice PDF and deciding which of three approval paths it belongs on.

Use a plain scenario when the input is a fixed form and the action is always the same. New row, send a text. New lead, create a contact. Routers and filters handle that in one credit per action with no model in the loop, and they do the same thing every time. Make's own introduction page says to keep agents away from sensitive data, high-stakes financial or strategic decisions, and anything with strict legal requirements. That is good advice. The agent should read and draft. A person, or a fixed rule, should send and spend.

A worked example: the quote-request agent for a six-crew landscaping company

Picture a landscaping company with six crews, 40 inbound quote requests a week, and an office manager who spends the first hour of every morning sorting them. The requests arrive from a website form, which Make can watch directly, and from a shared Gmail inbox, which Make can also watch. Both feed one scenario.

The scenario's trigger fires on each new form submission or email. The Run an agent module receives the text as Input. Its Instructions say: you sort inbound requests for a landscaping company; decide whether this is a new quote request, an existing customer, a vendor or spam, or something we do not do; look the sender up before deciding; never promise a price or a date. It has four tools. A Google Sheets search module finds the sender in the customer list. A Google Calendar search module checks whether a job for that address already exists. A Gmail module creates a draft reply, never a sent email. A Slack module posts a one-line summary to the office channel when the agent is unsure. Its Knowledge is one PDF: the company's service list, service area and the standard quote-visit wording.

A new request from a new address becomes a draft reply offering two estimate windows plus a Slack line that says new quote, draft ready. An email from an existing customer about a scheduled job becomes a Slack line that says existing customer, job on the 14th, needs a human. A mower vendor's pitch is labeled and ignored. The office manager opens Gmail, reads the drafts, edits two, sends all of them, and is done in ten minutes instead of an hour.

Each run uses roughly eight to twelve operations: the trigger, the agent step, two to three tool calls, the draft and the Slack post. Call it ten. Forty requests a week is about 400 credits a week, or around 1,700 a month, well inside the 10,000 credits that Make's Core plan includes. This is an illustration of the method, not a documented client.

Step by step: build it this week

Step 1: Write the job on paper first

Make's own guide starts with planning and so should you. Write one sentence for the goal, a list of the decisions the agent may make, a list of the actions it may take, and a list of things it must never do. If the list of actions includes sending, paying or deleting, move those to a human.

Step 2: Create the scenario and its trigger

In Make, click Create scenario and add the trigger that brings in the work: Gmail Watch emails, a Webhooks module behind your website form, or Google Sheets Watch changes. Set the schedule under the clock icon. On the Free plan the minimum interval between runs is 15 minutes; Core and above can schedule down to the minute. Then click the plus after the trigger, search for Make AI Agent (New) and add Run an agent.

Step 3: Pick the provider and model

In the Run an agent module, set Connection and Model. On the Free plan you must use Make's AI Provider. On a paid plan you can also connect your own OpenAI or Anthropic Claude account with an API key, which changes how you are billed; see the costs section. Start with a capable reasoning model while you are testing, because a weak model fails in ways that look like prompt problems, then step down to a cheaper one once the tools and instructions are stable.

Step 4: Write the instructions and map the input

Instructions are the agent's job description. Keep them to the role, the decision list from Step 1, the order of operations (look up before you decide), and the never-do list. Map the email body or form fields into Input. If the request can include attachments, map them to Input files; Make's guide says file inputs need Make's AI Provider, OpenAI, Anthropic Claude or Gemini and a model that supports files. Leave Response format as text unless a later module needs JSON.

Step 5: Add tools, and keep the list short

Hover the plus on the Run an agent module and click Add tool to attach a Make module as a one-step tool, such as a Google Sheets search or a Gmail create-draft. For anything with several steps, build it as its own scenario, turn on the On demand toggle in that scenario's toolbar, end it with a Return outputs module, and attach it with the Call a scenario module. Write a clear Description for each tool, because the model chooses tools by reading those descriptions. Make also supports MCP servers and sub-agents, and warns that many MCP tools drive up token use. Three to five tools is plenty for a first agent.

Step 6: Upload knowledge the agent must not guess at

Click Knowledge on the module and upload the documents that answer the questions your agent will get: the service list, the price sheet it may quote from, the policy wording. Make accepts JSON, TXT, CSV and PDF and stores them in a vector database for retrieval. Uploading costs tokens once per file, so keep the files tidy and re-upload only when something changes.

Step 7: Decide whether it needs memory

By default the agent forgets everything between runs. Make's module hint says that if you leave Conversation ID blank, the agent has no memory of previous interactions. For the quote sorter that is correct: every request is judged on its own. If you are building something conversational, such as a Slack assistant that follows a thread, set Conversation ID to the thread and set Maximum conversation history to the number of replies it should remember.

Step 8: Test in chat, then run once, then read the reasoning

Use Chat with Agent on the module to talk to it directly with a dozen real past requests, including the awkward ones. Then use Run once on the scenario. Open History, click Details, and read the Reasoning tab, which shows each tool it called and why. When it calls the wrong tool, fix the tool's Description before you touch the instructions. Use Disable tool to take one away and see whether the agent degrades gracefully or invents an answer.

