From the RJ·AJ lab in Riyadh · services and prices
Private AI agents that run on hardware you own and read only your documents. Front desks trained for your sector. Websites, stores and iOS apps engineered by hand. The machines themselves, sized, configured and installed. Every price on this page is the price, in the open, and every consequential step waits for a named person in your company. That last part is not a feature. It is the architecture. Anything on this page can be combined, and anything not on it can be built to the same rules.
Founding customers. The first five companies get 30% off every list price in exchange for a named reference.
Nothing matches that. Try one word, or tell us the problem and we will name the package.
A private agent needs a private machine, and the right size is decided by memory, not by marketing: the model must fit in memory whole, with room for your documents and for the number of people asking at once. We size it with you, order it, harden it, and install it in your office with the model already running on your data. The machine is yours from day one; if care ever stops, the machine and the data stay with you. Five installations a month, because we do every one ourselves. When a month is full, the waitlist keeps your place in order.
The rule from the benchmarks: for one user or a handful of agents the Mac wins on memory bandwidth. When twenty or thirty people ask at the same time, the NVIDIA machines win on batched serving. Before you buy, we run your real documents at one, three and six concurrent users and show you the numbers. That test is part of the pilot.
Every product here is paid for once and returns something every year. But they are not all used the same way, so they are not all measured the same way. An executive’s tool returns that one person’s hours, and multiplies if their team uses it too. A team’s tool returns a share of each member’s week. A company-wide tool grows with headcount. An always-on desk returns cover, 168 hours a week against 45, and the enquiries that arrive after closing. Pick a product; the questions change to fit it; every number is yours to set. The answer is in months, not promises.
Setup and machine paid once, against what it returns each month net of care.
At your figures above. Hours returned to people, and enquiries that are answered instead of lost.
Some of them faster work. Some of them work that was not being done at all.
Costs are our list prices before the founding-customer discount, plus the machine, plus care for three years; integration add-ons and electricity (under SAR 50 a month) are excluded. Role pay is typical total monthly pay in the Kingdom for staff with three to seven years’ experience, before employer GOSI and benefits. The share carried, the value of an executive’s hour and the worth of a customer are yours to set; we only suggest a starting point. Hours are counted on a 45-hour week and 48 working weeks. The seat price used for reference is SAR 79 a user a month. Everything here is capacity returned to your people; what you do with it is your decision.
Every build starts with a written scope, a price and a delivery date, and an acceptance test that says what the system must do, how accurately, and what it must refuse. The 30-day pilot is judged against that test on day 30, together; if the agent fails it, 20,000 of the 25,000 is refunded and only the scoping fee is kept. Half is paid upfront and half at hand-over, and nothing is charged that was not agreed in writing.
Your machine, your data, your prompts, your code. A hand-over pack with documentation, admin training and a video. If care stops, the agent keeps running.
Care Plus replies within four working hours. A monthly health report. A quarterly review with usage numbers. Models refreshed twice a year.
Record of processing, data map, access log, retention schedule and the consent text for WhatsApp. Fines under the Saudi data law reach SAR 5 million and there is no exemption for small companies.
Models run on your machine. Full-disk encryption, access by named accounts, an audit log of every question and every action, and remote care access you can switch off. No cloud model unless you sign a written exception that names what may pass through it and what may not.
Arabic and English in the product and the paperwork. Hijri dates. GOSI, Qiwa, Muqeem, ZATCA and Nitaqat understood, not translated.
WhatsApp Business, ZATCA e-invoicing, Qiwa and GOSI exports, Odoo, Zoho, QuickBooks, Microsoft 365, Google Workspace, Jisr, Bayzat: each a fixed add-on quoted before the work.
The same architecture runs every agent on this page, and it is the architecture of the systems already in production on the lab site. No part of it depends on a single vendor, and every part of it is replaceable as the field moves.
Current open models run entirely on the machine in your office, chosen by task: a large model for reasoning, smaller ones for speed, and specialist models for Arabic speech, vision and document reading. Refreshed twice a year, with the acceptance test re-run before the swap.
Answers are retrieved from your own indexed documents and cited to the paragraph, not generated from memory. Permissions follow your directory, so each person sees only what they are cleared to see. Long documents are read whole, not skimmed.
Agents plan multi-step work and use tools: calendars, ledgers, portals, mail. Every tool call passes a policy engine that tiers the action by risk. Reversible reading runs alone; anything consequential is prepared in full and held for a named person. The refusal is a separate mechanism the model cannot reach.
Each agent carries an evaluation suite built from your real cases: accuracy, refusal of what it must refuse, latency, and behaviour under bad input. It runs before launch, after every change, and at each model refresh. The results are yours to see.
Connections to your systems use open connector standards rather than one vendor's plug-ins, so an integration built today survives a model change tomorrow. Each connector is scoped to read and write only what the acceptance test allows.
Every question, every retrieved source, every tool call and every approval is written to a log your auditor can read. Memory is scoped to the company and to the person; nothing about your business trains a model outside your building.
One workflow, one department, one named approver. A written scope, price and date. Nothing starts without them.
What the system may never do is written first and enforced by a mechanism the rest of the system cannot reach.
On your machine, on your documents, in both languages. Sources shown, evidence attached, every step recorded.
The agent runs against real work without touching anything real. Your people inspect its record until it passes the acceptance test.
Reversible chores run alone. Consequential steps wait for a signature. Documentation, training and a kill switch outside the agent.
Is this for us?
