OCTOPUSCustom AI Solutions · 13 min read
How an AI agency in Saudi Arabia runs itself on AI: the Octopus case study
Two founders, twenty-seven defined AI roles, one owner instruction per project. The figures we have, where they come from, and the parts that broke.
Key takeaways
- Octopus Mena runs on 27 defined AI roles, two founders, and a written loop: one instruction, one ticket, verification, then a human decision.
- The figures on file are build times, test counts and credits; hours saved and cost per deliverable are not measured yet.
- What broke taught the most: shipped placeholders, first-run agent mistakes, and no time log from day one.
Contents
- The situation: two founders who had to run like thirty
- What we built: named roles, a portal, dashboards, a WhatsApp agent, and a search engine
- How it runs day to day: one instruction, one ticket, a pipeline, and a reserved list
- What it changed: the figures we have, and the ones we do not
- What broke, and what we learned
- What a Saudi company can copy from this
- Frequently asked questions
This is a case study of an AI agency in Saudi Arabia that runs its own company on AI. Octopus Mena has two founders and no staff beyond them. Since the summer of 2026 the company has run on a team of defined AI roles, a job ticket system, a client portal we built ourselves, and a short list of decisions that only a human takes. What follows is the record: the figures we have, where each one comes from, and where a figure does not exist yet. If you are a Saudi owner deciding whether this is real, read it as a working file, not a pitch.
The situation: two founders who had to run like thirty
Octopus is the AI sister company of an events agency in Riyadh. By August 2026 it had client accounts in real estate, fitness and retail, a website to build and launch, proposals to write, weekly client reporting to deliver, and a content and search program to run. A conventional agency covers that with ten or more specialist roles plus an account team per client: a company of thirty.
We had two people, and we did not want to hire ahead of revenue. So we made a bet: build the company as the same kind of system we would build for a client, and run our own operation on it first. Everything described below is the architecture we offer under custom AI solutions, and the first entry among the systems we run in production. If it did not work for us, we had no business offering it to anyone else.
What we built: named roles, a portal, dashboards, a WhatsApp agent, and a search engine
The agent team. Twelve Octopus agents, each with a name, a job description, a knowledge pack, guardrails, a taste file and a task log: Bryan (strategy and business development), Ryan (creative director), James (content marketing), Mark (social content), Seol (SEO), Omar (paid media), Ayman (web development), Rami (AI video), Zaid (graphic design), Mira (HR), Arwa (finance) and Noor (workspace administration). Above them sits Octo, the Head of Staff, which opens tickets, names owners and verifies work between stages. Two client-specific teams run inside the client folders with their own toolkits: six roles for JRE Adyah, led by a team leader, and eight roles for The Forum. That is 27 defined roles in total.
The client portal. We stopped sending PDF proposals and shared spreadsheets. A proposal or a status report is now a private link on our own domain: an email gate per recipient, dwell time measured per section, an admin view that shows who opened what and for how long, readable link names, and a PDF only as a fallback after approval.
Client dashboards and a WIP tracker. Each client's status lives on a gated dashboard on that portal, refreshed weekly on the same link. Two are live: JRE Adyah with 11 deliverables across 5 phases, and The Forum with 38 deliverables across eight phases. Behind them is one master work-in-progress tracker that states status in plain text, names functions rather than people, and carries no costs.
The WhatsApp agent. An AI assistant on a business number that greets, names itself as an AI, asks three qualifying questions, captures name, company and work email, scores the lead, and hands off to the founder with a summary email. It masks national IDs and card numbers before anything is stored, respects office hours, stops on STOP, and has a kill switch. It has passed 15 of 15 QA tests and is not live yet; it stays behind an off flag until the founder's account steps land.
