Blog
The Thero blog
The founding team of Syntriq writes here about using AI in mid-sized companies. Every article is published in German and English.
All articles
Free Agent or Fixed Workflow: When an Agent Should Decide for Itself
“Frei” or “Ablauf”: every entry in the agent list carries one of the two labels. We walk through four criteria that show which task can afford freedom, which one needs fixed steps, and when a hybrid is the most honest answer.
Fewer Hallucinations at Work: Why Mandatory Sources Beat a Better Model
An assistant that sounds certain and is wrong costs more than one that admits a gap. Why mandatory sources, one citation per statement and the answer “not in the documents” beat a better model, and how to catch four kinds of error in everyday work.
How Enterprise Search Actually Finds Things: Chunking, Embedding, Reranking
When an AI disappoints inside a company, the cause is rarely the language model. It is usually the search in front of it. A walk through the five decisions that determine whether the right passage is found, with examples of how searches fail.
Audit and Data Protection Will Ask: What an Audit Log Can Prove About AI Use
Sooner or later internal audit asks who started which AI process and when, and data protection asks whose data was involved. A catalog of questions, the chain of evidence from permission to approval to log, and an honest list of what an audit log does not do.
AI Rollout in 90 Days: A Roadmap for Mid-Sized Teams That Also Plans the Exit
Thirteen weeks, three phases, three decision points: this roadmap shows how a mid-sized team can introduce AI without promising results nobody can guarantee. With roles, an example calculation of the effort and clear criteria for stopping.
Works Council and AI: Why the Early Meeting Is the Fastest Route
A pilot is running, the department is enthusiastic, and then one email from the works council stops everything. Where AI typically touches co-determination, what a council sensibly wants to see, and how a works agreement sets the frame.
The Data Traffic Light: Which Data Belongs in Which AI?
A ban does not answer the data question, it moves it to private accounts. We show a three color traffic light and a decision tree that tell employees in seconds which data may go into which AI, and how permissions and approvals enforce the rule in daily work.
What an Enterprise AI Really Costs: A Calculation Model Instead of a Price List
The price per seat is only one of six cost lines. This article shows the cost blocks of an enterprise AI, the formulas behind them and a worked example with round assumptions in which the offer with the higher seat price turns out cheaper.
Connecting the Systems Is the Real Project: Where Time Disappears When AI Meets Company Data
The AI feature is quick to demonstrate, connecting the systems is not. Six stations from sign-in to source data quality, an example calculation for the first sync, and the question of whether a vendor shows the state of its connectors openly.
Documents AI Cannot Read: Check Your Data Before You Connect Anything
A scan without a text layer, a table without column headers, a file name like “final_v2_new”: to an AI these are gaps nobody notices. We show how to check data quality before you connect anything, and what an assistant should report instead of failing silently.
Approval Is Not a Formality: The Four-Eyes Principle for AI Actions
An approval only protects when someone really looks. When a workflow step needs one, how to design title, deadline and approvers, what the four-eyes principle achieves, and why too many approvals produce the opposite of safety.
The Inbox: When AI Sorts Requests and Humans Decide
Letting AI sort incoming requests hands over preparation, not a decision. We show how a workflow routes requests to sales, service, purchasing and accounting, when it hands back to a person, and how to measure misrouted requests.
Why Wikis Go Stale and What Changes When an AI Makes the Gaps Visible
Wikis go stale not out of neglect but because nobody can see which page is no longer true. An assistant that answers from your documents makes gaps and contradictions visible. With owners and review dates, upkeep becomes a small, steady effort instead of a big project.
AI in Procurement: Comparing Quotes, Questioning Framework Contracts
Two quotes, one framework contract with two amendments and a deadline a few days away: a fictional example shows how an assistant lays out prices, tiers and notice periods with evidence, where PDFs mislead, and why the decision stays with procurement.
The Error Code Is Only the Symptom: Fault Reports as the Memory of the Plant
The same error message can have two different causes. Using the fictional example of filler FL-200 and error E-217, this article shows how fault reports become earlier causes and actions, in which order to check, and why the decision to intervene stays with the technician.
Support Replies With Sources and Approval: From Inbox to a Checked Draft in Four Steps
AI does not speed up support by skipping the check, but by making the check cheap. How a reply draft with sources and approval comes together in four steps, which topics should never go out automatically, and what to measure.
Prove the Value, Do Not Claim It: The Metrics an Enterprise AI Needs
“People like it” is not an answer to the question of what the AI has delivered. Five metrics, a before value recorded ahead of the start and an honest line between measured, estimated and judged turn a feeling into evidence that finance can recalculate.
