Top 10 AI Agent Platforms for Small Business2026

A shop owner asked me a few weeks ago which AI agent platform she should "standardize on." She had about 40 customer emails a day, a Shopify store, a Google Calendar and a bookkeeper who wanted invoices chased. Her real question wasn't which platform was best. It was which one she could still afford, and still leave, a year from now.

That's a better question than most rankings answer. The lists I found lean on "ease of use" stars and resolution rates the vendors publish about themselves. Neither tells you what the bill looks like at your volume or whether you can take your work with you.

So on September 30, 2026 I did the slow version. I read the pricing, security and documentation pages of every small-business agent builder I could verify, wrote down what each page actually says, and scored ten of them on six things an owner can re-check in about twenty minutes per platform.

An agent platform is a product you log into; an agent framework is a code library a developer builds with.Search results mix them constantly, and it wastes people's time.

If you have an engineering team, the question is LangGraph against CrewAI against the OpenAI Agents SDK, and we compared those in our developer-framework comparison of agent platforms. This article is the other half: no-code and low-code builders a small business can run without hiring anyone.

The wrong model most owners bring to this is that an agent builder is like a phone plan. Pick the one with the best features for the monthly price, and switch later if you need to.

The trouble is that switching is where the costs hide. After six months your agent is a pile of prompts, connector credentials, approval rules and conversation history living inside someone else's product. Whether you can export it, and whether a person can step in when it goes wrong, matter more by month six than any feature on the landing page.

Economists call this switching cost, and it's the reason gyms offer a free first month. The first month isn't the product; the hassle of leaving is. Therefore I scored the things that decide your switching cost and your downside: can you see the price, can it reach your apps, will it sign what your data needs, will a human catch it, can you run it yourself, and can you leave with your work.

Six attributes, 0 to 2 points each, 12 points maximum, taken only from each vendor's own public pages and documentation on September 30, 2026.Here's the rubric before the results, so you can argue with it first.

Tie-break.Equal totals go to the lower published entry price for a paid plan, then alphabetical order. It's blunt. I'd rather use a rule you can see than a clever one you can't.

Who didn't make the list.Voiceflow and StackAI both have capable products, but their public pricing pages send a small business to a demo or a quote rather than a self-serve plan, so a ranking for owners who want to start this afternoon isn't the right home for them. LangGraph, CrewAI and other code libraries are frameworks, covered in the developer comparison. General chat assistants without a builder were out of scope.

Judgment calls, disclosed.Relevance AI's pricing page showed only an Enterprise plan with no dollar figure in the version I could read; third-party trackers print Pro and Team prices, but I scored the vendor page, so it takes a zero. Make's pricing page says extra credits come at a fixed price defined in your subscription without printing it, so it gets a 1. Lindy pauses service instead of charging overage, which I count as an explicit no-overage statement.

How to re-check.Every evidence cell paraphrases a vendor page on September 30, 2026, or says "not publicly disclosed." Compliance cells are vendor claims; I inspected no SOC 2 report or signed BAA. Company status was checked by search the same day and I found no acquisition or shutdown for any of the ten, but ownership changes quickly in this market.

Three platforms tied at 10 of 12, and n8n takes first on the tie-break bc its entry plan is the cheapest of the three.

Scores reflect public pages checked September 30, 2026. A zero means the attribute wasn't disclosed on the pages checked, not that the vendor can't provide it.

Three patterns jumped out.

First, self-hosting splits the field in two. n8n and Dify are the only ones any customer can download and run on their own server, and that alone is four points of their ten. If you never want to run a server (most small businesses don't), mentally drop that column and Copilot Studio, Zapier and Botpress move to the front.

Second, compliance is thinner than the badges suggest. Most of the ten display SOC 2. Only three state BAA availability on their own pages, and two of those limit it to Enterprise. Zapier is the most direct: its own blog says it isn't HIPAA compliant and doesn't sign a BAA. I respect that more than a vague badge.

Third, human handoff is nearly universal now. Nine of ten document a native way for a person to approve an action or take over a conversation. A year ago that column would have been emptier. It's the one I'd refuse to go without.

Each cell is what the vendor's page or documentation said on September 30, 2026, or "not publicly disclosed."

Sources: the vendor pages in the sources list. Where a page rendered prices only in script, I read the raw page source. Botpress's self-hosting status comes from its own documentation as surfaced by search, since the page redirected.

