Which AI Agent Platform Should Your Business Build On?

An AI agent for an Australian business runs on one of three layers: a low-code platform such as Microsoft Copilot Studio or n8n, a developer framework such as LangGraph or the Claude Agent SDK, or the model provider’s API called directly from code you own. Pick the layer by who will maintain the agent, where your data is allowed to sit, and what one completed job costs at your real volume. The layer decides what the agent costs in year two, who can change it, and whether the OAIC’s due-diligence questions have answers. The model it runs on matters less than any of those, because the model is the part that gets swapped.

The question usually reaches me as “which framework should we build on?”, and my honest answer is that framework is the wrong word for most businesses. A framework is a kit for developers. What a business is choosing is the layer the agent lives on, and that choice sets the running cost, the lock-in and who fixes it when it stops on a Friday afternoon. I have written up how to choose the agent itself: the job, the form, the supplier and the controls. This post is the next decision down.

What are the three layers an AI agent can run on?

The three layers are a low-code platform, a developer framework, and the model API called directly. A platform gives you a visual builder, hosting and connectors for a subscription plus usage. A framework gives a developer the building blocks and leaves hosting, testing and monitoring to them. The API gives you the model and nothing else. Each moves cost and control in a different direction, and the table is the shortest honest version of the trade.

Layer	Examples	Who it suits	What you pay	Where it stops
Low-code platform	Microsoft Copilot Studio, n8n, Zapier Agents, Salesforce Agentforce	A business with a capable operator and mainstream systems	Subscription plus usage, billed in credits, executions or conversations	Anything the connectors do not reach, and logic that needs to be tested like software
Developer framework	LangChain and LangGraph, CrewAI, AutoGen, the Claude Agent SDK, Strands Agents SDK	A business with a developer on staff or an agency on retainer	Developer time, model usage and hosting	The abstraction hides the prompt when something breaks
Model API directly	Anthropic, Google Cloud and AWS model endpoints	A narrow, high-volume job that must be cheap per run	Model usage and hosting, plus a developer who owns the code	Retries, logging and the hand-off to a person are yours to build

Microsoft describes Copilot Studio as “a graphical, low-code studio for building and managing AI-powered agents and workflows” (Microsoft Learn), and that sentence is a fair definition of the whole platform layer: a studio, not a codebase. The framework layer is what Anthropic’s engineering team was writing about when it listed the Claude Agent SDK, the Strands Agents SDK from AWS, Rivet and Vellum among the tools that make agents easier to build, and then set out what they cost you:

“These frameworks make it easy to get started by simplifying standard low-level tasks like calling LLMs, defining and parsing tools, and chaining calls together.”

Anthropic, Building effective agents (Anthropic)

The same page adds that they “can also make it tempting to add complexity when a simpler setup would suffice” and suggests that developers “start by using LLM APIs directly: many patterns can be implemented in a few lines of code” (Anthropic). I agree with that, and I would go one step further for a business rather than a developer: the simplest setup is often no framework and no custom code at all, because the job turns out to be a rule with one judgement step in the middle. The three ways to automate a process is the test I run before any of this.

When does a low-code platform pay off?

A low-code platform pays off when the job is standard, the systems are mainstream, and someone on your team will own the agent as part of their role rather than as a side project. Copilot Studio, n8n, Zapier Agents and Agentforce all sell that arrangement, and Copilot Studio, n8n and Agentforce each publish a usage-based unit rather than an hourly rate, which is the first thing to read before the feature list.

The pricing units differ, and the unit tells you how the vendor expects the agent to be used. Microsoft notes that “Starting on September 1, 2025, the common currency for agents changed from messages to Copilot Credits” (Microsoft Learn). n8n prices “based on monthly workflow executions, regardless of complexity”, with a Starter plan at 20€ a month billed annually and a Pro plan at 50€ a month for 10K workflow executions, and it keeps “a standard, self-hosted version of n8n” available on GitHub (n8n). Salesforce lists Agentforce at AU$2.80 per conversation, or AU$700 per 100k Flex Credits for pay-per-action (Salesforce). Zapier’s product page says its agents “perform tasks across 9,000+ apps” (Zapier).

Platform	Pricing unit on the vendor’s page	Published price, checked 28 September 2026
Microsoft Copilot Studio	Copilot Credits (replaced messages on 1 September 2025)	Depends on licence and pack; Microsoft publishes a usage estimator
n8n	Workflow executions per month	Starter 20€ a month billed annually; Pro 50€ a month for 10K executions; self-hosted Community Edition on GitHub
Salesforce Agentforce	Conversations, or Flex Credits per action	AU$2.80 per conversation; AU$700 per 100k Flex Credits

Prices are the vendors’ own published figures on the date shown and change without notice; check the linked page before you budget on them.

