techone --guide=ai-agents
AI agent: software that does a task across your systems
Lead research, an ERP query, data checks across systems. Where rule-based automation stops at fixed rules, an agent begins.
TL;DR
- What it is
- An agent does a task: it finds something out, verifies it, decides, and writes to a system. A chatbot only answers. An agent acts.
- Agent vs rules
- Fixed rules (Power Automate) handle a deterministic flow. An agent handles context and ambiguity. In practice they combine.
- How it connects
- Through MCP to ERP, CRM and registries. Read and write within a defined scope, full audit trail, running on your infrastructure.
- Squad
- Several specialized agents under an orchestrator do more than one universal agent. Easier to tune and control.
- Where to start
- One task with high volume or repetition. Writes limited to approved steps, a human in the loop for ambiguous cases.
What an AI Agent Is (and Is Not)
An agent is defined by what it does. It gets a task, finds out what it needs, decides, and performs an action in one of your systems. It is more than a chatbot that only answers a question, and more than robotic automation (RPA) that runs through fixed steps.
One example for all: a new contact comes in. The agent verifies it against a registry, pulls public information about the company, checks whether it is already in the CRM, and ranks it as a priority or not based on criteria. Between the question and the action there is a decision, and that is the difference.
Agent or Power Automate? When Each
This is a decision almost every project faces. Power Automate and similar tools run a process: when A happens, do B. Cheap, visual, a non-developer can set it up. They work as long as the rules are unambiguous.
An agent steps in where a rule is not enough, because context has to be understood and a call made in the gray zone. Unstructured input, ambiguity, exceptions. In practice the two combine: fixed rules run the flow, the agent makes the judgment in the steps that need it.
Fixed rules are enough
A deterministic process. Clear if X, then Y. Approval by amount, moving a file, a notification. Here an agent is unnecessary and more expensive.
An agent pays off
Context and decisions. Is this value correct? Does this case belong in this category? What about the exception nobody set up? This is where a rule fails.
The point is knowing which part of the process rules run and where the agent decides.
How an Agent Safely Connects to Your Systems
To do anything, an agent has to reach the data. That is what MCP (Model Context Protocol) is for: a standard interface through which an agent reads from and writes to your ERP, CRM or registries. This is where trust is decided.
The key is scope. The agent gets only what it needs: reading reference data, writing only approved records, no access beyond that. Every operation has an audit trail, so you can see what the agent did and when. And it runs on your infrastructure, not somewhere outside. Here is an agent answering an operational question:
>query: how many of item X are in stock and what is the status of order 4471?
- ERP stock 320 units, 80 reserved
- order 4471 in production, due in 5 days
- access check read-only, audit trail
→ answer240 units free, the order will meet its deadline
A Squad of Agents: Why Many Small Beat One Big
The tempting idea is one agent that does everything. In practice the opposite works: several specialized agents, each on its own task, run by an orchestrator.
One agent researches a company, another verifies data against registries, a third prepares the brief, the orchestrator runs them and keeps the order. The reason is practical. A small specialized agent is easier to tune, easier to check, and at each step you know what it does. One big monolith is a black box nobody will fix.
What Agents Actually Do
Real tasks agents handle in operations. One example from sales, where the agent gets a rep ready for a meeting:
Sales prep
The agent researches a company before a meeting, enriches the CRM contact from registries and public sources, and suggests a priority. The rep arrives prepared instead of spending an hour searching.
Querying your systems
Instead of clicking through the ERP, you ask in plain language: how much is in stock, what is the status of an order. The agent looks it up and answers. People stop interrupting each other for a number.
Data checks across systems
The agent compares records between CRM, ERP and registries, finds mismatches and flags them. Payment matching and document intake via EDI run at Tecam.
Document processing
Incoming invoices and orders are read, verified against the ERP and posted. The document automation guide covers it in detail. Marketing data has its own guide.
>agent: new lead, Vacek Engineering Ltd.
- business registry active, 45 employees, 3 years of growth
- web and news building a new plant, an investment signal
- CRM no duplicate, contact is free
→ qualificationpriority: manufacturing, growth, investment signal
Where to Start and What to Watch
Do not start by handing the agent the keys to everything. Pick one task with high volume or annoying repetition. It will show whether this makes sense and let you tune the setup.
The most important decision is scope and control. Keep writes to approved steps only, limit reading, and leave the final word to a person on ambiguous cases. An agent speeds up the routine. Decisions that carry weight stay with people.
When a rule is enough and when AI pays off is also covered in the Automation and AI service.
Frequently Asked Questions
What is the difference between an AI agent and a chatbot?
A chatbot answers a question. An agent does a task: it finds something out, verifies it against your systems, decides by your rules, and performs an action, such as creating a record or flagging a mismatch. Between the question and the result there is a decision and an action, not just text.
Do we need an agent, or is Power Automate enough?
When the process is deterministic (clear if X, then Y), Power Automate is enough and cheaper. An agent pays off where context has to be understood and a call made in an ambiguous situation. In practice they combine: rules run the flow, the agent makes the judgment.
How does an agent safely access our ERP?
Through an MCP interface with a defined scope. You set which modules it can access, typically reading reference data and writing only approved records. Every operation has an audit trail. It runs on your infrastructure, the data stays with you.
Can we put an agent on one task and expand?
That is exactly what we recommend. You start with a high-volume task, verify the result, and only then add more. A smaller start means a faster result, easier control and less risk.
We are building our own AI product and need to connect it to customer ERPs. Can you do that?
Yes, that is a common case. We set up the connection via API or MCP to Dynamics 365 and most other ERPs, with a defined scope and an audit trail. We handle data mapping, validation and error handling so your product gets clean data.
Want to see an agent run?
We have a working solution built on Anthropic Claude and MCP that we adapt to your task. On a call we show an agent at work on a real scenario.
Book a consultation