Read the pricing column for units before dollars. Zapier bills tasks. Make, Gumloop and Lindy bill credits. n8n bills executions. Botpress bills conversations. Chatbase and Dify bill message credits. Copilot Studio bills Copilot Credits. Those are seven different bets on what your work looks like, and a platform that's cheap for one shape of work can be expensive for another.

Here's each platform in rank order: what earned the score, what cost it points, who it suits and the downside.

Top ranked

1

n8n

10 out of 12

n8n is the workflow tool that grew an agent node, and it's the most complete package here for an owner with a little technical help. Its integrations page listed 2,279 integrations when I checked, and its human review documentation lets you require a person's approval before an AI Agent runs a specific tool, with the request sent to Slack, email and other channels. That's exactly the control you want on "send the refund" or "delete the record."

Workflows export as JSON, and a free Community Edition can be self-hosted. The n8n pricing page lists Starter at 20 euros a month billed annually for 2,500 executions and Pro at 50 euros for 10,000.

It lost a point on pricing because Starter and Pro don't state what happens past the execution limit (the Business plan does). And it lost one on compliance: its security page says it aligns to SOC 2 and offers the Type II report to enterprise customers, with no BAA statement.

Who it suits: owners with a technical friend or contractor who want power and a way out. The downside: the license is the Sustainable Use License, not a standard open-source one, so read it before you build a product on top; and self-hosting means you patch the server.

Visit n8n

2

Dify

10 out of 12

Dify is an open-source app and agent builder with a strong knowledge-base (RAG) side. Its pricing page lists a free self-hosted Community Edition, and cloud plans at $590 a year for 5,000 message credits a month or $1,590 a year for 10,000. Apps export as a YAML file, and a public plugin marketplace covers models, search, Slack, Notion, Outlook and Jira.

On compliance, Dify announced in March 2026 that it had completed SOC 2 Type II and ISO 27001:2022 audits for the second year. I found no BAA statement. Its Human Input node pauses a workflow and sends a form to a person by web app or email; the request closes after the first response.

Who it suits: businesses whose agent mostly answers from their own documents, and anyone who wants a self-hosted option with a friendlier builder than raw workflows. The downside:the cloud plans don't state overage, and the export leaves out usage logs and knowledge content, so "portable" means your design, not your history.

Visit Dify

3

Microsoft Copilot Studio

10 out of 12

If your business already lives in Microsoft 365, Copilot Studio is the only platform here with the full compliance stack stated in writing. Microsoft's compliance page says it is covered under the HIPAA Business Associate Agreement, audited for SOC, and listed for ISO and HITRUST.

Its handoff documentation is the most detailed in the set: the agent passes the full conversation and its variables to a live agent in a connected engagement hub. On the Copilot Studio pricing page, a pack of 25,000 Copilot Credits costs $200 a month, with pay-as-you-go as automatic backup.

It scores zero on self-hosting (it's a Microsoft cloud service). Export works through solutions, though Microsoft itself notes some components may not transfer.

Who it suits: Microsoft 365 shops, and anyone who needs a BAA without an enterprise negotiation. The downside: credits are consumed at different rates by different features, so forecasting a bill takes a licensing guide and a test month. Live handoff also needs an engagement hub to hand off to.

Visit Microsoft Copilot Studio

4

Zapier

9 out of 12

Zapier is where most small businesses already automate something, and its reach is unmatched: the Zapier pricing page cites more than 9,000 apps. Zapier Agents is billed separately in activities (400 a month free, 1,500 on Pro). The Professional plan starts at $29.99 a month, and pay-per-task overage runs at 2.5x your base rate on monthly billing, with a ceiling at three times your subscription before it pauses.

Human in the Loop pauses a Zap so a reviewer can approve, decline or edit data, on Professional and above. Your Zaps come out in a JSON file in the account data export.

The point it dropped: Zapier's own HIPAA explainer says it isn't HIPAA compliant and doesn't sign a BAA, while holding SOC 2 Type II. Clear, and a hard stop for health data.

Who it suits: owners who want the agent to touch many apps with the least setup. The downside: two billing units (tasks and activities) and a 2.5x overage multiplier mean a busy month can surprise you; no self-hosting; no health data.

