Guide

What are AI agents?

An AI agent is software that pursues a goal on its own: it plans the steps, uses your tools, does the work, and checks its own output — instead of just answering a single prompt.

Sounds simple. Yet Gartner expects over 40% of all agent projects to be cancelled by the end of 2027. This guide explains what AI agents really are, how they work — and how to recognise the ones that deliver in daily operations, not just in the demo.

What is an AI agent?

A classic AI tool — a chatbot, a prompt in ChatGPT — does one step: question in, answer out. An AI agent works in a loop: goal, plan, act, check, learn.

It breaks a task into steps, reaches into real systems (e-mail, CRM, calendar), runs several steps in sequence, and checks itself before a human sees the result. The difference is not how smart the model is — it is the checking step. That is what decides what you can safely automate.

→ What is agent looping — and why one pass is never enough

AI agent, chatbot, or automation?

	Chatbot	Automation	AI agent
Triggered by	a user question	a fixed rule	a goal
Scope	one answer	a rigid flow	a multi-step workflow
Tools / systems	usually none	hard-wired	picks as needed
Checks its own output	no	no	yes (eval loop)
Handles the unexpected	no	breaks	re-plans

Rule of thumb: answer standard questions → chatbot. Take over a whole process (read the request, draft the quote, file it in the CRM) → AI agent.

How do AI agents work?

AI agents are built from four parts — plus the mechanism that holds them together:

Model (the brain) — a large language model that understands language and proposes steps.
Orchestrator (the planning) — breaks the goal down and routes subtasks to specialist agents.
Memory (the company brain) — versioned knowledge — rules, prices, processes — that every agent reads before acting. Without it, the agent guesses what your company already knows.
Tools (the hands) — access to CRM, e-mail, calendar and internal APIs, often via standards like MCP.

Where AI agents deliver in the Mittelstand

Agents deliver measurably in clearly defined, recurring processes — not everywhere:

Sales — lead qualification around the clock; read inbound requests, enrich them and prioritise in the CRM.
Customer service — standard requests and scheduling outside office hours, handed to humans when needed — case reports show cost per interaction dropping sharply.
Back office & finance — process e-mails, extract requirements, prepare quotes or invoices. One trades business cut time per request from 90 to 12 minutes this way.

→ Adding an agent to any vertical: the 4-week pattern

Why most AI agents fail

Only around 16% of German Mittelstand companies use AI agents — and most projects fail. But almost never because of the model. They fail on poor data quality, underestimated integration costs and missing internal processes.

Gartner names the root cause „agent washing": existing chatbots and RPA tools rebranded as agents; of thousands of vendors, Gartner counts only about 130 as real. Between an impressive demo and an agent that delivers unattended at 6am on Monday lies a gap.

The agents that deliver have three things demos don't show: a company brain that holds company knowledge; an eval loop that catches the slop; and an operator who runs the system. Werkshift builds exactly these three — and runs its own company on them. When we say a closed agent can run your weekly reporting, it is because one runs ours.

Frequently asked questions

Software systems that complete a task on their own: they plan steps, use your tools and check their output — a whole workflow, not just one answer.

Through four parts — model, orchestrator, memory (company brain) and tools — plus an eval loop that checks each result before it moves on.

A chatbot answers one question. An AI agent takes over a multi-step process, reaches into your systems, and checks itself.

Start with one clearly scoped, recurring process, fill the company brain, and build a closed agent with an eval loop — for us, usually in four weeks.

Broadly two: open agents (human in the loop) and closed agents (run on a schedule and check their own work).

Would an AI agent pay off for your operations?

The fastest way to find out is fifteen focused minutes. We name the one or two workflows with the highest leverage — even when the answer is „not yet".

Book an Agent Audit