Pillar 02 · Artificial intelligence
Sooatek offers SMEs AI assistants, agents and models connected to their tools: a costed assessment, a prototype measured on your data, integration, local AI when your data must not leave. Already running: a vision model in production at a client, our internal assistant Tsiaro, and an open Malagasy speech recognition model (12.7% word error rate with a language model, against 21.5% for our previous model).
§ 01 / The steps
Each step has a written deliverable. The workshop, assessment, integration and operation are at a fixed price, each with its own quote; the measured prototype is the first stage of the integration and is part of its quote. You can stop after the workshop or the assessment; the assessment is deducted from an integration signed within 30 days, above a minimum amount stated in the quote.
Half a day
Demonstrations on real cases, your questions, what AI can and cannot do for you.
3 weeks
Inventory of uses and data, three costed use cases (gain, cost, risk), GDPR and AI Act review.
2 to 3 weeks
On your data, before any production release: the measurement decides whether we continue.
4 to 8 weeks
Connected to your tools, with safeguards, an action log and documentation.
Monthly
Hosting, monitoring, re-indexing, updates; model usage re-invoiced at actual cost.
§ 02 / What we integrate
For SMEs whose teams search for information, re-type data, sort emails or requests.
An AI that answers from your documents, procedures or knowledge base, and cites its sources. Updated without retraining.
Connected to your tools (APIs, email, CRM, n8n or code): the agent prepares, a human validates every outgoing action, every action is logged.
When a general-purpose model gets your language, vocabulary or images wrong: fine-tuning, speech, vision, measured before and after.
Open models running on your server or on our hardware, when your data must not leave. European models are also possible.
§ 03 / Offers
From the assessment to the specialized model, up to 3 times cheaper than the European market for the same standards. AI model usage and GPU compute are re-invoiced at actual cost.
Offer 06
3 weeksFind out where AI saves you time, what it costs, and where to start.
Included
Pricing
Fixed price, on quote
Deducted from an integration signed within 30 days, above a minimum amount stated in the quote
Offer 07
4 to 8 weeksAI connected to your documents and tools, put into production, under human control.
Included
Pricing
Fixed price, on quote
Model usage re-invoiced at actual cost
Offer 08
6 to 10 weeksA model adapted to your language, vocabulary or images, measured before and after: fine-tuning, speech, vision.
Included
Pricing
Fixed price, on quote
GPU compute re-invoiced at actual cost
Agencies and IT services companies: −20% off the direct rate, excluding costs re-invoiced at actual cost. The agency program
Compare all offers§ 04 / Safeguards
The user information required by the AI Act (Article 50) has applied since 2 August 2026; the marking of generated content is postponed to 2 December 2026 for systems already on the market (digital omnibus, subject to the final text). We build these obligations in from the design stage, together with rules of our own.
A human validates every outgoing action
Sending an email, writing to a tool, replying to a client: nothing leaves without validation.
An action log
What the AI suggested, what was validated, edited or rejected, and by whom.
The “AI-generated content” notice
Your users know when they are dealing with an AI (AI Act, Article 50).
Your data stays yours
European or local models are possible; your training data is only used for your model.
A measurement before production
The prototype is measured on your data: with no measured gain, we stop.
What we do not do
§ 05 / Proof
What is public, and what runs in-house; the only client case cited, a vision model, is anonymized. The figures below come from the model’s public card.
Public model
We fine-tuned and openly released a Malagasy speech recognition model: 12.7% word error rate on 11,625 short utterances never seen in training (with a language model; 14.3% without), against 21.5% for our previous model, a Whisper medium fine-tuned on the same data; 10.1% on long clips.
The model card on Hugging FaceClient case · anonymized
A mobile app recognizes an object among thousands of near-identical references from a photo: a retrained image model that runs on the server and on the phone.
The case in detailMeasurement rigor
We publicly corrected a wrong measurement of our previous model (non-normalized text): the card says so and points to the new model.
The model and its measurementIn production at our company
In production at Sooatek since October 2026: a local model sorts our emails, chats and tickets, a human validates every message sent.
Discover Tsiaro§ 06 / Madagascar
Same offers, same deliverables: quotes in local currency on request, workshop and assessment on site in Madagascar (in France: remotely by video call, or on site with travel costs extra). Our Malagasy speech recognition model is open: it can be a starting point for your projects in Malagasy.
§ 07 / Questions
With a half-day leadership workshop, then an assessment that costs out three use cases: expected gain, cost, risk.
To answer from your documents, a document assistant is usually enough: it cites its sources and is updated without retraining. Fine-tuning is for when a model gets your language, vocabulary or images wrong, as with our Malagasy speech recognition model.
Not necessarily. We offer European models or local models, on your server or on our hardware; the choice is written into the assessment, with its cost. Usage of hosted models is re-invoiced at actual cost.
We build in its transparency obligations (Article 50): your users know when they are dealing with an AI (applicable since 2 August 2026) and when content is generated (marking postponed to 2 December 2026 for systems already on the market, subject to the final text of the digital omnibus). A system classified as “high-risk” requires legal support, which we do not provide.
Describe the task that takes up your teams’ time: we tell you in writing whether AI can help, how, and for how much.
With your consent, we measure the site’s audience with Google Analytics, in anonymised form. The data is sent to Google, including outside the European Union. Learn more