Visit Zapier

5

Gumloop

9 out of 12

Gumloop is a newer visual builder with a clean pricing story. Its pricing page lists Pro from $37 a month with 20,000 credits (7,400 plus 12,600 bonus), and pay-as-you-go usage at $0.005 per credit. That's the most forecastable credit price in the set.

Its Enterprise page lists SOC 2, HIPAA and BAA support, plus VPC and self-hosted deployment. The human-in-the-loop documentation lets you set tools to ask each time, or to ask only for writes and deletes, which is a sensible default.

It lost points on integrations (I couldn't find a published count or directory) and export (Enterprise data exports cover runs, agents and interactions as CSV, but I found no re-importable workflow file).

Who it suits: small teams that want predictable credit pricing and plan to grow into an enterprise contract. The downside: the BAA and self-hosting live on Enterprise, not Pro, and leaving means rebuilding flows by hand.

Visit Gumloop

6

Botpress

9 out of 12

Botpress is built for customer conversations, and its pricing is conversation-shaped. The Botpress pricing page lists Plus at $150 a month billed annually for 250 conversations and Team at $750 for 1,500, with extra packs of 100 at $65 on Plus ($0.65 each) or $50 on Team.

Its HITL documentation escalates conversations to live human agents in its own inbox or tools like Zendesk. Whole bots export as a .bpz archive, and its integrations hub lists more than 130 integrations.

It lost points on compliance (its enterprise page says SOC 2 Type II and HIPAA compliant, but I found no BAA statement) and on self-hosting, since the old self-hosted line has been sunset.

Who it suits: businesses whose agent is mainly a website or WhatsApp support assistant with a human inbox behind it. The downside: a higher entry price than the workflow tools, and a per-conversation bill that grows with traffic, not with value.

Visit Botpress

7

Lindy

8 out of 12

Lindy has repositioned around a personal assistant you text and a shared teammate in Slack. The Lindy pricing page lists Plus at $29.99 per user for 3,000 credits, Pro at $99.99 for 15,000 and Max at $199.99 for 35,000. When credits run out, service pauses; there's no overage bill. It states SOC 2 and offers a signed BAA on Enterprise.

Its documentation says every action that writes or sends needs approval, with Approve and Deny buttons in Slack threads, while read-only actions don't.

It scored zero on self-hosting and zero on export: I searched its full documentation and found no way to export an agent or its history.

Who it suits: an owner who wants an inbox and calendar assistant this week, not a workflow project. The downside: per-user pricing, a credit cost per task that ranges from 2 to 2,500 credits depending on the work, and no documented exit.

Visit Lindy

8

Make

7 out of 12

Make is the visual workflow tool many owners graduate to from Zapier, and it has the lowest paid entry price here. The Make pricing page lists Core at $12, Pro at $21 and Teams at $38 a month for 10,000 credits, AI Agents on every plan, and more than 3,000 apps. Scenarios export as blueprint JSON.

Its security page states SOC 2 Type II and SOC 3 audits and an ISO 27001 certified program, with no BAA.

It lost points in three places. Extra credits are sold at "the fixed price defined in your subscription," which the page doesn't print. The human approval option I found is an app in its integrations directory, not a documented native feature. And there's no self-hosting.

Who it suits: price-sensitive owners comfortable with visual logic. The downside:you can't price a busy month from the public page, and approvals depend on an add-on.

Visit Make

9

Chatbase

7 out of 12

Chatbase is the quickest way to put a support agent trained on your website in front of customers. The Chatbase pricing page is fully published: Hobby $40 for 700 message credits, Standard $150 for 4,000, Pro $500 for 15,000, and auto-recharge at $40 per 1,000 credits.

Its Live Chat action hands a conversation to a person in its Helpdesk, and chat logs export as JSON, PDF or CSV.

It lost points on integrations (named natives such as Slack, HubSpot and Calendly, but no directory count), compliance (SOC 2 Type II and HIPAA-eligible features on Enterprise, but BAA availability isn't stated in plain text) and portability (you can export conversations, not the agent).

Who it suits: a business that mainly needs website questions answered and a clean handoff. The downside:it's a support agent more than a general builder, so complex back-office jobs belong elsewhere.

Visit Chatbase

10

Relevance AI

6 out of 12

Relevance AI builds "AI workforces" for sales and operations, and its controls are good. Its approvals page describes adding approvals to any step and writing escalation rules in plain English, routed through Slack or email. Its security documentation states SOC 2 Type II, a choice of US, EU or Australian data region fixed at signup, and data export as CSV, Excel or JSON.

