All-in-One Sovereign Agentic AI Platform
Language models and AI agents on your own servers. They read mail, work through contracts and tenders, keep the CRM and close tasks by morning — while company data never leaves your walls. We install, tailor to your processes, train people and support.
Companies in the UAE and the region handle data that cannot be sent abroad, and volumes at which the cloud API bill becomes a budget line. A local platform solves both — and adds a third thing: agents that do the work rather than just answer questions.
77% of GenAI users paste company data into it, 22% paste personal and payment data, 82% do it through personal accounts with no control at all. Bans do not work — your own AI that is pleasant to use does.
Source: LayerX, 2025SAR 27M per incident; shadow AI adds another $670k to each. 97% of those hit had no AI access controls.
Source: IBM Cost of a Data Breach, 2025Health data stays in the UAE (law 2/2019, up to AED 1M). Bank customer data only with CBUAE approval. ADGM up to $28M, Saudi PDPL up to SAR 5M, doubled on repeat.
Source: Chambers, 2026A US court in 2025 ordered all ChatGPT logs preserved and 20M chats handed to plaintiffs. Enterprise contracts were carved out — the rest were not. There is nowhere to send a local model an order.
Source: terms.law, 202584% call a week of vendor downtime critical. Vendors retire models (GPT-4.5 after 4.5 months) and double prices. An open model on your server will not vanish or get dearer.
Source: IBM IBV, 2026The median company spends $11 per person — it does not need its own server, and we will say so. Active teams spend $611, teams running agents up to $7,450. With 50 active staff the Business server pays back within a quarter. The calculator below runs on your numbers — honestly, including the cases where the API is cheaper.
Source: Ramp AI Index, 2026$60M saved a year, two thirds of contacts, 82% faster replies. An agent that does the work — not a chat that answers questions.
Source: Klarna Q3 202572% fear data leaking into third-party AI tools. Public cloud as the main inference environment fell from 56% to 41% in a year.
Source: Dell, 2026 · Broadcom, 2026Not a feature list, but what the agents do from the first month. Every action waits for approval where a person should decide.
One server or a cluster. No layer calls outside: models, the document index, agents and the log live on your hardware. Tools connect over MCP — the open protocol handed to the Linux Foundation in December 2025.
Public cloud models stay outside the perimeter. Connected on your decision and only through the masking gateway.
The choice depends on whose data you hold and whether you have a server room. The line can move later: the platform is the same.
Data never leaves the building. Your servers, inside your walls, every key in your hands. For banks, healthcare, government and anyone whose data must not travel.
Dedicated servers in a UAE data centre — Khazna, Equinix, Moro Hub — run by us. The same platform, no server room of your own. For companies without premises, small ones included.
When you need the strongest public model: a local gateway masks names, numbers and amounts before sending, restores them on the way back and refuses when unsure.
The cloud wins on time to start and access to the strongest models — and we do not hide it. For data, law, budget and control — the local platform. Sometimes the right answer is the cloud, and we will say so in the assessment.
| Criterion | Cloud API | On-premise platform |
|---|---|---|
| Where the data is | With the provider, often abroad | Inside your perimeter |
| Regulators: PDPL, CBUAE, DoH | DPA, risk assessment, consents, approvals | Localisation by construction |
| Cost as usage grows | Grows with every request and seat | Fixed: hardware + subscription |
| The very strongest models | Available on release day | Open ones trail by months |
| Time to start | Minutes | Weeks: hardware and setup |
| Tailoring to the company | Prompts and RAG | Fine-tuning on your data |
| Latency and offline | The link and the provider’s queue | Local network, works without internet |
| Dependence | Prices, limits, model retirement — the provider decides | A model swaps in a week |
| Audit trail of AI actions | Provider-side, partial | Complete, yours |
| Maintenance | Not your concern | In the subscription — but it is people and hours |
Three typical answers are in the table. Below — a picker for your company: move the sliders, click the dots. You may buy the server yourself from any supplier; we specify, install and tune it. Prices are an October 2026 guide, VAT excluded; NVIDIA quotes change monthly.
