Most people know Copilot as a chat and expected more from it. The bigger lever is elsewhere: with Copilot Studio, you build your own AI agents that get work done in your processes – with the AI model that fits the task, and under your control.
Ambit Group · Microsoft Solutions Partner, AI Business Solutions
30 minutes, free of charge, directly with Christian Schipp.
Invoice PDF arrives in the AP mailbox · 6:02 a.m.
AI has been rolled out, the work has stayed the same. We see three patterns again and again:
Employees have texts written and summarized. Invoices, orders and requests still run exactly as before.
Teams compare whether ChatGPT, Claude or Copilot answers better. What decides the value is whether the AI is connected to data and systems and is allowed to take on work.
Many ideas and a few pilots, but no figure that shows leadership what it delivers. So AI stays an experiment.
of agentic AI projects will be canceled by the end of 2027 – due to escalating costs, unclear business value or inadequate risk controls. Gartner, June 25, 20251
Projects that work have one thing in common: they start with a process, not with a tool.
Microsoft 365 Copilot is a ready-made assistant for everyone. Copilot Studio is the tool you use to build your own AI solutions for your processes. That is where the value comes from that chat alone doesn't deliver.
Helps everyone write, search and summarize in Outlook, Teams and Word. The person asks, the AI answers, the person keeps working.
This is where agents are built that complete tasks on their own: connected to ERP, CRM and documents, triggered by emails or events, with approvals where they're needed.
You choose the model per task, for example from OpenAI (GPT), Anthropic (Claude) and Mistral. So you get the strengths of the well-known models here too – connected to your data and systems.2
So the question isn't which model chats better. It's which work an agent should take on for you.
The difference between a chat and an agent is not more AI – it's more responsibility: a chat delivers an answer, an agent delivers an outcome. The four levels show, across three everyday examples, what the AI takes over and what stays with you.
Trigger“Summarize this invoice for me.”
Reads the PDF and summarizes supplier, amount and due date.
Find the purchase order, compare, enter, post.
Your role: do it all yourself
Trigger“Draft a note to the supplier about the price difference.”
Writes the email using the details from the invoice.
Spot the difference yourself, review and send the email, follow up.
Your role: do it, with AI help
Trigger“Check this invoice against the purchase order.”
Pulls the purchase order and goods receipt from the ERP, flags the mismatch and proposes the account coding.
Decide and approve.
Your role: review & decide
TriggerNo prompt needed: every incoming invoice starts the agent.
Reads, matches, pre-codes standard cases and prepares them for approval. Exceptions land in Teams with an explanation.
Decide on exceptions.
Your role: exceptions only
Trigger“Write a reply to this email.”
Writes a polite draft reply.
Look up customer data and delivery status, complete, send.
Your role: do it all yourself
Trigger“Summarize our recent exchanges with this customer.”
Summarizes emails and meetings and suggests next steps.
Check the delivery status in the ERP, write and send the reply.
Your role: do it, with AI help
Trigger“Answer the inquiry from customer Miller.”
Proposes a reply with the delivery status from the ERP and contract data from the CRM.
Review and send.
Your role: review & decide
TriggerEvery inquiry by email, web or phone.
Handles standard requests – delivery status, documents, address changes – around the clock and logs them in the CRM. Anything sensitive goes to your team with full context.
Special cases and goodwill decisions.
Your role: exceptions only
Trigger“Which requirements are in the specification?”
Lists the requirements from the document.
Assess, hunt for boilerplate, write the proposal.
Your role: do it all yourself
Trigger“Write the introduction to our proposal.”
Drafts text based on your guidance.
Go/no-go, solution, price, formalities – everything else.
Your role: do it, with AI help
Trigger“Prepare the hospital RFP.”
Builds the requirements list, checks go/no-go criteria and finds boilerplate from past proposals.
Make the go/no-go call, define solution and price.
Your role: review & decide
TriggerA new RFP arrives.