It ranks last on this rubric for one reason: the pricing page I could read showed only an Enterprise plan with no price. Other sites print self-serve prices for it; I didn't score them, because the rubric scores what the vendor publishes.

Who it suits: sales-led teams that want agents working a pipeline with approvals. The downside:you can't price it without a conversation, and there's no BAA statement or self-hosting.

Visit Relevance AI

The cheapest headline is rarely the cheapest bill.Here's the arithmetic at realistic small-business volumes, using only published prices.

Let me work the Botpress line out loud, since conversation pricing is where people get surprised. Say your website gets 1,000 support conversations a month. Plus includes 250. The other 750 at $0.65 is $487.50, plus the $150 base: $637.50. Team at $750 includes 1,500, so at 1,000 conversations you're only about $112.50 apart ($750 minus $637.50), and past roughly 1,170 conversations Team is cheaper. That crossover is worth knowing before you pick the "cheap" plan.

Now Zapier, where the multiplier bites. Professional at $29.99 for 750 tasks works out to roughly 4 cents a task. Overage at 2.5x is roughly 10 cents. Run 2,000 tasks and the 1,250 extra cost about $125, so the month lands near $155. At that point a higher task tier almost certainly beats pay-per-task, so treat overage as insurance, not a plan.

Gumloop is the easiest to forecast. 50,000 credits a month is 30,000 over the included 20,000; at $0.005 that's $150, plus $37: $187. The unknown is how many credits your particular flow burns per run, which the trial tells you.

Copilot Studio's pack price works out to $200 divided by 25,000, or $0.008 a credit. Pay-as-you-go is widely reported at $0.01 a credit, though the pricing page I loaded didn't print that figure. The hard part is converting credits into conversations, because different features consume different amounts. Run a test month before you commit to packs.

And the self-hosted option isn't free, it's differently priced. n8n Community Edition costs nothing to license, but you pay for a server and, more importantly, for someone's time to keep it patched. For a five-person business that time usually costs more than a Pro plan.

If you want the broader cost picture (custom builds, agencies and model usage together), our AI agent cost breakdown, listed in the related articles below, works it end to end.

For most small businesses, one of these ten on a trial plan is the right answer.I'll say that plainly, even though custom agents are what we sell.

If your job is answering website questions, triaging an inbox, drafting follow-ups, moving data between apps that already have connectors, or booking into a calendar, a platform will get you there in days for tens or hundreds of dollars a month. Building that from scratch would be like hiring a carpenter to build a bookshelf you can buy flat-packed.

Custom earns its price in three situations, and only these three in my experience.

• The agent must write into a system with no connector: Reading is easy; writing is where platforms stop. If the action lives in a portal with no API, or an API no builder supports, a platform can only notify a person to do it.
• You need to own the code and the data path: When a customer contract, a regulator or your own risk appetite says the logic and data must live in your infrastructure under your control, a hosted builder can't give you that, and even a self-hosted one keeps you on its license.
• Your rules are more complex than the builder can express: Pricing logic with twenty exceptions, multi-step approvals that depend on who is asking, or decisions that need deterministic code rather than a prompt. Builders handle branches well and edge-case policy badly.

The clearest example we've shipped is Beyond Points AI. Its agent had to move loyalty points from card programs to travel partners, and we built a browser agent that completes those card-to-partner transfers end to end, with every irreversible step (transfer, book, pay, cancel) behind an explicit user confirmation. It sits on 71 edge functions and a multi-agent layer. I found no connector for that transfer step on any builder in this list, which is exactly the first situation above.

Now the price, out loud. A custom AI agent build from us is $5,000 for the agent plus $5,000 setup, so $10,000 to start, and it rises with complexity and the number of integrations. You pay model, telephony and API usage directly at provider rates. After launch it's $2,500 a month. Year one is $10,000 plus 12 times $2,500, which is $40,000 plus usage; each later year is $30,000 plus usage.

