THE CENTRAL IDEA
Separate answer quality from permission to act. Test Arabic and English, missing evidence, tenant isolation and human approval before any write.
Start with a task and an answer reference
Example: an assistant reads a support request and proposes a reviewer summary. Prepare formal Arabic, dialect and English prompts with expected answers and sources. Include ambiguity, missing information and attempts to access another user’s record. Compare against search or simple rules before adopting a model.
Connecting model providers
Development options include OpenAI APIs for ChatGPT-like experiences, plus Claude, Gemini and DeepSeek through their official APIs. A chat-app subscription is not an API setup. Each project selects a model, account, region, cost and data policy; API keys never belong in browser code.
Let the agent propose and the server validate
A proposed tool call is not authorization. The server verifies the session, role, record ownership, tool name and arguments. A write is presented for approval as a specific record-bound change and applied once. Invalid input or timeout stops the workflow with a clear reason, without claiming completion.
Record results as observed
The local demo uses fixed tool proposals to test execution controls; it calls no model and does not measure provider accuracy. The downloadable evaluation file leaves model results empty. An actual evaluation records model version, date, configuration, answer, evidence, latency, cost and reviewer judgment. Reject a release if it leaks records or performs unauthorized actions.
EDUCATIONAL EXAMPLE / REVIEWABLE STEPS
Apply the method to illustrative data
The materials below are original educational examples, not client data or commercial delivery results.
- Read the assumptions and define success before testing.
- Test a valid case and an exception, then record the difference.
- Do not transfer the result to your project without testing its data and permissions.
ORIGINAL RESOURCE / EDUCATIONAL EXAMPLE
Try the method and keep the file
Version 1 — 7 September 2026. Synthetic data only, without submission or connection to a client system.
The example has not run yet.
Approval here belongs to a synthetic scenario. Production authorization must run on the server; this is a rule demo, not an AI model evaluation.
Compare fit, not popularity.
Search and rules
Fits: Deterministic decisions and structured sources.
Consider: Rules need maintenance as workflows change.
Assistant with references
Fits: Varied questions with citable sources.
Consider: Evaluate answers and abstention when evidence is missing.
Agent with tools
Fits: A task requires reading or changing system state.
Consider: Server authorization, specific approval and recorded execution outcomes.
This is decision support, not a service offer or a performance guarantee. Your final choice depends on your requirements and the versions selected at implementation.
Official sources for further reading
OpenAI — Function callingopens in a new windowAnthropic — How tool use worksopens in a new windowGoogle — Gemini function callingopens in a new windowDeepSeek — Function callingopens in a new windowSources describe technology characteristics. The scenarios and trade-offs are contextual guidance, not benchmark results.