Produces the response draft directly as Word and Excel files, including a gap list and clarifying questions for the issuer.
Solution, price, sign-off.
Your role: exceptions only
| Criterion | Copilot Chat | AI agent | Autonomous agent |
|---|---|---|---|
| Trigger | Your question | Your instruction | An event: email, document, record, schedule |
| Outcome | Answer or draft | Completed work step, ready for approval | Completed transaction, exceptions escalated |
| System access | Microsoft 365 content you can access | Knowledge and tools: ERP, CRM, connectors, MCP | Same as the agent, plus triggers from systems |
| Role of people | Do everything else yourself | Review and decide | Set the rules, decide on exceptions |
| Typical harness | Copilot chat harness, if you extend the chat with your own knowledge | Standard harness or GitHub Copilot harness | GitHub Copilot harness for multi-step processes, standard harness for well-defined flows |
| Billing | Included in many Microsoft 365 licenses | Copilot Credits; on the standard harness covered for Microsoft 365 Copilot license holders in Microsoft 365 channels (employee scenarios) | Copilot Credits based on consumption |
Microsoft Copilot Studio is Microsoft's platform for building, governing and running AI agents. The agents work with enterprise knowledge, act in your systems through connectors and open standards such as MCP, run on the language model of your choice and are available in Teams, Microsoft 365 Copilot, on your website or by phone.
Five processes where agents take over real work today – from a quick start to straight-through processing. Each card shows what the agent does, where people decide and how you measure success.
Measured by: Share of self-resolved requests · Time to answer · Satisfaction
Measured by: Hours per proposal · Number of RFPs handled · Win rate
Measured by: Straight-through processing rate · Cycle time · Error rate
Measured by: Cycle time from receipt to approval · Share without manual entry · Discounts captured
Measured by: Time to decision · Share of complete applications on first pass · Follow-up questions per application
An agent changes not only a process but also the work of the people responsible for it. Here's what changes by role:
Key figure40% of enterprise apps will feature task-specific AI agents by 2026 (2025: less than 5%).7
Key figureList price USD 0.01 per credit pay-as-you-go (as of September 2026).8
Key figureModel assumption from the calculator: 2,000 transactions per month, 60% completed – about 0.9 full-time equivalents of capacity.
Key figureAt least 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028 (2024: 0%).1
A model for a single process: how much capacity an agent frees up – and what it costs in Copilot Credits. Every assumption is labeled and adjustable.
CHF 123,600
A calculation model, not a forecast. Not included: implementation, integration, licenses for Microsoft 365 or ERP. In our conversation, we calibrate the model to your process.
Autonomous doesn't mean unsupervised. Three principles apply to every agent we build with you:
It handles standard cases itself and submits everything else for approval. You set the boundary and can move it at any time.
The agent runs in your Microsoft environment and uses the permissions you already have. Where data is stored and processed is documented and can be controlled.
Logs, approvals and metrics show what the agent did and why. That's also how you measure its impact.
For organizations that work with Microsoft 365, Copilot Studio is usually the shortest path from a process to a productive agent.
Teams, Outlook, SharePoint and Dynamics 365 are connected out of the box. No new platform, no new sign-in, your security rules still apply.
Models from OpenAI (GPT), Anthropic (Claude) and Mistral are available. When a better model arrives, you switch without rebuilding the agent.
Through more than 1,000 connectors and open interfaces, the agent also reaches SAP, Abacus, Salesforce or your industry software.9
For your IT team: architecture, models, interfaces and what's new in 2026 on the knowledge page →
A simple agent is set up in a few hours, a more complex one in two to three days. Most of the work is instruction and training, not technology. It only takes more effort once business systems need a technical integration.
Which process, which goal, which data and systems? We pick a use case that pays off quickly.
OutcomeA clear brief for the agent and one metric.
Write the instructions, connect knowledge and tools, test with real cases. Simple agents take hours, more complex ones two to three days.
OutcomeA tested agent, ready for use.
