AI Agents in Your Business: Where to Actually Start (and Why 90% Fail)

Grzegorz Olszowka•April 15, 2026•10 min read

AI AgentsOpenAIAzureOllamaGDPRSMBBeginner Guide

An honest pillar guide to getting started with AI agents: the three paths (OpenAI, Azure, local), when each one fits, and the five traps almost everyone falls into.

AI Agents in Your Business: Where to Actually Start (and Why 90% Fail)

By Grzegorz Olszowka • April 15, 2026
#AIAgents #OpenAI #Azure #Ollama #GDPR #SMB

Three weeks ago, I sat down with the managing director of a Berlin trades business. He'd spent a weekend reading LinkedIn posts about AI agents. "I want this too," he said. "Where do I click?"

That question is the reason most AI agent projects in mid-sized companies fail. Not because the technology doesn't work. Because the question gets asked too early.

By the time you finish this article, you'll know three things: what an AI agent actually is, which path you should take, and whether you're even ready. Spoiler: most companies aren't. That's not bad news -- it's an honest one.

The Wrong Question

Gartner forecasts an average of 12 AI agents per company by the end of 2026. That's the number that lands in every other board meeting. The number that kicks off projects. The number that brings us the calls.

What we see is different: three out of four unsupported pilot projects fizzle out within 90 days. Not because of the technology. Because of the question that was asked wrong at the start.

The usual question is: Which tool should I pick?

The right question is: Which process do I understand well enough that I could hand it off to someone else -- human or machine?

If you can't answer the second question, don't build an agent yet. No matter how impressive the demo looks. Answering the wrong question first gets you the most expensive answer.

What an AI Agent Really Is

By 2026, the term "AI agent" has become so loaded that it barely means anything anymore. Marketing departments call every chatbot an agent. That's wrong.

An agent has a goal, not just an answer. A chatbot replies to a question; an agent works toward an outcome ("handle the customer inquiry so that it's either resolved or qualified and handed off to sales").

An agent picks its own tools. It queries the database, writes emails, reads files, escalates to humans. That's the technical difference -- and the hard part.

And an agent plans across multiple steps. "First I'll check the CRM, then verify the invoice, then reply." If step one fails, it figures out a step two.

What that looks like in practice: a chatbot answers "When are you open?" An agent reads an incoming quote request, checks the CRM to see whether the sender is an existing customer, looks up your service catalog spreadsheet, drafts a first response, and drops it into your inbox for approval.

The first takes half an hour to set up. The second takes two weeks -- and after that, it replaces half a full-time role.

Before You Start: The Honest Prerequisites

Before we talk about tools, three reality checks. If you have to pass on one of them, that's not a defeat. It's the reason 90% fail: they start anyway, even when one of these three boxes isn't ticked.

Dieses Thema vertiefen? 32 KI-Rezepte mit Kostenrahmen als kostenloses PDF.

PDF holen

1. The process is documented.

If you can't write down in two sentences what the agent should do, neither can a human. And if no human can, an AI definitely can't. Run the test: on half a sheet of A4, write down what should happen -- input, intermediate steps, outcome. If you stall, the agent is the wrong tool. You need process clarity first.

2. The data is reachable.

The agent needs access to whatever the human currently looks up by hand -- CRM, email, spreadsheets, accounting. If your customer data is sitting in an Access file from 2014, the agent has nothing to work with. Check before the project whether your systems have APIs or at least automatable exports.

3. Your GDPR position is settled.

If the agent works with personal data -- and it almost always does -- you need a legal basis, a data processing agreement (DPA) with the LLM provider, and a data protection impact assessment. That decides which of the three technical paths you can take. Not every path is legal for every type of data.

If you can check off all three, you're ready. If not, invest there first. It's boring, but it's the difference between an agent that runs and a project that fizzles out after six weeks.

The Three Paths -- and How to Pick the Right One

Search "AI agent" and you land on OpenAI right away. That's understandable. It's also the wrong answer half the time.

In 2026, there are three paths to running an agent in your business. They don't differ in intelligence -- the models are often the same -- they differ in where your data is processed and how fast you can get going.

Path	Setup Time	Where Data Sits	Cost/Month	Good For
OpenAI direct	30 min	USA, DPA available	from EUR 20	Pilot, public data
Azure OpenAI	3-10 days	EU, DPA standard	from EUR 50	Production with customer data
Ollama / local	1-3 weeks	Your hardware	One-time EUR 2-5k HW	Sensitive data, full control

Path A: OpenAI Direct

You go to platform.openai.com, set up a business account, generate an API key, add a credit card. You're up and running in 30 minutes.

The catch: the servers sit in the United States. You can sign a data processing agreement (OpenAI offers one for Business, Team, and API customers), and your data isn't used to train models. But a GDPR auditor will ask why you didn't take the European option. For publicly available data, prototypes, and internal experiments, it's still the fastest path.

When do we use OpenAI direct at kiba? For proof-of-concepts that decide within two weeks whether a use case even works. If it does, we move it to Azure.

Path B: Azure OpenAI

Microsoft runs the same OpenAI models in its European data centers, including Sweden Central with the EU Data Boundary enabled. Same API, same models (GPT-4o, GPT-4 Turbo, o1), same costs -- but GDPR-compliant infrastructure out of the box.