Step 9: Put a person between the agent and anything irreversible

Make's own marketing says you can add manual approvals or stop the agent at specific points, and you should. Drafts instead of sends. A Slack post asking for a yes before a calendar event is created. A Google Sheet row the owner reviews before a quote goes out. Make's introduction page is explicit that agents are not for sensitive data or high-stakes decisions; design the scenario so the agent cannot take one.

Step 10: Watch credits for two weeks

Because the model decides how many tools to call, credit use varies from run to run. A Make community thread on agent pricing suggests running at least 50 sample inputs and looking at the spread before you trust a monthly estimate. Check your organization's credit usage after a week, and set the Extra credits auto-purchase option only once you know the number.

What this costs

Make's pricing page, read on November 21, 2026, lists the plans at 10,000 credits a month: Free at $0 with 1,000 credits a month, Core at $16 a month billed monthly or $12 a month billed annually, Pro at $28 or $21, and Teams at $51 or $38. Extra credits come in bundles of 1,000 or 10,000 at your plan's rate, and unused credits expire at the end of the term.

Agent runs are charged on top of the plan in credits. Make's help center lists the rules. With Make's AI Provider, Run an agent costs 1 credit per operation plus credits based on AI tokens, and each tool call costs 1 credit per operation plus tokens. With your own OpenAI or Anthropic connection on a paid plan, Make charges only the operation credits and your provider bills the tokens directly. How many tokens a credit buys depends on the model; Make's credits page lists, as examples, 18,080 input tokens per credit for GPT-5 nano and 452 for Claude Sonnet 5, and says the rates change as models change. Uploading a PDF to Knowledge costs 1 credit per operation plus 10 tokens per page plus embedding tokens.

For the landscaping example, Core at $16 a month covers the roughly 1,700 operation credits comfortably. Token costs on a short email with three tool calls are cents, not dollars, on a small model, whether Make or your own provider bills them. A realistic total is $16 to $20 a month, plus the afternoon it takes to build and a week of reading the reasoning tab.

When not to do this

Do not use the agent module for a job a filter can do, because you will pay tokens for a decision that never varies. Do not run it on the Free plan for anything time-sensitive, since the 15-minute polling interval means a quote request can sit for a quarter of an hour. If you expect thousands of runs a day, compare n8n, which this site's n8n vs Make vs Zapier guide covers, because per-execution pricing gets cheaper there at volume and you can self-host. And remember this is a beta: Make says features and pricing can change, so avoid building a client deliverable on it that you cannot rebuild elsewhere.

Frequently Asked Questions

What is a Make.com AI agent?▾

It is a module called Run an agent, from the Make AI Agent (New) app, that you place inside a scenario. You give it instructions, an input, uploaded knowledge files and a set of tools, which can be Make modules, other scenarios, MCP servers or sub-agents, and a language model decides which tools to call to finish the job.

Are Make AI agents free?▾

The feature is available on all plans, including Free, as long as you use Make's AI Provider. Each run still spends credits: 1 credit per operation plus credits based on tokens with Make's provider, or 1 credit per operation if you connect your own OpenAI or Anthropic key, which requires a paid plan.

How much does Make cost in 2026?▾

As of November 2026 Make's pricing page lists Free at $0 with 1,000 credits a month, and at 10,000 credits a month Core at $16 billed monthly or $12 billed annually, Pro at $28 or $21, and Teams at $51 or $38. Enterprise is custom.

Can a Make AI agent remember previous conversations?▾

Only if you set a Conversation ID in the module. Make's help says that with the field blank the agent has no memory of earlier interactions, and Maximum conversation history sets how many replies it keeps.

Can I use my own OpenAI or Claude key with Make agents?▾

Yes, on paid plans. You create a custom AI provider connection with your API key. Make then charges only operation credits and your provider bills the tokens.

Keep Reading

Explore Related Niches

Related Strategies

FeaturedLead GenLandscapingFreelance/Agency

Landscaping company's lead-to-booked rate jumped 42% after deploying an AI agent that responds to every inbound lead in under 60 seconds

Intermediate$50-200/moVery High ($10K+/mo)
View Strategy
E-commerceHome ServicesSaaS Product

An OpenClaw agent given $1,000 in startup capital generated $14,700 in revenue in three weeks by autonomously building and selling digital products.

AdvancedUnder $50/moHigh ($2K-10K/mo)
View Strategy
Dev ToolsMarketing AgenciesInternal Tool

Seven AI agents ran a company on a single VPS for a week. They pivoted the business, shipped an MVP with 158 passing tests, then forgot basic housekeeping.

AdvancedUnder $50/moLow (under $2K/mo)
View Strategy

Keep Exploring the Directory

Browse more case studies, compare tools, or dig into the full tool stack behind these workflows.

Want more strategies like this?

Get weekly AI agent case studies in your inbox.