| Size | Staff | Models | GPU | Server | Power |
|---|---|---|---|---|---|
| Starter Budget alternative — 2× RTX 5090 (64 GB) |
10–30 | Qwen3-32B · gpt-oss-120b · Falcon-H1-Arabic-34B | 1× NVIDIA RTX PRO 6000 Blackwell 96 GB — a workstation | $20–23k | 0.6–0.9 kW, a normal socket |
| Business 2,000–3,000 tok/s on 70B, up to 7,000 on MoE; 50 agents at 32k context |
50–200 | Llama 3.3 70B FP8 · Qwen3-235B INT4 — several in parallel | 4× RTX PRO 6000 Server (384 GB), a 4U server | $105–120k | 3.4 kW, a rack and a 32 A line |
| Enterprise ~9,900 tok/s per server; 200–500 sessions |
200+ | DeepSeek-V3.2 FP8 · Qwen3.5-397B · Mistral Large 3 | 8× H200 (1,128 GB) or 8× B200 (1,440 GB), a cluster | $275–562k | 10–14 kW, colocation |
Horizontal — server price, vertical — the largest model that fits. Dot size — how many people work at once at normal speed (≥15 tok/s). The hatched zone is your conditions. Hover or click.
Left — how many people at once, right — one person’s speed on the recommended model. Configurations in ascending price.
Cumulative cost of the selected configuration versus renting the same GPUs in the cloud and versus API fees at your token volume.
How we counted. One person’s speed is memory bandwidth divided by model size, scaled by 0.7 for NVIDIA and 0.85 for Apple; concurrency comes from vLLM/SGLang measurements (databasemart, 08.2026; SemiAnalysis InferenceX, 09.2026) and memory left for contexts. Spark, Mac mini and Mac Studio stacks follow Exo Labs and NVIDIA measurements: memory adds up, single-user speed barely grows because a token passes through every unit in turn. Mac mini M6 Pro and DGX Station GB300 are estimates from the 2026 line-ups. Rental is about 10% of the purchase price a month (RunPod, CoreWeave, 10.2026); electricity at the top DEWA tariff; maintenance 1.5% a year. API prices are blended (80% input, 20% output) as of October 2026. All of it is a guide, not an offer.
Every step ends with a tangible result and a decision — go further or not. The pilot is measured in minutes per task and in money, not in impressions. The first agent in production — within 10–14 weeks; timings are an estimate from published 2025–2026 rollout plans.
Interviews with leaders, inventory of systems and data, baseline metrics, network and security review. Together we pick the first three use-cases with a measurable effect.
We install the server in your building (yours or ours), deploy models, connect SSO, an isolated VLAN, connectors to mail, CRM and Jira, and load the documents.
Two weeks in shadow mode: the agent works alongside people, people grade it. Then 10–15% of live traffic and a pilot group of 20–50 staff. Metrics are written down in advance.
Full traffic with a rollback button, the personal CRM, training for all staff (5+ hours), champions in each department. Up to 5–10 agents within 3–6 months.
All keys and documentation are yours. Monitoring, model updates, a new use-case every quarter, server access only with your approval.
No sovereign vendor publishes prices: Netveva, Cohere North, Mistral Enterprise, Aleph Alpha — all “on request”. We publish ranges. Two sums per package: a one-off rollout and a care subscription. Hardware is separate: flip the switch to see the total with it. Use is unlimited.
Prices exclude VAT (5% in the UAE); dirhams at 3.6725. Hardware is quoted for purchase from a supplier, plus a 5% duty outside free zones. For private-cloud hosting instead of purchase — a monthly rental, calculated during the assessment. For banks and healthcare — a 15–30% compliance surcharge.
