Live demo · fictional company
Four ways to make an AI useful inside a company, running live on one invented logistics firm, Delta Freight. Change the prompt, the context or the documents, and watch the answer change. The fine-tuning results come from a real training run.
Four levers
Each lever goes one layer deeper into the model. The deeper it goes, the slower it is to change and the narrower the job it should do.
What the model is told
What the model sees, right now
What the model can look up
What the model is
The lab
Delta Freight is made up: a Cairo logistics firm with a warehouse in 6th of October, drivers, and ten short policies. Everything the model writes below is generated live, nothing is scripted.
A customer complains about a late, damaged shipment. On the left the model gets one line. On the right it gets the building blocks you switch on. Both run on the same model at the same time.
An employee asks the company assistant a simple question. Whether the answer is right depends entirely on what gets put in front of the model. Switch the pieces on and off.
“Can I take next Thursday off, and the Sunday after?”
Ticks come from which context is switched on, not from reading the answer.
Ten company documents. Every question is searched against them, and only the best passages go to the model, with the rule: answer from these, cite them, say so if they don't cover it.
Press “Search & answer”. You'll see which documents it finds and how strongly each one matches.
Only the top passages go to the model, with the rule: answer from these, cite them, and say so if they don't cover it.
The answer appears here.
Ask about annual leave. Watch it find HR-04, build the prompt, and answer with the source.
Edit HR-04 from 27 to 30 days, save, then ask again. The answer changes at once. A fine-tuned model would need retraining.
Maternity leave and international per diem. The fine-tuned model invented answers. Retrieval finds the document.
Ask about the dress code. A good system says the documents don't cover it instead of guessing.
Ask about salary bands as warehouse staff, then switch “Asking as” to HR manager and ask again.
Loss falling to near zero means the model memorised its 40 examples. It does not mean it learned the company.
8 policies × 5 phrasings. Every answer in the same house format: ANSWER / POLICY / NEXT STEP.
Grey is the original model. Amber is the same model after fine-tuning. The line above each pair is what the company's documents actually say.
It learned the format perfectly. Every answer now arrives as ANSWER / POLICY / NEXT STEP. That is what fine-tuning is good at: style, format, terminology.
It learned facts only where it had seen them, and invents where it hadn't. Asked about something it was never taught, it answers in the same official format with a made-up rule and policy code. That is the most dangerous failure: an invention that looks like company policy.
Ask the RAG lab the same “never taught” questions. It finds the right document and cites it, or says the documents don't cover it.
Choosing
In practice they stack. An Enterprise Brain is a good prompt, engineered context and retrieval together, with fine-tuning only where format or cost demands it.
| Which lever for which job | Prompt | Context | RAG | Fine-tuning |
|---|---|---|---|---|
| Follows a fixed format and tone | ||||
| Knows your company's facts | ||||
| Stays right when a document changes | ||||
| Knows who is asking and what's happening today | ||||
| Respects who may see what | ||||
| Shows where the answer came from | ||||
| Says “I don't know” instead of inventing |
● Strong◐ Partly○ No
The Enterprise Brain is all four levers, built on your files, your permissions and your systems. It runs on Orcastly, our own platform, which you can start on for free.
Where to go next