QWE AI Academy One Make-powered setup at Stellantis &You UK already handles 20,000 AI voice calls a month and saves £200,000 a year. Not a demo. Production scale: a virtual receptionist wired into live customer data.
Solo consultants and tiny agencies drown in the same three piles – lead qualification, draft content, ticket triage. Juniors are expensive. Rent is worse. A Make.com AI agency is how you stand up a virtual bench that does more than chat: agents reason, pick tools, and execute inside the apps you already run.
Agency floor, not a single chatbot
Make AI Agents live on the same visual scenario canvas as everything else. You write a goal in plain language, attach tools (modules, other scenarios, MCP connections), and the model chooses the order. Fixed scenarios follow your flowchart. Plain ChatGPT stops at text. These agents push buttons across 3,000+ apps.
According to Make’s AI Agents page, every plan can run agents through Make’s own AI Provider. Paid plans also accept your OpenAI or Anthropic keys. The practical win is the Reasoning panel on the canvas – each decision is visible, so you audit misfires without treating the run as a black box.
Picture a small agency floor. One seat qualifies inbound leads. One drafts outreach. One writes CRM fields and pings Slack. Hand-offs happen through shared records or a Conversation ID, not magical group chat.
Pro tip: Ship one tightly scoped agent and one boring trigger first. Grow the bench only after that agent returns stable structured output. A fat tool list on day one makes the model thrash.
First hire: lead-gen agent in under 30 minutes
Concrete starter – not inventory checkers, not generic support triage. Trigger on a new Tally form or Airtable record. The agent researches the company, scores fit, drafts a personalized email, logs the row.
- Create a scenario. Add the trigger (Watch Records in Airtable or a form webhook).
- Drop in Make AI Agent (New) → Run an agent.
- Connect Make’s AI Provider (works on free/all plans per pricing) or your own key. Start with a strong reasoning model; swap cheaper once the brief is stable.
- Write instructions like a job brief: “You are a B2B lead qualifier. Score 1-10 on fit using company size and industry. Only use provided tools. Return JSON with score, reason, draft email.”
- Map form fields into Input. Optional knowledge files: ICP doc, tone guide.
- Attach tools: a research scenario (web or Perplexity), a Gmail Drafts creator, an Airtable updater. Every tool scenario ends with Return output and carries a crystal-clear name + description – the agent literally reads those labels to choose.
- Conversation ID if you need thread memory across runs; leave blank for stateless. Default step timeout is 300 seconds.
- Run once. Read the Reasoning panel and the execution history before you trust it with a full form queue.
Official help follows the same planning → scenario → configure → tools → knowledge → test path in Create your first AI agent.
Multi-agent hand-offs and the credit math that bites
Clone what works. High scores route to a closer agent that books calendar slots. Low scores hit a nurture agent that queues a short email sequence. Shared Airtable rows or Data stores carry state; Conversation ID + max history keeps thread context without pasting the full window every run. For anything downstream, set Response format to a Data structure so you get JSON, not polite paragraphs.
| Plan (annual billing) | Price (as of late 2025 / early 2026) | Agency-relevant limits |
|---|---|---|
| Free | $0 | 1,000 credits/mo, 2 active scenarios, 15-min min interval, 5-min max execution |
| Core | from $9/mo annual | 10k credits base, unlimited scenarios, 1-min interval, 40-min max execution, API access |
| Higher paid tiers | see Make pricing | Priority execution, richer logs, roles, shared templates – check current page before you budget |
Billing unit is credits (operations renamed August 2025, 1:1). Ordinary modules still cost 1. AI does not. Make’s credit-usage docs spell it out: with Make’s provider, Run an agent = 1 credit + variable token credits; knowledge files add embedding and description tokens. Custom provider keys keep Make at a flat 1 credit per operation while the LLM vendor bills tokens on their side – so you can forecast Make spend when volume climbs. Extra credits have carried a 25% markup since the November 2025 update. Polling a watch module every five minutes can quietly chew a Core allotment before any agent thinks.
The catch is spike behavior. One long-reasoning run, a fat knowledge file, and three tool calls can cost several normal scenarios. Free-tier math collapses under real agency load. Prefer custom keys once you leave sandbox if predictability matters more than convenience.
Where agency builds break
Long chains drift. Vague tool names? Wrong scenario gets picked – or you loop until timeout. Default 300s step limit plus the paid 40-minute scenario ceiling (5 minutes on free) will cut a multi-hop hand-off mid-flight if a research tool drags. No cross-run memory unless you set Conversation ID or park state externally; blank ID means a fresh brain every trigger, which is how you get duplicate tool calls and lost thread context.
You’re still on the hook for guardrails. Approval steps on high-stakes sends. Rate limits. Human eyes on anything customer-facing. The Reasoning panel helps debug; it does not replace policy. Test ugly inputs – empty company names, duplicate forms, partial CRM rows – before you call the bench “hired.”
FAQ
Do I need coding skills for a Make.com AI agency?
No. Visual builder only. If you can finish a normal scenario, you can ship an agent.
How much does heavy AI agent usage actually cost?
Provider choice dominates. Make-hosted path: base 1 credit + token credits that swing with model and prompt length. Custom OpenAI/Anthropic key: Make stays at 1 credit per op; tokens bill outside. Knowledge file size moves the needle hard – trim docs before you attach them. Run your real prompts once, open the dashboard credit table, then decide. Extra credits cost 25% more (as of Nov 2025).
Can multiple agents talk to each other like a real agency team?
Indirectly. Pass outputs as the next input, share Airtable/Data store rows, or reuse a Conversation ID for threaded context. Scenarios can also surface as MCP tools so something outside Make calls into your stack. There is no one-click “agent group chat.” Map qualifier → closer on paper first; orchestration gets obvious after two live hand-offs, not before.
Open Make today. One lead-qualifier. Two tools. Five real form submissions. Watch the Reasoning panel. That is hire number one – small, measurable, and yours to fire if the brief is wrong.