Most of the page is built for companies between twenty and five hundred staff, where one person carries a whole function and cannot take a holiday. A front desk, the HR specialist or the government relations assistant earns its place at forty people. The calculator above tells you honestly where the arithmetic works at your size and where you would be buying it for privacy and capacity rather than savings. Start with the readiness audit or a single pilot; nobody should buy a bundle first.
A general assistant answers the person typing. An agent on this page runs a workflow: it reads your policies, prepares the letter, checks the deadline, files the record, and holds the consequential step for a signature. It also never sends your data to someone else's cloud. Keep the assistant for writing emails; use an agent where the work has rules, records and consequences.
That is exactly the case these agents were built for. The model runs on a machine inside your building, reads only the documents you index, and every question, source, tool call and approval is written to a log your auditor can read. Nothing about your business trains a model outside your walls. The data protection pack that comes with every agent gives you the record of processing and retention rules the Personal Data Protection Law expects.
Yes. The Custom agent product is exactly that: a week of scoping with the people who do the job today, then a build to the same rules as every specialist here, with an acceptance test written first. Anything larger becomes a Custom system, scoped after the readiness audit and quoted as a fixed price with milestones.
What exactly do we get?
Yes. It is the list price before any discount, in riyals; the dollar figure is at a fixed rate for reference. The founding-customer offer takes 30% off every list price for the first five companies, in exchange for a named reference once the system is running. There are no hidden usage fees: where a product carries usage, the card says so.
Included: the scoping, the build, the acceptance test, installation on your machine, the documentation, one training session for your team, and the data protection pack. Not included: the machine itself, which is priced separately and owned by you; integration add-ons, which are quoted per connection before the work; and third-party fees such as WhatsApp conversation charges or an Apple developer account.
Care, not rent. It keeps the agent alive and current: monitoring, security updates, backups, the model refreshed twice a year with the acceptance test re-run, a monthly health report, a set number of hours of changes, and a reply within one working day on Care Basic or four working hours on Care Plus. Stop paying and the agent keeps running; it simply stops being looked after.
Every card names the actions the agent prepares but does not take. Paying, signing, sending in someone's name, changing an employee record: the agent drafts, attaches the evidence, and a named person in your company presses the button. That refusal is a separate mechanism the model cannot switch off, and the agent always says it is an AI.
Money and risk
Before day one we write the acceptance test together: the twenty questions the agent must answer from your documents, the pass mark as a number, the speed, and the actions it must refuse. You pay 25,000, of which 5,000 is a scoping and installation fee that is never refunded. On day 30 we run the test together and write the result the same day. Pass and continue: the whole 25,000 is credited against the setup. Fail: 20,000 is refunded, the agent is removed, your data is deleted from our machine, and you keep the written findings. A change of mind after a pass is not a fail.
Half at the signed scope, half at hand-over, in riyals. Online payment through a Saudi gateway is being connected; until it is live, press Reserve and we send a payment link for the first half within 24 hours. Nothing is charged before you have the scope in writing.
Not today. We work as registered freelancers below the VAT registration threshold, so no VAT is added to invoices. If that changes, this page will say so and prices will show VAT separately.
You can, at any time, with one month's notice on the care plan. The machine, the data, the prompts, the code and the documentation are yours from day one, and the hand-over pack is written so that another engineer can take over. We would rather earn the renewal than lock the door.
Machines
Because it is yours. You own it from day one, the warranty is in your name, and if we ever part ways it stays with your data on it. We buy it at retail, add installation and configuration, and show one price. If you already have a suitable machine, we use it and you pay only the setup.
Size is decided by memory: the model must fit whole, with room for your documents and for the number of people asking at once. The sizing tool above gives a first answer; before you buy, we run your real documents at one, three and six concurrent users and show you the figures. If a machine we recommended cannot pass the acceptance test, we exchange it at no charge.
We install every machine ourselves and test the model on your documents before we leave. Five a month is what we can do properly. When a month is full, the waitlist keeps your place in the order requests arrived, and we tell you the month before we ask for anything.
After launch
Current, and kept current. The agents run open-weight models on your hardware, retrieve from your documents with citations, use tools through a policy gate, and are measured by an evaluation suite before every release. Care Plus refreshes the models twice a year and re-runs the acceptance test. Nothing here depends on one vendor, so every part can be replaced as the field moves.
It will, sometimes, which is why nothing consequential runs without a person. Every answer carries its source, so the error is visible; every action is logged, so it is traceable; and the evaluation suite is extended with the case, so it does not repeat. On Care Plus you have a reply within four working hours, and a fix in the next change window.
No. The clinic desk never gives medical advice, the contracts assistant never gives legal advice, and the trade hunter is a tool built to your own written strategy, never a recommendation. Each card states its boundary, and the boundary is enforced by the policy gate, not by good intentions.
Who you are dealing with
RJ and AJ, a two-person lab in Riyadh. One finds the opening, the other builds the system. The five systems on the lab site, a trading system that runs inside written limits, a strategy unit, an HR specialist in pilot, a retail platform and a live AI-visibility service, are the track record. We take a small number of companies at a time because the two of us do the work.
Press Start on any card, or tell us the problem in the form below. Within 24 hours you get a reply from a person with a scope, a price and a date in writing. If you are not sure what you need, the readiness audit is three days inside your company and a written plan, credited against the first build.
Every product on this page prepares the work and waits for your signature. Pick one, or tell us the problem and we will name the package.
Start a conversation