Here is how I read that table for a business of five to fifty people. A conversation-priced agent suits a customer-facing job with countable conversations and a CRM already in the same vendor’s hands. An execution-priced platform suits back-office work that runs on a schedule, because you can forecast the bill from the calendar. A credit-priced agent inside a Microsoft tenant suits a business whose documents and staff already live there. The platform that wins is the one whose unit matches the shape of your job, and the operator who will own it should be able to name a workflow they built last quarter. If nobody can, the subscription is paying for a capability nobody is using.

When do you need a developer framework, and does it matter which one?

You need a developer framework when the job is specific to your process and your systems, when the agent has to be tested like software before it touches real records, or when a platform’s connectors stop short of what the job needs. Which framework matters less than whether a developer will still be around to maintain it. LangGraph, CrewAI, AutoGen and the Claude Agent SDK are all real choices with different trade-offs, and none of them removes the need for someone to own the result.

IBM’s own guide to the field is a useful map of what each one is. It describes LangChain as “another open-source framework for building LLM-powered applications, including chatbots such as ChatGPT and AI agents”, AutoGen as “an open-source framework from Microsoft for creating multiagent AI applications to perform complex tasks”, and CrewAI as a beginner-friendly option with “a no-code interface for rapid prototyping and ready-made AI agent templates for swift deployment” (IBM). Those are accurate descriptions and they are also the point: every one of them is a tool for a developer. A business does not choose a framework. It chooses a developer, or an agency, and inherits their framework.

The research on agents backs the plain approach over the elaborate one. Sayash Kapoor, Arvind Narayanan and their co-authors benchmarked published agents on cost as well as accuracy and found:

“As a result, SOTA agents are needlessly complex and costly, and the community has reached mistaken conclusions about the sources of accuracy gains.”

Sayash Kapoor et al., AI Agents That Matter (arXiv)

That matches what I see in quotes that cross my desk: a multi-agent design for a job that is one classification and one lookup. The evidence on who should do the building points the same way. MIT’s NANDA initiative found that buying from specialised vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded only one-third as often (Fortune). So, my honest opinion: if you are going to the framework layer, do it with someone who has shipped on it before, scope the agent to one measurable job, and insist that the code, the prompts and the accounts end up in your name. That is how our own AI agent builds are handed over, and it is the difference between owning an agent and renting a developer.

Where will the agent’s data sit, and does the platform have an Australian region?

The data question decides the platform for a medical practice, a law firm or an accountant before any feature does. Microsoft lists Australia East (New South Wales) and Australia Southeast (Victoria) among Copilot Studio’s data locations, Google Cloud offers Sydney (australia-southeast1) for its Gemini models, and n8n can be self-hosted on your own infrastructure. The regulator expects you to have checked, and to keep checking.

The detail matters. Microsoft’s data-location page states that “If a tenant’s location isn’t listed in the data locations table, data is stored in the United States” (Microsoft Learn), which means residency is a property of your tenant setup, not of the product. Google’s region list for the Gemini Enterprise Agent Platform includes Sydney (australia-southeast1) alongside Singapore and Tokyo (Google Cloud), so a build on that layer can be pinned to an Australian region if the developer pins it. A self-hosted n8n instance keeps the workflow data wherever you run it, and the model call is then the only thing that leaves.

The Office of the Australian Information Commissioner is direct about how to treat all of this. Its guidance on commercially available AI products says that “AI products should not be used simply because they are available”, and that “Due diligence for AI products should not amount to a ‘set and forget’ approach” (OAIC). The OAIC received 532 data breach notifications in the January to June 2025 reporting period, and the health sector reported the most at 18% (OAIC); those figures sit in our Australian SMB AI index alongside the adoption numbers. The Privacy Act also covers some small business operators with an annual turnover of $3 million or less, including private sector health service providers (OAIC), so a small practice does not get the small-business exemption its neighbours do. For those businesses the platform question and the data question are the same question, and what it takes to run a private AI agent is the version of this post where the answer is “on your own hardware”.

What does each layer cost over a year?

Over a year, a platform costs its subscription plus usage, a framework costs developer time plus model usage plus hosting, and the API route costs the least per run and the most in engineering you have to own. The comparison that matters is the cost of one completed job at your real monthly volume, and the sticker price rarely predicts it.

