It writes the query. It does not invent the answer.
That distinction is the entire product. A model that computes your fuel figures is a model that can be wrong about them without anybody noticing — so ours is not allowed to do arithmetic.
Six questions, two refusals
Including the ones it will not answer, because a page that only shows an assistant succeeding is describing a demo.
Ask it something
Two of these six get declined. That ratio is deliberate.
SELECT driver_id, name, events_per_1000km, score
FROM driver_scores
WHERE period = 'last_90_days' AND distance_km > 2000
ORDER BY events_per_1000km DESC
LIMIT 5| Driver | Per 1,000 km | Score |
|---|---|---|
| W. Nasser | 4.36 | 26 |
| E. Kone | 3.89 | 34 |
| Z. Ahmed | 3.59 | 39 |
By events per 1,000 km, W. Nasser. Ranked on raw event counts you would get a different name entirely.
Choices it made
- Normalised per 1,000 km. Without that the list ranks mileage rather than driving.
- Drivers under 2,000 km in the period are excluded as too small a sample.
- "Worst" was read as the behaviour score. It is not a judgement about the person.
2 answered cleanly, 2 answered with the choices it made printed above the result, and 2 declined outright. Every answer arrives with the query that produced it, because a number you cannot check is a rumour with a chart attached to it.
Four steps, and the model only does one
Which model sits behind it matters far less than what it is permitted to touch. Translation is a job language models are genuinely good at. Arithmetic over your operational data is not their job.
This is also why the answers do not drift. Ask the same question twice and the query is regenerated, but the numbers come from the warehouse both times — so two people asking the same thing on the same day cannot get two different figures.
- 01
You ask in words
Arabic or English. No query language, no report builder, no ticket to the analytics team.
- 02
It writes a query
The model's job is translation — plain language into a query against your data. That is the only thing it produces.
- 03
The database answers
The numbers come from the same tables the dashboards read. The model never computes a figure and never guesses one.
- 04
You see the query
Every answer shows the query that produced it. A number you cannot check is a rumour with a chart attached.
What it will not do
This section usually does not appear on pages like this one. It is here because the limits are what make the rest of it usable — an assistant that answers everything is one you cannot trust on anything.
- It does not compute
- If the model were adding up litres you would have no way of knowing when it added them up wrongly. It writes SQL; the warehouse does arithmetic.
- It declines rather than guesses
- Questions the data cannot answer get a refusal and an explanation, not a confident-sounding paragraph. Two of the six examples above are refusals, which is roughly the real ratio.
- It shows its assumptions
- Every ambiguous question involves a choice — which normalisation, which period, which exclusions. Those are printed above the answer rather than hidden inside it.
- It does not make decisions
- Disposal, hiring, discipline and contract questions are yours. It will assemble the inputs and stop there.
Where your questions go
The first question every operations director asks about this feature, and the one most vendors answer in a footnote.
Security and data handlingYour data does not train anything
Not our models, not anyone else's. Queries and results are yours and are not retained for improvement.
It only sees what you see
The assistant runs with your own permissions. A dispatcher cannot ask it for payroll-adjacent data they could not open directly.
Every question is logged
Who asked what, and when, in the same audit log as everything else — visible to you, including for our support staff.
It stays in the Kingdom
Inference runs in-region. Your fleet data does not leave to answer a question about it.
The rest of the platform
Live tracking
Position, speed and ignition every ten seconds.
Routing and dispatch
Plan the day, send it to the driver's phone.
Asset and trailer tracking
Trailers, generators, plant and containers.
AI dashcam
Road and cabin, clipped to the event.
Driver behaviour
Scored per driver and per trip.
Incident centre
Footage, telemetry and witnesses in one file.