The perfect triad: generative AI + human judgment + deterministic tools | Xyclos
- Carlos Altamirano
- Apr 24
- 3 min read
It all started with this idea today, which I made to ChatGPT o3:
"LLM's are predictive, and you can't expect consistent or equal answers, as you would expect from an application like Excel, for example, where you enter 1 + 1 and it will always give you 2."
And this is what ChatGPT answered me. o3: That's right: a LLM (large language model) doesn't work like a deterministic spreadsheet. When you type =1+1 in Excel, it runs an explicit mathematical algorithm, and you always get back 2. An LLM, on the other hand, generates the next word based on probabilities learned during training, and those probabilities are "sampled" each time you ask for an answer.
Hence:
There is inherent randomness
The temperature parameter (or top-p ) controls how much variation is allowed when sampling. At temperature = 0, randomness is greatly reduced, but it never disappears completely: details such as hardware, context length, or hidden system messages can influence this.
The output depends on the entire context
If you change a word in the prompt, add older messages, or use a model from a different date, the probability distributions change, and so does the resulting text.
There is no single “correct outcome”
For 1 + 1, there is only one canonical answer; for "explain why innovation is important," there are infinitely many plausible answers. LLMs choose one of many.
Why is it useful to know?
Hybrid Systems Design : Use Excel, SQL, or code when you need fixed, numerical rules; turn to an LLM for drafting, flexible classification, or idea generation.
Correct expectations : When presenting results to users or clients, make it clear that they “may vary” and offer temperature control if needed.
Testing and QA : For critical tasks (e.g. contracts), add human reviews or automated validations.
I continued with this observation:
"LLM's work just like humans; in each interaction we behave differently, we give responses based on the present context, and that varies. I gave you one answer yesterday, and a very similar one today. That's why the combination of AI, humans, and tools (like Excel or a calculator) provide an optimal environment for performing any task. If you leave one of them out, the formula doesn't work. 3 actors...not one, not two...three."
And this is what ChatGPT o3 answered me:
Exactly: when you think about “the triad” — generative AI + human + deterministic tool — each actor brings a complementary type of intelligence:
Actor | Superpower | Weakness | When to use it |
LLM / IA | Creativity, massive text synthesis, translation between formal ↔ natural languages | Not 100% deterministic, risk of hallucinations | Devise, write, classify, “converse” with data |
Human | Deep context, ethical criteria, intuition and final responsibility | Fatigue, bias, limited speed | Define objectives, validate, decide, adjust prompts |
Deterministic tool (Excel, SQL, Python, Power BI) | Mathematical precision, reproducibility, traceability | Little linguistic flexibility, requires explicit rules | Calculation, structured data transformation, auditing |
How they are orchestrated
The human defines the goal and success metric. E.g., "I need a monthly sales report with actionable insights."
LLM translates the objective into solution sketches.
Generate a Power Query prompt or a DAX draft.
Write the narrative report.
Deterministic tool runs the calculations.
Power BI calculates KPIs; Excel simulates scenarios; a script validates each number.
Human reviews/refines/approves.
Adjust filters, correct AI biases, decide the final storytelling.
Iterative cycle (“triple feedback loop”).
Results from the tool are returned to the LLM for new hypotheses.
Human observations become new prompts or formulas.
Key advantages of the 3-player ecosystem
Speed with control : AI accelerates, the tool ensures accuracy, the person governs.
Bias reduction : each actor compensates for the blind spots of others.
Sustainable scalability : You can automate without losing critical human oversight.
The question is not generative AI vs human nor generative AI vs Excel.
Competitive advantage arises from combining the probabilistic mindset of LLMs, the precision of spreadsheets, and human judgment.
Implement the triad and transform your workflows into faster, more creative, and more reliable processes.
Next step, get trained.
At Xyclos Academy we teach you how to master each actor:
Are you ready to work faster, with more creativity and precision?
The trio of generative AI, human judgment, and tools like Excel gives you a competitive advantage you can't ignore. Train with us and master every aspect of this winning combination. Are you already applying this trio in your company? Or will you be left behind?
Contact us and design your perfect learning path.
Comments