Volume is the number to get right first. n8n’s own pricing page gives the arithmetic: “a daily schedule would result in 30 or 31 executions per month, while one that runs every 5 minutes would result in about 8,600–8,900 executions monthly” (n8n). The same job costs a few dollars or a few hundred depending on how often it polls, and that is a design decision, not a pricing one. On a conversation-priced platform the arithmetic is simpler: at Salesforce’s published AU$2.80 per conversation, a thousand customer conversations a month is AU$2,800 in usage before anyone has built anything, and ten thousand is AU$28,000. That is my arithmetic on their figure, and it is the reason a customer-facing agent needs a volume estimate before it needs a demo.

A build has the opposite profile: most of the cost is up front and the running cost is model usage and hosting. Our own AI and automation projects start at $1,990 per project on a fixed quote, with a full workflow audit before anything gets built, and the running cost is estimated from your volumes in that audit. The honest comparison is a three-line table for your own job: platform subscription plus usage at your volume, a build’s fixed cost plus its usage at the same volume, and the hours the job takes a person today. Whichever line is lowest in year two, not month one, is the platform.

How should an Australian business decide?

Decide in this order: who will maintain the agent, where the data may sit, and what one completed job costs at your volume. The Australian Bureau of Statistics counts only 12% of Australian businesses using AI in 2024–25, up from 1% in 2021–22 (ABS), against the 78% of organisations that Stanford’s 2025 AI Index reports using AI in 2024 (Stanford HAI). Most of the businesses I talk to are choosing their first layer, not replacing one, and the first choice is the cheap one to get right.

The staff, meanwhile, are ahead of the business. Roy Morgan puts 13.6 million Australians, 58% of people aged 14 and over, on AI tools in an average four weeks (Roy Morgan). So the platform decision is rarely about whether AI enters the business. It is already in the building on someone’s phone. The decision is whether it runs on a layer the business chose, with the data where the business can account for it.

My rule of thumb, for a business of five to fifty people: take the platform layer if the job is standard and a named person will own it; take a build on a framework or the API if the job is specific to how you work, and have it built by someone who has done it before; and take the API directly only for a narrow, high-volume job where cost per run is the whole game. Whichever layer you land on, the agent should be scoped to one job with a baseline taken before anything is built, because that is the only way the year-two comparison can be honest. If you want a second opinion on which layer fits, the workflow audit is where we work that out, and if a platform subscription is the right answer we will say so. You can see how we scope and price a build or start a conversation. Choose the layer for the job in front of you, keep the data where you can account for it, and the agents you add later will sit on the same foundation rather than fighting it.

Frequently asked questions

What is the difference between an AI agent platform and an AI agent framework?

An AI agent platform is a hosted product with a visual builder, connectors and a subscription, such as Microsoft Copilot Studio, n8n or Salesforce Agentforce, and it is operated by a capable non-developer. An AI agent framework is a software library for developers, such as LangGraph, CrewAI, AutoGen or the Claude Agent SDK, that supplies building blocks and leaves hosting, testing and monitoring to the person writing the code. A business chooses a platform directly; it chooses a framework by choosing a developer.

Do I need a developer to use an AI agent framework?

Yes. Frameworks such as LangChain, LangGraph, CrewAI and AutoGen are code libraries, and Anthropic’s own guidance to developers is to start by calling the model API directly and add a framework only when the job needs it. A business without a developer on staff should either use a low-code platform or have an agent built and handed over by an agency, with the code, prompts and accounts in the business’s own name.

Can an AI agent keep my data in Australia?

Often, but only if it is set up that way. Microsoft lists Australia East and Australia Southeast among Copilot Studio’s data locations, and states that a tenant not listed in its data-location table has its data stored in the United States. Google Cloud offers a Sydney region for its Gemini models, and n8n can be self-hosted on your own infrastructure. The OAIC expects an organisation to do due diligence on an AI product before adopting it and to keep doing it afterwards, so confirm the region in writing rather than assuming it.

How much does an AI agent platform cost per month?

It depends on the unit the vendor bills in. On 28 September 2026 n8n’s Starter plan was 20€ a month billed annually and its Pro plan 50€ a month for 10K workflow executions, Salesforce listed Agentforce at AU$2.80 per conversation or AU$700 per 100k Flex Credits, and Microsoft bills Copilot Studio agents in Copilot Credits that depend on the licence. Multiply the unit by your real monthly volume before comparing any of them with a fixed-price build.