A CRM you do not have to fill in. The core is open-source (Twenty: PostgreSQL, built-in agents and MCP), the agents run on your local model. Contacts and deals appear from mail, calls and messengers by themselves; the database is yours, on your server, zero per-seat licences.
| HubSpot Breeze · Salesforce Agentforce | Your local CRM | |
|---|---|---|
| Where data is processed | At OpenAI and the vendor’s partners | On your server |
| Price | $1 per lead, $0.50 per conversation — on top of the per-seat fee | Included, no limits |
| Telephony and WhatsApp in the UAE | Via third-party connectors | e& / du, WhatsApp Business API — direct |
| Source code and database | The vendor’s | Yours |
Salesforce State of Sales 2026: 54% продавцов уже работают с агентами, экономия на черновиках −36%
88% of companies already use AI, yet only 39% see an effect on profit. The gap is not in models but in choosing tasks, data, rules and people. That is what we sell: assessment → strategy → rules → people → care. The platform joins at step two or three — as one option, not the goal.
| Service | Duration | Result | Price |
|---|---|---|---|
| AI Pulse — rapid assessment | 2–3 weeks | Maturity across six domains, risks, 10 use-cases with impact estimates and priorities, a hardware spec | $8–15k |
| Use-case workshop | 1–2 days | Leaders and key staff; a ranked use-case list, an effect calculation, a pilot plan | $5–10k |
| AI Blueprint — strategy & roadmap | 4–8 weeks | Strategy, budget, KPIs, a 12-month quarterly plan, team requirements, a data model | $25–60k |
| AI policy, PDPL & Dubai AI Seal | 3–6 weeks | An AI use policy, procedures, a risk register, UAE PDPL and AI Charter alignment, a Dubai AI Seal application | $15–40k |
| AI Academy — training | 1 day · a 4–6-week programme | Three tracks: leaders ($6–12k a day), staff ($3–6k a day), IT. Hands-on with your own tasks, 5+ hours per person; a programme for a 50–250-person company — $15–40k | $3–12k |
| Fractional Chief AI Officer | per month | Leading the rollout: 2 days a month (advisory) or 1 day a week (embedded) — priorities, vendors, budget, board reporting | $6–20k |
| The “Sovereign AI Launch” bundle | 4–5 months | Assessment + strategy + policy + training + 3 months of fractional CAIO + a pilot of one agentic use-case. Hardware and subscription excluded | $80–150k |
Our day rate is $1,200–2,000: a strong boutique’s level without the Big Four mark-up, where an hour costs $400–800 — $3,200–6,400 a day. All package prices are AI Finder estimates from market corridors; the quote is fixed after the assessment.
Nothing. Models, the document index, agents and the log run on your server. If a specific task needs a public model, data passes through a masking gateway: names, numbers and amounts are replaced before sending and restored after. This is enabled separately and only on your decision.
Open models are updated every few months and the subscription covers it: we evaluate new releases on your tasks, size them to your hardware and apply them when you are ready. The hardware lasts for years: data centres depreciate GPUs over 5–6 years.
We do, under the subscription: monitoring, updates, new use-cases, engineer support. Keys and access are yours; we connect only with your approval. One or two of your people are trained so day-to-day tasks do not need us.
Yes. For Arabic — Falcon-H1-Arabic (TII, Abu Dhabi) and Jais 2 (G42), with dialects and correct right-to-left layout. Russian and English are native to all major open models. Agents reply to customers in the language they wrote in.
The first agent in production — 10–14 weeks after the start; 9 weeks under ideal conditions. Starter in full — 10–14 weeks, Business 3–5 months to 5–10 agents, Enterprise 4–8 months including hardware lead time (6–14 weeks) and regulator approvals.
That is why we start with a pilot whose metrics are written down in advance: minutes per task, accuracy, the share of mail handled without a person. The full-rollout decision is made on those numbers. The hardware stays yours either way and suits any other workload.
Not for Starter: a workstation around a kilowatt and a normal socket. Business needs a cooled rack and a 32 A line; Enterprise needs colocation. With no room of your own, the platform goes into a private cloud in the UAE — data stays in-country.
Three things. Data does not go to the vendor. You pay for the rollout and a subscription, not $30–75 per seat — for 100 staff that is $36–90k a year. And the agents are built for your processes and systems, not for an average office: they work with your ERP, your contracts and your rules.
Two or three weeks — and you have a report: where AI pays off, what hardware is needed and what it costs. No obligation to buy the platform.
Discuss a project