Best AI agent builders for a small business

There’s no single best AI agent builder. The useful way in is to notice that “AI agent builder” covers two different kinds of product. There are assistant-style builders aimed at people who never want to touch code: Lindy (from $49.99 a month), Relay (free, then from $19) and Gumloop (free, then from $37). And there are mature automation platforms that have bolted an agent onto a workflow tool: Zapier (agents from a $33.33 add-on), Make (from $9) and n8n (free if you self-host). Which group you want depends far less on features than on who is going to build and look after the thing.

The two groups come at the problem from opposite ends, so lining all six up on price alone tells you nothing until you know which kind you need. This page sorts them that way, gives you the real current price and the metering model for each (the part that actually decides your bill), plus two things to settle before you spend anything: for a lot of jobs you don’t need a separate subscription at all, and a fixed, same-every-time task doesn’t need an agent in the first place. Whichever you land on, start with one narrow job, not a swarm of agents. One agent doing one thing well is how this goes right.

The line-up, and what each tool really is

Sort them by who builds and maintains the agent, because an assistant you set up by texting it and a self-hosted node graph are different purchases, and ranking them on price alone hides that.

Tool	What it really is	Entry price and model	Best for
Lindy	Assistant-style agent builder, no code	7-day trial, then from $49.99/mo (no free tier)	Non-technical operators who want it running today
Relay	Automation with a human-approval step built in	Free; paid from $19/mo	Jobs that need a human yes partway through
Gumloop	Visual no-code agent canvas	Free (5,000 credits/mo); paid from $37/mo	A friendly first build
Zapier	8,000-app automation, agents bolted on	Free; Zapier Agents from $33.33/mo add-on	Widest integrations, you already use Zapier
Make	Cheaper visual automation, AI Agents built in	Free; paid from $9/mo	The most power per dollar
n8n	Source-available, native AI agent node	Free self-hosted; cloud from €24/mo	Technical teams who want control

Those prices are US dollars before tax (n8n bills in euros) and they shift fast here, so read the table as the shape of the market, not a quote. If you’re in Australia, you pay that USD figure in AUD plus your card’s currency margin, so budget closer to one and a half times the sticker, and claim the 10 percent GST back if you’re registered and the vendor gives you a compliant invoice. And hold one thing in mind as you read on: every tool here meters usage somehow, by credits, tasks, activities or steps, so the headline price is the floor, and an agent can run the rest up faster than a fixed automation can. More on that below.

Two kinds of “AI agent builder”, and which you need

The single most useful thing to understand is that the six tools split cleanly into two camps, and you choose your camp before you choose your tool. Get this straight and the category falls into order.

The first camp is the agent-first builders: Lindy, Relay and Gumloop. These were designed from scratch around an AI that decides, for an operator who is never going to open a developer console. You describe the job, connect your apps, and you have an agent. They trade some power and some control for being genuinely usable by a non-technical person on day one.

The second camp is the automation platforms with agents added: Zapier, Make and n8n. These are established workflow tools, the ones people have used for years to wire apps together, and each has bolted an AI agent feature onto a mature engine. You get deeper integrations, more control and (with Make and n8n) a lower price, at the cost of a steeper tool that was built for fixed automation first and agents second.

So the question isn’t “which is best”, it’s “who’s building this”. A non-technical owner who wants an agent handling their inbox by Friday is in the first camp. A business that already runs ten Zapier automations, or has someone technical who wants to self-host, is in the second. If you want the plain-English version of what an agent even is before you pick, our explainer on AI agents covers it without the jargon, and the agentic AI guide goes deeper on what they can and can’t do.

Lindy: the assistant-style agent for non-technical operators

Lindy is the one to start with if you want an agent running today and never want to see a build canvas. It works as a personal AI assistant you mostly talk to: it manages your inbox, drafts replies in your voice, takes and summarises meeting notes, schedules and reschedules, chases leads and updates your CRM, driven from chat or even text message. There’s nothing to wire together. You tell it the job and approve what it proposes, which is the right shape for someone who wants the outcome, not the tool.

