AI Adoption & Operations

AI Agent Platforms Compared: Copilot Studio, Agentforce and the Criteria for Selection

Marin Niehues28 May 202612 min read

Server fans in a data centre

The right platform for AI agents is the one that sits where the data and permissions of the process live and gives you control over every action the agent takes. This article sets out the four platform types, names the selection criteria and shows how to calculate the cost per completed case. This article provides an overview and does not constitute legal advice.

An AI agent differs from an assistant in that it acts: it plans steps, calls systems and changes data, for instance by triggering a refund. With an assistant, answer quality matters most. With an agent, what matters is access to the right systems and control over every action. Choosing a platform on the strength of a product demo means assessing the wrong property.

Which platforms are there for AI agents in the enterprise?

AI agent platforms fall into four types: agents within business applications, low-code platforms from workplace vendors, developer platforms from the major clouds, and open frameworks. The types differ less in the performance of the language models than in proximity to data and in how they are built and run.

Agents within business applications. Vendors of CRM and ERP systems offer their own agent platforms. Examples include Salesforce Agentforce, the agents from ServiceNow and SAP Joule. Their strength is direct access to the data and permissions of the application concerned. A service agent in Agentforce knows the customer, the contract and the open case without any additional interface. The limits show as soon as a process touches several systems from different vendors.

Low-code platforms from workplace vendors. Microsoft Copilot Studio is the best-known example. Business teams build agents through a graphical interface, use the identities and permissions from Microsoft 365 and connect further systems through prebuilt connectors. Google offers comparable tools for its workplace and cloud environment. The strength is speed to the first agent. The limits lie in complex logic and high-volume processes, where the pricing model and testability gain weight.

Developer platforms from the major clouds. Amazon Bedrock, Google Vertex AI and the AI services of Microsoft Azure provide building blocks for custom agents, from model access through orchestration to monitoring. The strength is full control over logic and model choice. The prerequisite is a development team that builds and runs the agent.

Open frameworks. Libraries such as LangGraph, CrewAI or AutoGen allow agents to be built independently of a platform vendor. The strength is freedom of architecture and hosting location, up to operation in your own data center. The downside: security and operation rest entirely with the company.

Eight criteria decide the selection

An agent platform is assessed against eight criteria, and the first is proximity to the data and permissions of the process. The remaining seven test control and operation as well as economics:

  1. Data proximity: Does most of the data the agent needs sit in one leading system? Then that system's platform has a structural advantage.
  2. Actions and integration: Through which interfaces does the agent reach other systems? Does the platform support open standards such as the Model Context Protocol and Agent2Agent, so that tools and agents from other vendors remain connectable?
  3. Control of actions: Can you set for each action whether the agent executes it alone or needs human approval? Does the agent act with the user's rights or with its own narrowly limited rights?
  4. Traceability: Does the platform log every step with its tool call and decision in a way that business teams and internal audit can review?
  5. Evaluation and testing: Can you create test cases with expected results and run them automatically after every change?
  6. Model choice: Can the language model be swapped without rebuilding the agent?
  7. Data protection and location: Where are inputs and logs processed, and what does the vendor commit to in the data processing agreement?
  8. Pricing model and exit: Is billing per user, per conversation, per action or by consumption, and how costly is a later switch?

Criteria 3 and 4 separate a platform for operation from a platform for demonstrations. An agent whose actions cannot be approved and traced individually has no place in a process with customer or financial impact. How to set the levels of human control is covered in the article on human in the loop, and the significance of the open standards in the article on Model Context Protocol and A2A.

Research shows why this caution is justified. In a benchmark study by Carnegie Mellon University, the strongest agent system tested completed around 30 percent of 175 work tasks in a simulated software company fully on its own. The conclusion for selection: the platform must support partial autonomy with human approval well, because that is the mode in which most productive agents run.

Copilot Studio or Agentforce: which platform fits?

Copilot Studio fits when the agent mainly works with documents and email in the Microsoft environment, Agentforce when it handles customer data and cases in Salesforce. The question is therefore one of the leading system, and three checks answer it.

First: where does the case the agent handles live? A service agent that matches customer requests against contract data belongs in the system where customer and contract are held. An agent that drafts quotes from templates and coordinates them internally belongs in the workplace environment. Second: who builds and maintains the agent? Copilot Studio is aimed strongly at business teams with low-code experience, the business application platforms at the administrators of those systems. Third: how high is the volume? At many thousands of cases a month, the pricing model decides the economics, and a comparison with a custom build becomes worthwhile.

