What Make is
Make is a visual automation platform built around scenarios: triggers, modules, routes, filters, iterators, data transformations, and actions. AI Agents add a model-driven layer that can choose from approved modules, reusable scenarios, and MCP tools. The same platform can therefore run predictable integration logic and more flexible agentic loops.
That does not make the two modes interchangeable. A fixed scenario is usually better for invoicing rules, record synchronization, notifications, and other processes whose path can be defined. Use an Agent when variable context truly requires interpretation or tool selection. Keep side effects in explicit scenario steps whenever possible so they remain testable and observable.
Agent tools and minimum access
A Make Agent combines a model, instructions, context or knowledge, and tools. Tools can expose broad business capability. A calendar tool that returns full event details is substantially riskier than one that returns only free/busy slots. A CRM tool that can delete or bulk-update records has a different blast radius from a read-only lookup.
Make's current Agent best-practices guide says to assume people may access information shared with the Agent, including tool data and knowledge. It recommends minimizing access and mapping only needed data into a tool. The guide also states that Agents may ignore or misinterpret explicit guardrails. This is an important procurement signal: a system prompt is not an authorization layer.
Limit output tokens, execution steps, thread history, retries, and tool count. Long history raises both cost and unintended disclosure. Step limits help stop loops, but they can also leave a business transaction partially completed. Route incomplete or low-confidence results to a review queue rather than pretending they succeeded.
Human review and recovery
Make can place a human-review request on a scenario route. A useful pattern lets the Agent classify and draft, then pauses before an email, order, CRM write, or external publication. The reviewer should see the original input, proposed payload, supporting data, and exact target—not only a generated summary.
Review how state resumes after approval or rejection. A timed-out approval should fail closed. Retrying must not repeat earlier side effects. Add unique transaction keys, error handlers, and a manual repair path. Test revoked credentials, API rate limits, malformed tool output, prompt injection, duplicate webhooks, and a model that never reaches a stopping condition.
Pricing and credits
At the review date, Make's pricing page listed a Free plan and priced Core at $12, Pro at $21, and Teams at $38 per month for a 10,000-credit configuration. Enterprise used custom pricing. Core added unlimited active scenarios and API access; Pro added priority execution, custom variables, and full-text log search; Teams added team roles and shared templates. Currency, billing cycle, credit level, and regional taxes can change displayed amounts.
Credits are not the entire cost. The selected module, data movement, AI Agent execution, external model connection, code, and retries can affect consumption. Paid plans can use custom AI provider connections, which move some model cost and data terms to that provider. Estimate credits with a normal run and a failure-heavy run; include tests, loops, human review, and provider token charges.
Security, logs, and data routes
Make's security page reports SOC 2 Type II and SOC 3 audits, encrypted credentials and storage, TLS for transport, and cloud infrastructure across two AWS availability zones. It says log data is stored for 30 days by default, with extended storage available in Enterprise. Enterprise runs in a separately managed AWS environment from self-service cloud customers.
Execution logs can contain mapped records, prompts, tool responses, and error payloads. A shorter history reduces exposure but can weaken incident investigation. Decide which fields may be logged, who can search runs, and where exports go. Disconnecting an app does not remove records already written into other systems.
Make's privacy notice says Google Workspace API data is not used to develop, improve, or train generalized AI or machine-learning models. That statement is specific and should not be expanded into an unsupported claim about every customer-data category or every external provider. Review the selected model connection, subprocessors, DPA, region, and deletion process separately.
Who should choose Make?
Make is a good fit for teams that value a visual canvas and want to keep deterministic automation beside bounded Agent decisions. Its official safety guidance is unusually direct about limitations, which helps responsible builders but also makes clear that careful architecture is required.
Compare Zapier AI for broad app automation and separate Agents, n8n for self-hosting and code flexibility, and Gumloop for an AI-native credit model. The AI automation and agent tools category explains how to test permissions, approvals, traces, and cost.
Visit Make AI Agents