The catch is the price and the lack of a free tier. Lindy runs a 7-day trial, then Plus at $49.99 a month, Pro at $99.99 (which adds computer use, so it can operate browser-based tools for you) and Max at $199.99 for heavier use across more inboxes. Usage is metered and scales up the tiers, and voice calling is billed separately by the minute. So Lindy is the most expensive way in here and the one with no free plan to linger on, which means you’re committing to spend inside a week. That’s the trade for the least setup of anything on this list. It’s also the tool we point to for outbound admin in our lead generation comparison, if sales follow-up is your sore spot. Pick Lindy when your constraint is time and setup, not budget.

Relay: automation with a human in the loop

Relay is the pick when you want an agent doing the work but a person approving the parts that matter. Its signature feature is the human-in-the-loop step: you build a flow and drop in points where it pauses for someone to review, edit or approve before it carries on. That maps exactly onto the one rule worth holding with any agent, which is to keep a human on anything that spends money or goes public. The builder itself is chat-based and runs on Claude, so you describe what you want and it assembles an editable visual workflow.

On price it’s gentle. Relay’s free plan gives one user 200 steps and 500 AI credits a month with every feature included, Professional is $19 a month on annual billing (closer to $38 month to month) for 750 steps, and Team is $59 for ten users and 1,500 steps. Usefully, it doesn’t charge you for triggers or for the plumbing between steps, only for the steps that actually do work, so the meter is more honest than most. The limit to watch is those step counts: a busy agent checking and acting through the day will eat them, and you’ll feel the jump from free to paid sooner than the headline suggests. Choose Relay when the jobs you want to automate genuinely need a human checkpoint, which is more of them than you’d think.

Gumloop: the friendliest first build

Gumloop is the easiest place to learn what building an agent feels like, because it puts the whole thing on a visual canvas you drag nodes around on, with a generous free tier to experiment in. You’re connecting steps by hand more than with Lindy, but it’s built so a first-timer can get something real working without a tutorial, and the free plan gives you room to make mistakes. Bigger teams use it too: Gusto, Instacart and Shopify are on the customer list, so it scales past the learning phase.

The free plan is 5,000 credits a month with one seat and unlimited agents and flows, and Pro is $37 a month ($29.60 on annual billing) for 20,000-plus credits, unlimited seats and the ability to host your own connectors. Like everything credit-based, your real cost depends on how AI-heavy your workflows are, since a run that calls a big model several times burns far more than a simple one. Gumloop’s own pricing slider lets you buy more credits as you grow, which is the honest read: this is a usage tool, not a flat subscription. Pick Gumloop when you want to actually understand your agent by building it, on a free tier roomy enough to get good.

Zapier: the integration king, now with agents

Zapier wins on reach: it connects to more than 8,000 apps, more than anything else here, so if a tool you use has an integration anywhere, it’s probably on Zapier. For years that’s been classic fixed automation, “when this happens, do that”, and the agent layer, Zapier Agents, is newer and sold separately. An agent here can browse the web, take actions across your apps and run in the background, billed on its own “activities” meter rather than the task meter the rest of Zapier uses.

Pricing comes in two parts, which is the thing to understand before you buy. The core automation plans are task-based: free for 100 tasks a month, then Professional from $19.99 a month on annual billing for 750 tasks, climbing with volume, and a five-step automation burns five tasks every run. Zapier Agents sits on top: 400 activities a month free, then a Pro Agents add-on at $33.33 a month for 1,500 activities. There’s a sensible guardrail worth noting, and it tells you something about agents in general: Zapier caps each agent run (40 activities on the paid plan, 10 on the free one) and makes it pause to ask permission if it gets stuck in a loop, so it can’t quietly drain your allowance. Pick Zapier when integration coverage is your deciding factor, or when you already run automations there and want agents in the same place. Just price the full stack before you commit, because the layers add up.