Many companies end up using several platforms. That is consistent, provided a central function sets the rules for approvals and permissions uniformly. Without that framework, agents with unclear rights multiply. When buying, adapting or building is the right choice is covered in Make or buy.

How much does an AI agent platform cost?

Measure the cost of an agent platform by the cost per completed case, because only this figure makes platforms with different pricing models comparable. License prices alone mislead, because they leave out model consumption and human review.

The monthly calculation has four items:

  1. Platform: licenses per user, allowances per conversation or action.
  2. Model and infrastructure: language model consumption, plus hosting and monitoring for a custom build.
  3. Operation and development: person-days for maintenance and testing.
  4. Human review: staff time for approvals and corrections.

Divide the total by the number of cases the agent closes without rework. An example with assumed values shows the effect. An agent handles 4,000 requests a month, 3,000 of them without rework. Platform and model together cost 6,000 euros, operation 4,000 euros, human review 5,000 euros. The cost per completed case is 5 euros. If an employee handles the same case in twelve minutes at a cost rate of 60 euros per hour, the case costs 12 euros. The comparison also reveals the strongest lever: if the share without rework rises from 75 to 90 percent, the cost per case falls markedly, because review time disappears.

If external support is added for selection and prototype, it is typically billed in the German consulting market at around 1,000 to 2,500 euros per consultant day, depending on seniority.

How does a robust platform selection work?

A robust selection takes six to ten weeks and ends with a test in which two platforms handle the same real process with real data. A product demo shows the best case, the test shows everyday reality.

The approach has four steps. First choose a process with high volume and limited damage in case of error, such as pre-qualifying service requests. Then create test cases with expected results. Difficult and impermissible requests belong in the set. Next, two platforms build the same agent in two to four weeks. Finally compare result quality, cost per case and the effort for approvals and logging.

The penultimate question is often forgotten. An agent that reads emails and may also execute actions is a target for hidden instructions, so its rights must stay narrow enough that a misstep remains contained.

Niehues Consulting as partner for platform selection

Niehues Consulting supports companies from choosing the first process through the platform test to the agent in operation. Marin Niehues is a management consultant with a delivery perspective. He builds organizations that deliver and brings AI into operation. His projects range from VC-funded scale-ups to DAX 40 corporations, he has worked in 20+ countries on six continents, and his home base is the DACH region.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Embed. Three of its review questions matter most for platform selection. Which business outcome should change measurably? This question determines the process and therefore the leading system. Does the prototype prove its value in the real process with real data? Only this answer justifies the platform decision. Who owns operation from day one? The answer decides whether the agent takes hold in daily work after the test. How this shapes the whole route of an AI application into operation is set out in the article From pilot to scale. The entry point is a free 30-minute initial consultation via the contact form.

Frequently asked questions

What platforms are there for building AI agents?

There are four types: agent platforms within business applications such as Salesforce Agentforce, low-code platforms such as Microsoft Copilot Studio, developer platforms such as Amazon Bedrock or Google Vertex AI, and open frameworks such as LangGraph. They differ mainly in data proximity and how agents are built. The choice follows the system in which the agent's process lives.

What is Microsoft Copilot Studio used for?

Copilot Studio is Microsoft's low-code platform for building custom agents and copilots. It uses identities and permissions from Microsoft 365 and connects further systems through connectors. It suits business teams that want to build agents without their own development team.

What is Salesforce Agentforce?

Agentforce is Salesforce's platform for AI agents that work directly with customer and case data in Salesforce, typically in customer service and sales.

Which AI agent platform is best for enterprises?

The best platform is the one closest to the leading system of the process that makes every agent action controllable. For service in the CRM that is usually the CRM vendor's platform, for processes in the workplace environment a low-code platform such as Copilot Studio. For processes across several systems with their own logic, a developer platform is the right route.

Do you need coding skills to build AI agents?

Simple agents on low-code platforms require skills in configuring workflows. Agents spanning several systems, with high volume or custom logic need a development team. Testing and logging need a named owner in both cases.

Should a company use more than one agent platform?

Yes, when processes sit in different leading systems. The prerequisite is a central function that sets rules for permissions and approvals uniformly across all platforms. Without that framework, agents emerge with unclear rights and no owner.

How long does it take to select an agent platform?

A robust selection takes six to ten weeks. It covers choosing the process, defining test cases and a test in which two platforms handle the same real process. Deciding on product demos alone is faster but does not test everyday reality.

Are open-source agent frameworks a viable alternative?

Yes, for companies with their own development team and specific requirements for hosting location or independence. Frameworks such as LangGraph or CrewAI carry no license fees but shift security and operation entirely to the company. Without a team for that operation, a platform is the more economical route.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation