Brilliants

Brilliants AI solutions integrated into existing systems—or built from the ground up as intelligent systems for different industries.

"We want our AI to know our internal policies."We hear this often. The instinct is to train the model on company documen...
01/08/2026

"We want our AI to know our internal policies."

We hear this often. The instinct is to train the model on company documents. But that's usually the wrong fix.

Here's the distinction: fine-tuning changes how a model behaves — its tone, its format, its habits. RAG (retrieval-augmented generation) changes what a model knows, by handing it the right document at the moment of the question.

Suppose a tendering team asks an AI about payment terms. Fine-tuning won't help if the policy changes next month. What they need is the current policy document, retrieved and cited, every time someone asks.

Training teaches a way of speaking. Retrieval delivers the manual, page by page, on demand.

Before you fix your AI's "behavior," ask a different question: is this really about how it answers — or about what it's missing?

Type the exact same sentence into a chatbot twice. You can get two different answers back. Why?Suppose a business owner ...
29/07/2026

Type the exact same sentence into a chatbot twice. You can get two different answers back. Why?

Suppose a business owner asks a chatbot to draft an email about a delayed order. Before the model even sees that sentence, it breaks the text into small pieces called tokens, and works within a context window — the amount of text it can consider at once.

A hidden system prompt may already tell it to sound formal, or apologetic, or brief. And a setting called temperature decides how predictable or varied the wording will be.

Same request. Different hidden settings. Different email.

None of this is magic — it's design, most of it invisible to the person typing.

Has a chatbot ever given you an answer that surprised you? What do you think caused it?

An AI model can give you an answer. On its own, that answer isn't a decision record.This week we sketched out a question...
29/07/2026

An AI model can give you an answer. On its own, that answer isn't a decision record.

This week we sketched out a question on the whiteboard: what actually needs to be captured when AI assists a business decision?

Not just the output. The model and version that produced it. The inputs it used. Its confidence, meaning how certain the system was. Whether a person reviewed it. Whether they approved it, corrected it, or overrode it. And the final outcome.

Imagine a hypothetical approval workflow: an AI recommendation, flagged for review, edited by a person, then acted on. Every one of those steps is part of the real record, not just the last one.

Where would you place the human checkpoint in your own workflow?

The best place for AI may be one step, not the whole workflow.We use a simple decision tree:→ Is there meaningful fricti...
29/07/2026

The best place for AI may be one step, not the whole workflow.

We use a simple decision tree:

→ Is there meaningful friction, risk or unclear information? If not, make no intervention.
→ Should the same input always produce the same output? Use deterministic software.
→ Do fixed steps repeat or move information between systems? Use basic automation.
→ Does the step require interpreting language, context, ambiguity or exceptions? Consider AI.
→ Does the decision carry meaningful consequences? Preserve human approval.

Using AI for predictable, rule-based work can add model cost, latency, inconsistency, maintenance and review burden without improving the decision.

Rules for certainty. Automation for repetition. AI for interpretation. Humans for responsibility.

Which repetitive process might be simpler than it currently appears?

Not every decision deserves the same level of trust from an AI system.Some are low-stakes: flagging a discrepancy betwee...
28/07/2026

Not every decision deserves the same level of trust from an AI system.

Some are low-stakes: flagging a discrepancy between two invoices, sorting a document, spotting a duplicate entry. Light review is enough. Let it move.

Others are high-stakes: approving a payment, releasing funds, committing to a contract. Here, the consequence is hard to reverse. Human approval belongs in the loop, not as a formality, but as the layer where responsibility lives.

The design question isn't "can the AI do this?" It's "what happens if it's wrong, and can we undo it?"

A system that flags an issue and waits can be more useful than one that acts on every judgment call.

Where would you draw the line between automate and approve?

Address

Beirut
12345

Website

Alerts

Be the first to know and let us send you an email when Brilliants posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to Brilliants:

Shortcuts

Share

Category