Make: the cheaper visual builder

Make does much of what Zapier does on a more visual canvas and noticeably cheaper, which makes it the value choice for the second camp. You build scenarios by connecting modules on a board, it reads more clearly than Zapier once a flow gets complex, and it reaches across 3,000-plus apps. Its agent feature, Make AI Agents, lets you drop an autonomous, reasoning agent into any scenario, so you can mix fixed steps with an agent that decides, which is a genuinely useful middle ground.

Make’s free plan covers 1,000 credits a month across two active scenarios, and the Core plan is $9 a month on annual billing for 5,000 credits, with pricier tiers above it. Since August 2025 it runs on a credit model where a standard action is one credit, the same as the old operation, but AI steps cost more because they’re billed by how much the model does. Make AI Agents went generally available in early 2026, so it’s past the beta stage, but two flags still apply. First, the credit meter bites hardest exactly where you’ll use an agent: a single AI Agent run can burn dozens of credits against the one credit a plain action costs, so the bill climbs much faster than the $9 sticker suggests. Second, the lower price comes with a steeper learning curve than Relay or Gumloop: the canvas is powerful and busier, and you’ll spend longer getting comfortable. Pick Make when you want the most capability per dollar and you’re happy to climb a slightly harder tool to get it.

n8n: the technical pick, free if you self-host

n8n is the choice when you have someone technical and you want control, lower cost at scale, or your data on your own servers. It’s source-available, it has a proper native AI agent node built on the open LangChain framework, and crucially it bills differently from the others: by workflow execution, where one complete run counts as one execution no matter how many steps are inside it. That makes it dramatically cheaper than Zapier’s per-task or Make’s per-credit model once your volume climbs, because a twenty-step agent run still only costs one execution.

The two ways to run it tell the story. Cloud plans start at €24 a month (n8n prices in euros) for the Starter tier and 2,500 executions, with a 14-day trial. Or you self-host the Community Edition, which is free software with unlimited executions and all 500-plus integrations, and you pay only for the server it runs on, commonly $3 to $7 a month on a basic cloud box. That free self-hosted route is the cheapest serious option anywhere on this list, with one condition: it’s a technical setup, and the node graph rewards someone who’s comfortable with how APIs and data flow. This is the developer’s pick, not the non-technical operator’s. If you want to see what building an agent in it actually looks like, AI Foundations has a clear 50-minute walkthrough, How to Build AI Agents in n8n for Beginners, that goes from the basics to a working agent with memory and tools. Pick n8n when control and cost matter more than hand-holding, and you have the skills (or someone who does) to run it.

Three more worth a look

The six above are the names you’ll keep meeting, but a few others fill gaps worth knowing before you commit. The one an Australian reader shouldn’t skip is Relevance AI, built in Sydney: a no-code platform for standing up a small team of AI agents (sales, ops, support) that hand work between them, free to start and from about $19 a month. It’s the credible homegrown option if buying local or having the vendor in your own timezone counts, and it sits in a slot the other six don’t, agent teams rather than single workflows. Watch the meter, like the rest it bills on two counters at once, actions and vendor credits.

If your business already lives in Microsoft 365, Copilot Studio builds agents right inside that stack, which is often the path of least resistance, though its per-credit billing on top of the licence makes the true cost hard to pin down. And if you’re technical and n8n still isn’t open enough, the open-source pair Flowise and Langflow are free, self-hostable builders for people happy to run their own server. None of these unseats the six for a typical small business; they’re the ones to know when one of those gaps is yours.

The pricing trap: an agent decides its own steps

The thing that makes agent pricing genuinely hard to predict, and the part the comparison tables miss, is that an agent chooses how much work to do. With fixed automation you can count the cost in advance: n8n charges per run, Make per credit, Zapier per task, so a five-step Zap is five tasks every time, predictably. An agent doesn’t work like that. It decides how many tool-calls and reasoning steps a job needs, and those AI steps cost more than a plain action, so the very same task can cost wildly different amounts from one run to the next depending on how much the model decides to chew on it.

This isn’t a theory, it’s why the tools build in brakes. Zapier capping each agent at 40 activities a run and pausing it if it loops is a direct admission that an agent left alone can burn through your allowance. So budget for the agent’s worst run, not its best, and treat any “from $X a month” headline on an agent tool as the starting line.

