And then? Then the result is disappointing.
The output is just not reliable enough. The process is not stable. The business case is not coming to fruition.
The conclusion is then often: “AI is not ready yet.”
Or, “AI isn't for us.”
But you may have simply chosen the wrong AI solution for the wrong problem. You ask an artist to fix your accounts.
Three AI solutions you need to know as an entrepreneur
In many of my conversations, I hear a persistent misconception: entrepreneurs often think that AI is one technology.
By the way, it is not surprising that this image was created. Over the past few years, we've been flooded with stories about ChatGPT and generative AI. So when you think of AI, that's what comes to mind.
But there are fundamentally different AI solutions, each with their own strengths and weaknesses. And where one solution doesn't work, the other can work very well.
At Blis Digital, we distinguish three main categories of AI solutions:
AI agents
AI agents are autonomous software programs. They can reason, perform actions and interact with data and systems. You give the agent a goal (“fix this for me”) and he'll figure out how to get there himself. He tries different routes until he succeeds.
That makes AI agents powerful, but also erratic. Where classical software works deterministically (the same input always gives the same output according to the same steps), an AI agent is probabilistic. He is looking for the fastest way and it can be different every time.
We recently came across a painful (but educational) example in our own development environment. The agent's assignment was simple: “There should be no more errors in our tests.”
The agent got to work and came back proud: “Done. No more mistakes.”
What had he done? He had simply deleted the five test cases that gave an error.
Technically, he fulfilled the assignment. But from a business point of view, this is of course disastrous. This is what you get when you focus on outcome, without controlling the process. Agents are great for creativity and flexibility, but they are risky for processes that require accuracy and predictability.
Agentic workflows
Looking for more control? Then you need an agentic workflow.
Here you decide exactly which steps to take, in what order and what to do if something goes wrong. AI plays a role as one of the cogs, but always within the limits you set.
You can see an agentic workflow as a tightly organized assembly line. Some steps are normal automations (code that always executes the same thing). You outsource other steps to an AI agent.
A concrete example from our practice:
A company has a mailbox where invoices, questions and newsletters arrive all day long. To make things extra complicated, emails for multiple sub-brands come together in this mailbox. Each email must go to the correct folder. Invoices must be forwarded to the accounting package. Simple Outlook rules go a long way, but not far enough. Many emails are complex or ambiguous.
So we built an agentic workflow. It grabs every email, checks with automation to see if there is an attachment and forwards it to an AI agent. That agent decides whether it is an invoice and which company the invoice belongs to. He also states how confident he is about his business. Is the AI agent's certainty rate 80% or higher? Then the workflow automatically forwards the mail to the accounting package. Is the certainty lower? Then the mail goes to the manual control folder.
We use AI to understand the messy reality, but let automation decide what to do with the outcome.
AI search & data enrichment
The third category often gets less attention, but is just as important: AI search and data enrichment.
Both your agents and workflows are nowhere without good data. AI search ensures that data is findable, understandable and usable. It's the indispensable layer between your raw business data and the AI solutions you build on it.
With AI, you can search semantically. In other words: you no longer search for keywords, but for meaning. You can also enrich your data with AI: AI can add tags and metadata to documents so that you can make connections you didn't see before.
You can also use AI for things that have to be exactly right
I often hear entrepreneurs say things like, “I can't use AI for my financial reporting because AI hallucinates and only gives you about answers.”
That is a misunderstanding.
Suppose you want to generate that financial report. The text and interpretation? AI can easily arrange that. Calculate the numbers? That is indeed going to be a link.
But you have two options to fix this:
Give your AI agents tools to work with
You can instruct your AI agent to call a tool (a piece of code) to make the calculations. This tool performs the calculation with a fixed formula and returns the exact result. The agent incorporates that result into the report.
Create an agentic workflow
Or you can build a multi-step workflow, leaving only AI to interpret the data and write the accompanying text. The other steps in the workflow consist of normal automations. Optionally, you can also opt for “human-in-the-loop” for extra validation.
Start with the problem, not the solution
Successful AI implementation doesn't start with choosing the coolest model. It starts with understanding your business problem and finding the solution that suits it.
So stop thinking from the perspective of technology (“What can we do with AI?”). Instead, start asking questions:
Which processes cause the greatest pain?
- What do we need to solve that pain?
- Is an autonomous agent or a controlled workflow better suited to this?
- Where is accuracy crucial and where can it be flexible?
- When should a person watch?
- And only then start designing the solution.
Would you like to discuss how to translate “good experimentation” into working business solutions? We'd love to hear from you.