Who Gets to Build Agents? Governance for Self-Service AI That Does Not Slow Everyone Down
When business teams build AI agents themselves, someone has to decide who may do what. A model of four roles, plus sharing, draft and publish, versions, tool rights per agent and a deputy, so approvals do not stall when the owner is on vacation.
Model Choice and Hosting: Criteria Instead of Dogma
Open or proprietary, run by you or by a provider, one model for everything or several: the argument often replaces the decision. This article lays out the criteria, two example calculations on speed and cost, and a way to keep models replaceable.
No Answer Means No: Why Our AI Never Sends an Email Without Approval
An assistant that sends the wrong email creates a mistake nobody can take back. Why we deliberately made TheroAI slower at the send step, and why five minutes of silence counts as a no.
One Policy That Applies Everywhere: How We Made More Than 300 AI Tools Governable
More than 300 individual AI tools sound like a loss of control. We decided early that the rules live in exactly one place: one organization-wide policy per tool, enforced on every call, whether it comes from chat, from an agent, or from a workflow.
From Question to Cited Report: How Our Research Mode Works
A web search returns a title, a URL and a snippet, and for a real question that is not enough. We walk through one complete run of our research mode: from the visible plan through parallel search and exactly one reflection round to the report with numbered citations.
The AI Only Sees What You're Allowed to See: How We Mirror Permissions into Retrieval
In our article on Copilot's sovereignty gap, the intern got the salary answer from a forgotten file, and building our own retrieval across 19 connectors confronted us with exactly the same question. Why we rejected post-hoc filtering and instead mirror source permissions into the retrieval query itself.
TheroAI, Microsoft Copilot and Langdock: criteria for your comparison
Compare data sources, permissions, deployment and total costs against your own use case.
AI pilots: how to evaluate value, data access and implementation
A guide to a verifiable AI pilot with real tasks, tested permissions and a complete cost assessment.
AI for mid-sized companies: choose a solution around your tasks
Questions to settle before buying an AI platform: tasks, data, responsibilities and costs.
The Public Sector in Transition: Why Open Source Becomes Mandatory (And Microsoft Isn't the Answer)
Germany has a goal: digital sovereignty. But reality in government offices often looks different. Dependencies on US vendors paralyze administration. TheroAI is the bridge to sovereign AI in government.
AI for Tax Advisors & Lawyers: Real Automation Despite § 203 StGB
It's the sword of Damocles hanging over every German law firm: § 203 StGB. While the free market uses AI, professionals often stand still out of fear. We solve this dilemma with local inference technology.
Onboarding on Autopilot: "The New Colleague Asks the AI Now"
Remember your first day of work? The excitement. All the new faces. And most importantly: the fear of asking "stupid questions". Onboarding is still often inefficient and frustrating.
The "Brain Drain" Is Here: How to Digitally Preserve Knowledge of Retiring Boomers
In the next five years, more experienced employees will retire than young talent will replace. When Herbert leaves after 35 years, he takes your department's operating system with him. We call it the "Brain Drain".
Cloud vs. On-Premise vs. Hybrid: The Architecture Decision for 2026
For years the dogma was 'Cloud First'. But in 2025/2026 the wind is turning. Facing AI regulation and exploding cloud costs, own infrastructure is experiencing a renaissance – the so-called 'Cloud Repatriation'.
"German Bureaucracy Agents": How We Tame Administrative Monsters
We have AI models that pass the Turing test, but German offices still have highly qualified specialists typing invoice numbers into spreadsheets. The reason? Most AIs are chatbots. What we need are agents.
RAG Instead of Fine-Tuning: Why You Shouldn't Train Your AI Model
"We need to build our own AI and train it with our company data." This sentence comes up in almost every first conversation. It's an understandable impulse, but in enterprise IT, it's often the beginning of an expensive misunderstanding.
EU AI Act & GDPR 2025: Is Your AI Strategy Actually Legal?
The grace period is over. Since mid-2025, the strict rules of the EU AI Act for "General Purpose AI" have been in full force. Combined with stricter GDPR interpretation, German executives face a new reality.
The "Sovereignty Gap": Why Microsoft Copilot Isn't Enough for German SMEs
Microsoft Copilot is here. The colorful button in Word and Excel promises magical productivity. But for German executives, a dangerous gap opens up: The "Sovereignty Gap".
Shadow AI: The Silent Security Gap on Your Employees' Laptops
It's Friday, 4:30 PM. Your sales manager urgently needs to finalize the proposal for a major client. He opens the browser and copies confidential data into a public AI tool...
New articles by email
We send you new articles as soon as they are published. Unsubscribe at any time.