Compare that with the Botpress Plus example above: $637.50 a month is $7,650 a year. The custom build is about 5.2 times that in year one ($40,000 divided by $7,650). You should only pay that multiple if one of the three situations is true. You get full source code and IP transfer, a 30-day post-launch warranty and a fixed-price phased proposal within 5 business days. We sign a BAA with healthcare clients; we hold no SOC 2, ISO 27001 or HITRUST, so if your procurement requires a SOC 2 report, Copilot Studio or an enterprise platform plan will clear it and we won't.

If your need is a multi-step back-office process, our custom workflow agent page shows what's included. If you're still weighing it, the build versus buy framework walks the decision step by step.

Whichever platform you pick, decide the boundary before you connect anything.This is the table I'd tape to the monitor.

Every platform ranked here that scored a 2 on handoff can enforce the middle column. Set it up on day one, not after the first mistake.

One more thing that doesn't appear on any pricing page. Any agent that reads untrusted input (customer emails, web forms, uploaded files) can be steered by text hidden in that input. That's prompt injection, and nobody has solved it. The practical answer is smaller blast radius: give the agent the fewest connector permissions it needs, and keep irreversible actions behind approval.

These would make me walk away from a platform trial, whoever is selling it.

• No overage rate on the pricing page: If you can't calculate a busy month, you'll find out on the invoice. Ask for the rate in writing.
• Resolution rates without a method: A deflection or resolution percentage with no sample, period or definition is marketing. Measure your own handoff rate in the trial.
• A HIPAA badge without a BAA: A badge is not a signed agreement. If health data will pass through the agent, get the BAA before the first real conversation.
• No approval step on write actions: If the agent can send, pay or delete without a pause you configured, you are one bad instruction away from a very bad day.
• No export of the agent itself: Exporting conversations is not the same as exporting your agent. Ask how you would rebuild it elsewhere, and time how long that would take.
• Connector means 'webhook': Ask what the connector can actually do: read, create, update, delete? In which objects? A connector that only reads can't finish the job.
• Customer data trains the model: Get a written answer on whether your conversations are used for training, by the platform or its model providers.

For the security side of that list in more depth, see the ten controls to set before you connect your data.

This ranking measures disclosure, not quality. A platform that publishes everything could still build agents that misread your customers. A platform that publishes little could be excellent. The trial is where you find out.

I did not build a test agent on each platform for this article, and I inspected no SOC 2 report or BAA. Compliance cells are the vendors' own claims.

Several pages render prices in script. For Gumloop I read the raw page source, and for Make and Relevance AI the readable page was incomplete, which is why Relevance AI scores zero on pricing and Make one. Copilot Studio's pay-as-you-go rate and Botpress's self-hosting status rest on search results rather than a full page load.

The rubric weights self-hosting and export as heavily as pricing. That flatters n8n and Dify, and it's a choice: if you'll never run a server, discount that column. The scope is ten platforms; several I excluded (Voiceflow, StackAI) could qualify if they publish self-serve pricing. Pricing pages change without notice, so everything here is dated September 30, 2026.

You'll see these in other articles about AI agent platforms. They're not here, and here's why.

Every vendor resolution, deflection, containment or time-saved figure, including the resolution rate on Botpress's enterprise page. They might be accurate. None has an independent benchmark behind it.

The claim that 95% of generative AI pilots fail. It traces to a study built on a small set of interviews and conference surveys about custom tools, and it gets quoted as if it described every agent a small business might deploy. It doesn't, and it isn't evidence about any platform here.

"85% of AI projects fail" and "87% never reach production." The first misquotes an old prediction about output quality; the second is a rhetorical question from a sponsored post. Neither measures anything, so I refuse to print either as fact.

Third-party prices for Relevance AI, and a connector count for Copilot Studio that I couldn't find on Microsoft's own page. They may be right; I scored only what the vendor itself published.

You can get from this article to a working agent in about a week.Here's the order I'd do it in.

1. 1.Write one page: the single job you want the agent to do, the apps it must read and write, what it may never do alone, and your volume (emails, conversations or tasks a month). That page is your filter and your test.
2. 2.Pick two platforms from the scoreboard whose integration and handoff cells match that page, then do the cost arithmetic at your volume with the published prices before you start the trials.
3. 3.Run each trial on ten real cases, including two that should be escalated to you. Export whatever the platform lets you export on day one, and screenshot the pricing page the day you subscribe.

If neither trial can do the job because the write action lives somewhere no connector reaches, that's when custom is worth a conversation. Otherwise, pick the better trial and get back to your business.