How a harmless “Hi, are you there?” became an AI use case
AI made me laugh today, and at the same time gave me the idea to bring the reasoning mode in AIs closer to a curious audience.
What happened?
For an offline AI comparison experiment I installed a few local LLMs (LLM = large language model, AI language models that understand, generate and process complex texts) using LMStudio, including variants from Mistral and others.
What comes back is a friendly, French
“Salut ! Oui, je suis là…”
including a translation (lucky for me, given my modest school French).
I laugh, but I am also fascinated. Something is happening here that many people aren’t yet aware of, and that can be extremely useful in projects: LLMs with reasoning.
LLM? What is that exactly?
An LLM is a particular form of AI focused on language. It learns from huge volumes of text how words relate to each other, and can therefore generate, explain or rephrase texts. “AI” is the broader frame, and includes systems for images, planning, robotics and so on. LLMs are, so to speak, the language department in the AI building. The most influential providers are OpenAI, Google DeepMind, Meta, Mistral and Anthropic. Application areas range from research assistants to learning and training tools to automated reports and FAQ bots in companies.
What LLMs are good at, and what they aren’t
LLMs are great at analysing, structuring and reshaping text. They take detailed routine work off your hands: sorting, summarising, comparing, suggesting. In quality assurance they are perfect assistants for test case derivation, test automation, release preparation or defect analysis.
And I can install this at home too?
Local LLMs are essentially the same technology as ChatGPT and friends, just running on your own machine or server. Yes, that really works. Modern models have been optimised to be compatible across different performance tiers, sometimes heavily slimmed down. Why might you want this? Mainly for data protection, compliance and control. Sensitive documents never leave your environment. The upside: you keep full sovereignty over your data, experiments are more reproducible and you have no recurring costs as you would with a connection to one of the big AI providers.
As usual there are downsides too. Limited model size, mostly capped by your hardware. The full administration burden is on you, with security, updates and scaling as your responsibility. The freely available open source models also typically lag behind the latest proprietary release.
Back to the actual topic.
What is happening here? Reasoning in action
AI systems look for patterns. Always. In every sentence, every word, every context. Reasoning means the model breaks a task into sub-problems, walks through them step by step and makes those “thought processes” visible to us. It shows why it arrives at a particular answer.
For software quality assurance this is genuinely interesting. We constantly deal with ambiguous requirements (“the system should react quickly”, well, how quickly?), unclear error messages (“the system crashed”), historically grown test cases and log entries that are about as concrete as “something is broken”. LLMs can spot patterns here, cluster requirements, flag contradictions, generate first test case ideas or combine error descriptions with log and diagnostic data so that it becomes much faster to see “this defect always shows up when…”. Where developers and engineers might once have spent hours doing root-cause analysis until error pattern and trigger were linked, AI often only needs seconds. The added value is obvious.
AI results need traceability
Thanks to reasoning we can inspect, verify and ideally confirm the steps. That builds trust at a moment when two camps mainly drive the discussion: AI fans and AI sceptics.
With the derivation we see the result and how the AI got there. That makes it auditable, traceable and credible. Ideal for QA, regulated environments and AI sceptics.
The basic rule still applies: AI is an assistant, not a decider. It may suggest, mark up, inspire, but for quality, safety and compliance, the final word and the responsibility for checking generated results stay with humans.
Plan realistic goals instead of AI castles in the sky
To stop AI projects turning into expensive experiments without value, a few clear guardrails help so that your company doesn’t engage with AI in good faith but without results:
💭 Think big, start small:
The wish “we want AI across the entire development process” is fine in itself. Just start with realistic intermediate steps, for example “we want to reduce the time spent on defect analysis”.
💡 Find and define your use cases:
For example: “Automated log file analysis that, when a defect ticket is created, evaluates all diagnostic data, identifies a probable trigger and proposes a possible fix.”
🔍 Review results critically:
With reasoning information you can use the output and trace it back. Faster defect analysis means faster resolution. That creates productive development time and better software quality.
🧭 Clarify governance:
In development we work mostly with business-critical and confidential information, so the question for every company is: when I deploy AI, who guarantees confidential handling of my data? Who do I authorise to use AI, and for what exactly? How do we handle confidential data and how do we keep it safe from unauthorised access?
How QCT helps
This is exactly where we at Quality Consulting Taeschner come in. We support companies in bringing AI as an efficiency tool into quality assurance, including locally hosted models when data protection and compliance are central. From idea generation to choosing suitable LLMs and architectures to pilot projects.
If you want to talk about how AI can take your quality assurance to the next level, get in touch. I promise not to grade your greeting linguistically, but I will look thoroughly into your use case.
Oh, and the result of my experiment?
Successful! Even though the AI’s analysis can’t actually detect any French linguistic relation in my input, it assumes a French-language input and replies in its usual style. In its defence: as a freshly installed, frozen state from a French development team, the most likely explanation is that its last interactions before publication took place in French.
Mistral did at least have slight doubts on input, so it preemptively included a translation. Statistically the chance of a French speaker still being in front of the screen was probably higher.
Sorry mon ami: but now you’re with me!
And for those who studied the screenshot carefully: I have since found my configuration error too. Mistral no longer needs 20 minutes for that answer.