The content and search engine. A content plan of five pillars, ten articles and three flagship pieces; a blog rebuilt to an SEO and answer-engine specification with schema, an llms.txt file and FAQ blocks; a launch keyword set of 27 English and 20 Arabic terms; and now a commercial keyword sprint of 33 English and 18 Arabic terms across seven clusters. A routine reads the inbox three times a day and routes content ideas to the writers. Nine presenter films, one per page, each between 52 and 55 seconds, were produced by the same team.
How it runs day to day: one instruction, one ticket, a pipeline, and a reserved list
Most "AI agency" stories go vague here. Ours runs on a written loop, and every step leaves a file.
- Step 1: the founder states a project to Octo in one instruction. Not a brief, not a meeting: one message.
- Step 2: Octo opens a job ticket from a template, names the pipeline and the owner of each stage, and puts the ticket on the Projects Board, the single source of truth for who does what.
- Step 3: each agent boots from its definition and knowledge pack, reads the board, finds the task it owns, and does its stage. It saves the output to a defined path, writes a handoff note on the ticket, and hands back.
- Step 4: before calling its work done, the agent writes a checklist of every change the job required and verifies each one on the real surface: the page opened, the video playing, the form sending. "The build passed" is not verification.
- Step 5: Octo verifies between stages, then presents the finished deliverable to the founder.
- Step 6: the founder makes only the reserved calls. On approval, the reserved action happens.
The reserved list is written down, and it is short: publishing, sending, spending, production deploys, tenant and security changes, client-facing delivery, hiring contact and offers, and per-fact clearance of anything about a client. No agent does these, whatever the instruction says. Octo prepares each one with a recommendation; a human decides.
Around the loop sit weekly routines. For JRE Adyah, for example, the client team collects on Thursday, refreshes the work-in-progress tracker on Saturday, and has the pack ready by 11:00 on Sunday, triggered by Octo. If you have read what an AI operating system is, this is what the four layers look like when they are actually running: the packs are the memory, the definitions are the skills, the reserved list is the governance, and the portal and inbox are the connectors.
What it changed: the figures we have, and the ones we do not
Every number below has a source file in our own records; where we have no number, we say so.
The client portal. The build plan for Phase 1 estimated roughly 36 to 49 hours of work across seven milestones, at conventional developer pace. The build and its local QA were logged at about 1 hour 10 minutes from the go. Two caveats: that is the developer agent's clock, not the planning and review around it; and production deployment then waited on a terms click that only the founder could make.
The films. The two inside-page films for the work and contact pages were rendered together in about 1 hour 10 minutes of internal time for 30 platform credits. The ninth film, for the start page, took about 35 minutes and 40 credits, because it needed three takes. The homepage brand film pipeline cost about 600 credits in total.
The website launch. The release was verified by a live test suite of 89 checks: 86 clean and 3 explained. The site went live on 21 August 2026; the brand launch date was Sunday 13 September 2026.
The search layer. The sitemap was submitted with 12 pages and 0 errors. The first article of the commercial sprint was drafted at 1,756 words with its on-page checklist filled in.
The WhatsApp agent. 15 of 15 tests passing on the local pipe, including the ones that matter to a Saudi owner: a request for pricing gets no number and a handoff, a national ID pasted into the chat is refused and masked, an angry opener goes straight to a human, and STOP ends the conversation and marks the contact for deletion.
What we do not have. We have no figure for hours saved against a human team, no cost per deliverable, no count of leads since the launch, and no search visibility number for our own domain yet. The rule to log time per creative task was set on 24 August 2026 and is not yet producing a table. The first search read is scheduled for 27 September 2026. We would rather publish this case study with those gaps stated than fill them with estimates.
What broke, and what we learned
Two pages shipped with a video block and no film. The work and contact pages advertised a runtime for a film that did not exist. The slots had been left as styled placeholders while the films were produced, and nobody re-checked the whole site before the batch shipped. The fix became the checklist rule above, and runtimes are now read from the video files; there is no hand-typed runtime anywhere on the site.
