What the AI does, and what it does not.

Because of Sakia, a smallholder in Kairouan who irrigates from their own well will know which day to irrigate and how much, from the weather forecast, through a short call or message that needs no smartphone and no reading, instead of deciding by habit or by waiting for rain. We know because, in a past SMS pilot, generic messages were judged “too general” and only about 15–16 % of 421 surveyed farmers said they arrived at the right time, and because replaying our rule on 11–12 seasons of observed weather (not on past forecasts, so forecast errors are left out) used 3 to 27 % less pumped water than a fixed weekly schedule we defined as a benchmark, with almost no water-stress days (a simulation for Kairouan, not a field measurement).

The simulation is on the proof page, with its assumptions. A small learned model, tested in public in shadow mode (it changes no advice), is on the Lab page.

The numbers come from a deterministic FAO-56 water balance, on purpose: it can be checked, and it cannot hallucinate. A spreadsheet could do that calculation.

What a spreadsheet cannot do is meet the farmer where the language is the barrier. 27.9 % of people aged 10 and over in Kairouan cannot read (17.3 % across Tunisia; INS, 2024 census), and many speak only Tunisian Arabic, which is spoken far more than it is written. So AI is used for three things: speech recognition (a farmer can speak instead of type), a voice agent that works out the crop and the region and reads our server's answer, and a voice that answers aloud in Tunisian-accented Arabic.

Understanding a typed message in English, Arabic, Latin-script Tunisian (“Arabizi”) or French is not AI: it is rules and fuzzy matching on a lexicon of crops and places.

Status, stated plainly: typed phrases were tested on sentences written by us, not by farmers, which is not an accuracy claim; speech recognition is not measured on real recordings.

• Tunisian Arabic comes first. The on-screen text and the recorded bulletins can be in Tunisian Arabic, written in Arabic letters, because it is the language of farmers who do not read standard Arabic comfortably. That text is not yet validated by a native speaker. Everywhere else the voice answers in Tunisian-accented Arabic.
• English is available in the web app, the Telegram bot and the SMS simulator; standard Arabic and French on every channel (the voice line speaks Arabic and French).
• A message typed in English, Arabic, Latin-script Tunisian (“zitoun kairouan”) or French is understood in the SMS simulator and the Telegram bot.
• Our answer to “what about a less-supported language?” Tunisian Arabic is exactly that case. It has no standard spelling, and the speech-recognition provider has not evaluated it. Our Tunisian Arabic text is built from fixed templates in common Tunisian words, with an automated check that rejects Moroccan and Egyptian forms. It is not yet validated by a native speaker.

• A human decides. The advice is indicative and must be validated by the regional agricultural administration (CRDA).
• “Not sure: ask a person.” When the tool cannot give a reliable answer it says so and does not guess. The plan and safeguard sentences are fixed templates, with the numbers filled in by the engine.
• The calculation is deterministic and the assumptions are always visible.
• Estimates are never presented as measurements; estimated days are labelled.
• The voice agent is the one place where wording is not fixed. It is a real ElevenLabs conversation with a language model (claude-sonnet-4-5) that is instructed to read our server's answer word for word: 15 of 15 plan answers were read word for word on 20 typed test phrases, but this is an instruction, not something we enforce.
• Personal data is limited to a Telegram chat identifier (pseudonymous, not anonymous) plus language, region, crop and a few settings. Our database stores no name, phone number, precise location or IP address. Raw IP addresses are held in server memory for rate limits and appear in the hosting provider's logs.
• Visits to this website are counted with Vercel Web Analytics: no cookie, no personal identifier, only totals (pages, countries, devices, where visitors came from).

On the home screen, a softer “old data” notice also appears after 5 hours (the forecast is refreshed about every 3 hours).

No large model runs on the device.

For every dataset: source, licence, size, and what it does not cover. Where a measurement has not been made, it says so. Nothing is rounded up.

A. Evidence that the problem is real

B. Data the tool works with

The irrigation advice uses no trained model. One research model (Sakia-ML, CatBoost, 43 KB) is trained and evaluated in shadow mode on the three satellite rows above; it changes no advice. Protocol and results: the Lab page.

C. What our data does not cover

• No field trial in Tunisia: crop coefficients are generic, so the advice is indicative.
• No local weather station and no soil-moisture sensor in the calculation; rain from the model is underestimated on heavy-rain days.
• No farmer registry, no market prices, no water salinity, no real irrigation calendar of the administration (the “fixed weekly schedule” in the backtest is a benchmark we built, not the State's).
• Relative yield is estimated only for the 9 crops with a published Ky, never for trees.
• Speech recognition of the Tunisian dialect is not measured on real farmers; the dialect text is not validated by a native speaker.
• Personal data is limited to a Telegram chat identifier (pseudonymous, not anonymous), the governorate and a few settings. Speech is processed by ElevenLabs under its own terms (retention not confirmed).
• The weather source was checked against Tunisian stations at Kairouan and one station 47 km away, over limited periods, not across the country.

D. Weather source checked against Tunisian measurements (3 October 2026)

Bias = Open-Meteo minus station. Stations: NOAA GHCN-D Kairouan; DGACTA stations (Oueslatia, 47 km from Kairouan; Zaghouan) and CRDA Kairouan on catalog.agridata.tn (Tunisian open-data licence, reuse allowed with attribution).

What we conclude. At Kairouan the weather source is good for evapotranspiration (within a few percent) and temperature (0.6 °C). A study in Morocco found an underestimation of 2 to 37 %: we do not find that here. The measured weakness is heavy rain: on the 43 days when the station measured 10 mm or more, the model saw about half of it, and the plan uses the model's rain as it is, with no correction. A local correction factor for evapotranspiration is not defensible today: the gap changes sign from one station to another, and the only complete series near Kairouan dates from 2019–2020. It would need at least a year of measurements. The analysis scripts are not included in this repository: treat these figures as our own unpublished analysis.

E. Evaluations: done and not done