The catch: the entry process is bureaucratic. You need an Azure account, then you have to apply for Azure OpenAI access (Microsoft reviews it), then configure a deployment, then set up endpoints and keys. That takes 3-10 days, depending on how fast Microsoft responds.

For production systems handling customer data, this is the default path. Period.

Path C: Ollama or Other Local LLMs

You install Ollama on a server with a GPU, pull down a model like Llama 3.3 or Mistral Large, and run the agent on your own hardware. No external servers, no API costs, no data leaving your network.

The catch: in 2026, local models are no longer far behind cloud models -- but they're still behind. For complex agent tasks, the quality is usually enough. For edge cases, sometimes it isn't. And you need hardware: at least an RTX 4090 for mid-sized models, or a server with 80+ GB of VRAM for the big ones. That's a one-time investment of EUR 2,000-10,000, with no recurring costs.

For us, this is the path for clients running law firms, medical practices, and accounting offices -- anywhere sensitive data should never, on principle, touch a third-party server.

And You Can Settle This with Three Questions

Question 1: Does the agent process data you wouldn't email to a US provider? If yes: Azure or local. If no: OpenAI is fine for the pilot.

Question 2: Is the subject sensitive enough that you'd reject EU servers too? If yes: local. If no: Azure.

Question 3: Can you absorb a one-time EUR 2,000-5,000 hardware bill? If no: Azure, even for sensitive data. At low volume, the recurring cost is below what hardware amortization would run you.

You don't need any more questions for the first decision. The details -- which specific model, which deployment, which configuration -- come in the follow-up articles in this series.

Your First Agent -- Concretely

Say you've picked a path. Azure OpenAI, let's say. What do you build first?

Our recommendation for the first agent has four properties: the use case is small, the human stays in the loop, the failure mode is recoverable, and the benefit is measurable.

A typical entry point -- and the one we use as a kickoff at Fortuna Elektro -- is pre-qualifying incoming inquiries:

Trigger: New submission via the contact form.

Agent task:

Read the inquiry.
Categorize: quote request / support / spam / other.
For quote requests: check the CRM to see whether the sender is an existing customer.
Extract: type of service, scope, timeline.
Draft a reply -- polite, in the company's tone of voice, with follow-up questions where things are unclear.
Drop the draft into the Outlook drafts folder of the rep responsible.

Human step: the rep reviews, edits, sends.

Why this use case? It hits all four properties. Small: one function, not ten. Human in the loop: nothing goes out without sign-off. Recoverable failure: a bad draft is something the rep spots immediately. Measurable benefit: time per quote, before vs. after.

Technically: the agent runs in n8n -- an open-source workflow tool you can either let us host or run yourself on a EUR 15 Hetzner server. The OpenAI node in n8n handles the LLM call. The CRM connection takes 20 minutes to wire up.

We'll publish a dedicated article on this setup -- we'll link it here once it's live.

The Five Traps Almost Everyone Falls Into

Drawn from six months of pilot projects we've seen or supported:

Trap 1: The Agent Without Boundaries

Teams build an "agent that does everything." Result: it does none of it well. Build one agent per clearly bounded process -- five specialists beat one generalist every time.

Trap 2: No Human in the Loop

Approval gets cut for the sake of efficiency. The first hallucinated quote sent to a major customer kills the project. Approval isn't a weakness -- it's insurance.

Trap 3: No Measurement

"The agent will save time." How much? Compared to what? Without a baseline, every success debate is a gut feeling. Spend a week counting manually: volume, duration, quality.

Trap 4: The Wrong Model

Teams reach for the biggest model "because AI." The smaller one is often faster, cheaper, and more accurate for the specific task. Test two models against the same test cases before you commit.

Trap 5: The Privacy Deja Vu

Nobody looped in the data protection officer or leadership. Four weeks after go-live, the email arrives: this thing isn't approved. Back to square one. Settle the approval before you write the first prompt.

Your First Step -- Today

Don't install the tool. Don't grab the API key. Don't order the GPU.

Instead, take 30 minutes, a blank sheet of paper, and a process that stresses you or your team every day. Write out what happens today -- as a step-by-step sequence in ten lines. Underline the steps that play out the same way every time (same kind of input, same kind of output). Those are your agent candidates.

If your sheet has two or more steps like that, you have a sensible entry point. If not, the process is either already optimized or too variable -- in either case, an agent is the wrong tool.

Only then comes the question: OpenAI, Azure, or Ollama? You answer that with the three questions from the matrix above.

When It Gets Complicated

The decision matrix above is the 80/20 grid. For the 20% where things do get tangled -- hybrid setups, industry-specific GDPR questions, integration with legacy systems -- we support SMBs through BAFA-funded AI consulting.

The consultation works out which path is right for your specific case before you invest. No tool sales, no SaaS subscription. Contact: info@kiba.berlin.

What Comes Next

This article is the overview. In the coming weeks we're publishing the deep-dive guides -- one per path:

OpenAI API for businesses: account, DPA, API key in 30 minutes. Including the two settings most people forget and then wonder why their GDPR review fails.
Azure OpenAI in Germany: the GDPR-clean default. The access request, the EU deployment, your first 100 requests.
Ollama on your own hardware: when data must never leave the building. From hardware selection to your first agent running on the network.
Your first n8n agent in 30 minutes. The template agent that qualifies inquiries -- with screenshots, code snippets, and a working starter file.

Which one to read first comes down to the three matrix questions. The right order saves you weeks.