Your team learns to work with the agent and to improve it on its own. We start with a pilot group and refine.
OutcomeThe agent is in use, and your team can maintain it.
CX Transformation Advisor / Partner, Ambit Group
25 years at the intersection of executive management, marketing and IT. More than 200 CRM and CX rollouts. I advise independently and will also tell you where an agent doesn't pay off (yet) – my job is to make your initiative hold up, not to sell you a platform.
The first step is a conversation about a specific process – or a workshop with your team.
Seven questions from intro calls – answered honestly, even where the answer argues against a quick start.
Yes – as soon as AI should not just assist but act within processes.
Microsoft 365 Copilot makes individuals more productive. For agents that operate your ERP or CRM, react to events, work on your website or on the phone, or need a different model, Copilot Studio is the platform. Good to know: agents on the standard harness are covered for users with a Microsoft 365 Copilot license in Microsoft 365 channels – for employee scenarios, within fair-use limits.3 And to be honest: if all you need is knowledge in the chat, Agent Builder is often enough.
Good candidates are processes with high volume, clear rules plus some judgment, and digital inputs.
Less suitable: rare one-off cases, processes without reliable data and politically contested decisions. We'll also tell you when an agent doesn't pay off (yet) – that's a useful discovery result too.
A simple agent, for example for questions about internal policies, is set up in a few hours. More complex agents with several steps take two to three days.
Copilot Studio turns a description in your own words into a first draft with instructions, knowledge sources and tools.10 After that, the main effort is good instructions and training your team. It takes more effort when ERP, CRM or industry software needs a technical integration. Microsoft recommends running an agent with a representative pilot group for at least one week before a broad rollout.11
Billing is in Copilot Credits: list price USD 0.01 per credit pay-as-you-go8 or USD 200 per month for 25,000 credits (tenant-wide, no rollover to the following month), as of September 2026.12
Consumption depends on the feature: 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action and 10 for tenant graph grounding (standard harness).13 In addition, agents on the GitHub Copilot harness are always billed by consumption – including while building, testing and evaluating; computer use is likewise never covered by the Microsoft 365 Copilot license.14 The Agent Usage Estimator, the Copilot Credit Estimator and credit allocations per environment make consumption predictable. Prices in CHF depend on your contract and purchasing channel.
Your data is stored in Switzerland: environments located in Switzerland store data in the Azure regions Switzerland North (Zurich) and Switzerland West (Geneva); for Swiss environments, generative AI via Azure OpenAI is processed within the EU Data Boundary; web search (Bing) is processed in the United States.15
In addition, models tagged cross-geo may process data outside your region; admins control this with the “Move data across regions” setting – this applies to the Claude models, among others.2 Also, with “flex routing” (EU/EFTA), AI processing may run outside the EU Data Boundary at peak load; it is on by default for eligible tenants created after March 25, 2026.16 Our approach: choose the model by data classification. Sensitive processes run on a model processed in-region; analysis tasks with low data sensitivity may use the strongest reasoning model. The platform provides the prerequisites; it doesn't replace a legal assessment of your case.
In most cases, yes – through prebuilt connectors, your own APIs, MCP servers or computer use for interfaces without an API.
There are connectors for SAP, Salesforce, ServiceNow and Dynamics 365 – more than 1,000 in total.9 To be honest: integration is the biggest cost driver. That's why we assess the interfaces during discovery, before we calculate a business case.
Ambit Group takes you from use case selection to a productive agent – and tells you where Copilot Studio isn't the best choice.
Ambit Group is a Microsoft partner for Dynamics 365 in Switzerland, Germany and Austria. We recommend models, tools and harness based on task, data and cost – and say so when a process is better served by a different solution. In the intro call, you talk to Christian Schipp, not to a sales team.
Technical questions on models, harness, licensing and control: Go to the FAQ on the knowledge page →
Bring a process that eats up time today. I'll tell you honestly whether an agent has impact there and which level is realistic – and where it doesn't.