Where the bill actually comes from. Not the subscription, the meter under it. Before you commit, work out the realistic cost of your agent doing its job a hundred times in a month, with the model taking a few steps each time, and compare that across the tools. A “cheap” plan with a stingy credit or step allowance often costs more in practice than a dearer one with a generous pool. Run one agent on one job for a fortnight first and read the actual usage, then you’re comparing real numbers instead of headline ones.

Do you even need an agent for this?

Before you build anything, ask: is this a job that actually needs an agent, or a fixed rule dressed up as one? Most of what a small business wants to automate is the same every time, “when a form comes in, add the row and send the welcome email”, and for that a fixed automation is cheaper, faster and completely predictable. An agent is the wrong tool for a job with clear rules, and you’ll pay more for less reliability by using one.

Reach for an agent only when the work genuinely needs judgement: the inputs are messy, the right next step changes case to case, or you can’t write the rules down in advance. That’s the agent-versus-automation line, and our agentic AI explainer draws it in full. If your job is a fixed, same-every-time process, the choice between these tools on price and speed alone is a different comparison, and the answer there is plain automation, not an agent. The same lean-team logic from doing more with a small team applies: match the tool to the task, and don’t pay for a thinker to do a typist’s job.

You might not need a builder at all

Here’s the move almost none of the rankings mention, because none of them make money from it: for a lot of small-business jobs you don’t need a separate agent-builder subscription, because the AI you already pay for now does the agent’s job. The clearest example is Claude Cowork, on Claude’s Pro and Max plans: it’s the autonomous engine behind Claude Code wrapped in a familiar chat-style workspace, so instead of wiring up a flow you hand it a brief and it works across your files and apps, decides its own steps, and comes back with the finished thing. You can save those jobs and run them on a schedule, and with Routines (in preview) they run in the cloud even with your laptop shut. That scheduled brief-and-go pattern, where you describe the outcome and the agent sets its own tasks, quietly does the work people buy a stack of dedicated builders for. On the other side a ChatGPT Custom GPT or its agent mode covers a lighter version of the same, and a Claude Project with connectors is the simpler “hold my instructions and reach my tools” option.

This isn’t a toy tier, either: Anthropic runs a genuine agentic platform, Cowork for non-technical operators and the Claude Agent SDK for developers wiring custom agents on the same engine, which is the serious version of the custom-build rung. The honest limits still keep it in proportion. ChatGPT’s agent mode works in a session, so nothing runs unless you open it, its scheduled tasks fire at most about hourly, and always-on, team-wide running needs the paid tiers. That gap is what the dedicated builders fill, an agent watching many app triggers or shared across a team with proper controls. There’s a durability point too: OpenAI said in mid-2026 it’s winding down its own standalone Agent Builder barely a year after launch, a reminder that a dedicated builder can be deprecated under you, while the assistant you already talk to is going nowhere. So the cheap order stands: try the job in the AI you already have, the way our agentic AI guide recommends starting, and only buy a builder once you’ve hit a wall that a builder is built to clear.

So, which should you pick?

Start from who’s building it and you won’t go far wrong. If you’re non-technical and want an agent working with the least setup, and budget isn’t the blocker, start with Lindy. If you want a free tier to learn in, Relay is the pick when your jobs need a human approval step, and Gumloop when you’d rather see the whole thing on a canvas. If you want the most power per dollar and don’t mind a steeper tool, Make from $9 a month is the value play. If you already live in Zapier or need its integration reach, add its agents there. And if you’ve got someone technical, n8n is free to self-host and the cheapest at any real volume.

Then do the one thing that decides whether any of this works: pick the single job that wastes the most of your week, build one agent to do just that, and watch it for a fortnight before you build a second. Keep a hand on anything it does that spends money or goes out under your name. The operators who get value from agents are the ones who started with one narrow job and proved it, not the ones who tried to build a digital workforce in a weekend. One agent, one job, supervised. That’s the part the hype leaves out.