aios · ai operating layer
AIOS is the layer between your people, your systems and the models, and it coordinates all three: which model runs, who may ask it what, which tools it may touch, who approves an action, and what is logged. I design it, build it on your cloud, and hand it over with an audit trail you own.
Every AIOS I build has the same six parts. Each one answers a question your business already asks, and each one is specified in the Blueprint before anything is built.
01 / context and data connectors
I specify each source, the scope it is read with and the permissions that travel with the data: your CRM, ticketing, ERP, documents and databases, connected read-only until a workflow earns write access.
02 / swappable model layer
I build one interface in front of Anthropic, OpenAI, Google and open-weight models on your cloud, so each workflow names its model in configuration and you are never locked to a single vendor.
03 / permissions and identity
I map your existing identity provider and roles onto every agent and tool, so an agent acting for a support colleague can do no more than that colleague could, and every call carries a named identity.
04 / agent runtime with approval gates
I build the loop as read, decide, ask, act, record: an agent may propose a side effect, and where it matters a person approves it in the interface they already use before the tool is called.
05 / audit and observability
I write an append-only log of every model call, tool call and approval, and put cost and quality dashboards on top of it, so you can answer an auditor or a finance question from the same record.
06 / interfaces
I connect the layer to Slack, Teams, WhatsApp, email and the web, whichever your people use today, so approvals and answers arrive where the work already happens instead of in a new tool.
Pick a scenario and run it. The agent reads, decides and proposes an action, then stops at the approval gate until you approve or reject it. Every step is written to an audit log you can download.
Simulation. Nothing is sent and no model is called. This is the shape of the loop I build: read, decide, ask, act, record.
Without JavaScript: the agent reads the ticket, checks its permissions, proposes a refund, and stops at an approval gate until a person clicks Approve. Every step is written to an audit log you own.
The same console runs on the Evidence page. The log it produces is a sample; the one your layer writes has the same shape and lives on your cloud.
The term gets stretched. Here is what you are not buying.
My stance, so you can check it against the Blueprint you receive.
The Blueprint compares platforms honestly: the options are named, the cost model is written out for three years at your expected volume, and the recommendation says which components to buy and which to build. Where a platform covers a component well and its pricing survives the three-year view, I say so and design around it.
Where custom on open source is cheaper over three years, or where a platform would own your data, your prompts or your audit trail, I build it on your cloud. I do not resell anything, I take no commission from vendors, and the layer carries no licence fee to me after handover.
Design first, at a fixed price. Build second, scoped in writing. Operate third, month by month. Each stage stands on its own and you can stop after any of them.
The architecture, build-vs-buy decision, cost model and 90-day plan for your AI operating layer. Credited in full against a Foundation build within 60 days.
$6,000
The governed layer in production: connectors, model layer, permissions, agent runtime with approval gates, audit trail, and your first three to five workflows live.
from $35,000
Model and prompt monitoring, cost and quality dashboards, eval re-runs, one new workflow a month, one-business-hour response, quarterly planning.
$2,500/month
Converted prices are approximate. You are billed in USD.
Foundation is typically USD 35,000–90,000. A starting point, not a quote: the written proposal fixes the number before work starts, never after. If the first milestone review is not what we agreed in writing, you do not pay for that sprint.
The outline is the same for every client; the content is yours. Download the outline to see the shape of the document before you buy it.
Download the Blueprint outline sample
Delivered as a document in a repository in your organisation, with the architecture diagrams and the cost model as editable sources.
Above roughly $25,000 I bring in vetted specialists under my contract. You still have one accountable engineer. Repos live in your organisation from day one and documentation is delivered at every milestone, so nothing depends on my availability. Read the full process →
The questions security, legal and finance ask before they sign. The answers are the design, not a promise.
Yes, and that is the default. The layer is built on your cloud account, whether that is AWS, Azure, Google Cloud or a VPC you already run. Data residency is a design input, not an afterthought: the Blueprint records which region every component lives in and which data may leave it. Nothing routes through infrastructure I own.
Whichever fit the job and your constraints. The model layer is swappable by design: Anthropic, OpenAI, Google, or open-weight models on your own cloud. Each workflow names its model in configuration, so you can change it without touching the workflow. The Blueprint compares the options on cost, quality and residency for your first workflows.
Every model call, tool call and approval is written to an append-only log with who asked, what was read, what was written and what it cost. Retention is a setting you decide at Blueprint stage, together with what the log may contain and who may read it. Prompts and outputs can be redacted or excluded where they contain personal data.
People and roles you name. Each side effect, such as sending an email, changing a record or making a payment, has an approval rule: who may approve, above which threshold, and what happens on timeout. Approvals are made in the interface your people already use and are recorded in the audit log with the approver's identity.
I design for them and name the obligations that apply to each workflow: lawful basis and data minimisation, records of processing, human oversight where the AI Act asks for it, and the ability to explain and reverse an automated decision. The Blueprint lists these per workflow with the evidence the layer produces. I am not a law firm; your counsel signs off, and the design gives them something concrete to sign.
You switch, or you negotiate with the numbers in hand. Because the model layer is swappable and every call is logged with its cost, you can see what a change would cost you per workflow and move that workflow to another model in configuration. The cost model in the Blueprint is built for exactly that comparison. There is no licence fee to me, so my side of the bill does not change.
Buy something small today, talk for twenty minutes, or write one line. All three end with a written answer.
A scoping call, no slides. You leave with a written recommendation within 48 hours, whether or not you buy anything.
One line about what you need is enough. Reply within one business day.
Replies within one business day