The WhatsApp agent got several things wrong on the first run. Asked "Are you a real person?", it handed off to a human instead of answering. The word "contract" in a message was treated as a price question. An angry opener got a cheerful greeting before the handoff. A follow-up question was stored as the person's name. Each was caught by the test set, fixed, and logged. That is why we say 15 of 15 rather than "it works".
A contraction cost a take. The start-page film needed three takes because the presenter said "Here's how" where the approved script said "Here is how". Verbatim means verbatim; two takes were lost to it.
The reserved list slowed us at exactly the right moments. The portal sat un-deployed until a human accepted a provider's terms. The WhatsApp agent still waits for the founder's account steps. That is not a flaw. The parts that commit the company wait for the person who owns it.
We did not log time from day one, so we cannot show you hours saved. The agents log what they did; the missing piece was how long each task took against what a human would have cost. Start that on day one.
What a Saudi company can copy from this
You do not need twelve agents. The parts that transfer have nothing to do with the model.
- Write roles, not prompts. Each agent has a job description, the folders it may write to, and the things it may not do. An agent without a lane will wander into someone else's work. The three columns of that scope, and who signs it, are in what to write down before an AI agent touches your company.
- One instruction, one ticket. The owner speaks once; a ticket carries the work; every handoff is written on it. If you cannot show the ticket, the work did not happen.
- Put the reserved list on paper before the first agent runs. Money, sending, publishing, security, anything client-facing. The six-step plan in how to implement AI in your company in Saudi Arabia starts with exactly this.
- Verification on the real surface is the product. Agents will tell you the work is done. Make them show it.
- Log time and cost from day one. We can show you build times and test counts; we cannot yet show you hours saved. Do not repeat that.
- Keep the client's view on a link you control. A gated page you can update and measure beats a PDF you cannot see being opened.
If you would rather bring a partner in to set this up, test us the way you would test anyone: how to choose an AI adoption partner in Saudi Arabia is the list of ten questions we ask you to put to every firm, with our own answers at the end.
Our WhatsApp lead agent is one of seven jobs an agent can take on in a Saudi company today. What AI agents can actually run in a Saudi company today lists all seven, with the approval line for each.
The roster, the ticket template and the reserved list are documents, not software; the portal, the dashboards and the WhatsApp agent are the parts we build. Setting the documents up inside your own company, with your people trained to run them, is the AI adoption program. If you want to see how a first workflow of this kind is scoped for a Saudi company, start with discovery.
Frequently asked questions
Is this a real company or a demonstration?
It is a real company with paying clients in Saudi Arabia, run day to day on the system described here since the summer of 2026. Every figure in this article comes from our own job tickets, task logs and build plans. Where we have no measurement, we say so rather than estimate.
How many people does it take to run an AI agency this way?
Two founders run Octopus Mena today. The twelve Octopus agents and the two client teams do the production work; one founder gives instructions, reviews finished work and takes the reserved decisions. Roles, tickets and verification carry the load that would otherwise need a staff of around thirty.
What does the human still do in an AI-run company?
The human owns the reserved list: publishing, sending, spending, production deploys, security changes, client-facing delivery, hiring, and clearance of any fact about a client. The human also gives the one instruction that starts each project and reviews the finished deliverable before it goes anywhere.
Can a Saudi company outside marketing run AI agents like this?
Yes, because the transferable parts are documents, not code: role definitions, a job ticket, a written reserved list and a verification step. A trading firm, a clinic group or a developer can start with one workflow, one agent and one owner, and add roles as the logs justify them.
What did it cost to build?
We do not publish a cost table yet. The figures on file are build times (a client portal built and tested locally in about 1 hour 10 minutes against a plan of 36 to 49 hours) and platform credits for films. A logged cost per deliverable is the next thing we are measuring.
Custom AI Solutions
The AI-native system, website, or product you need, built and run. Blueprints, websites, apps, AI operating systems, and custom agents.
