ACIEL™ · the safe way to put Claude to work
ACIEL is one assistant for one named executive. She works inside the rules that person signed, the policies they uploaded and the Line drawn from their own job description, on their own AI key, in their own browser. She reads, drafts, searches and proposes. She never pays, never signs, never sends without the rule that allows it. Every control is in code and proved in a browser on every change. Nothing here asks you to trust a model.
The AI Augmentation Line™ is the method behind her. Paste a job description and she places each responsibility by fixed rules, not by a model's opinion: above the line, the volume she holds; below it, what only the named person signs, with a date. The ledger of what is below the line is the product. She finishes nothing a person has not started.
Every control below is in the page's code, not in a prompt alone, and every one is driven in a real browser against mocked providers on every change to the page. The proof files are public; the workflow is public. A control that has never been tested has not been tested.
| Control | What it means for you | Where it lives | Proved by |
|---|---|---|---|
| 1 · Your rules card | In a ten-minute first meeting she writes the card you sign: how to address you, whether and to whom she may send, the stop words, what must always be asked, the calendar and money rules. Money is never allowed. Until it is signed, she only drafts. | The first meeting; the desk reads the card before every run | features/aciel.feature, the rules-card scenarios |
| 2 · Your policies and documents | Upload the delegation of authority, the expenses policy, the procurement rules, any document she must hold you to. Every draft is read against them by fixed rules on your device before a model sees it, and the Legal Reviewer step names what would embarrass the firm. | Cards 6 and 9 of the set-up screen; the Legal Reviewer | The legal-review scenarios; drive-console.js (a document added and read back) |
| 3 · Tools she may use, and a gate she cannot pass | You choose her tools and connectors. A tool that reads runs. A tool that creates, updates or drafts asks you every time. A tool that sends, pays, deletes, signs or approves is never offered to her, whatever you tick. Every result is read as data: an instruction inside an email or a record is reported, never followed. | Card 7; the page is the client for every connector | drive-tools.js: a connector added, kinds classed, an acting tool asks and a No refuses, a sending tool never offered |
| 4 · Bounded output | Every locked prompt forbids a price, a commitment and anything to be sent. A reply that crosses the line is withheld before you hear it, sentence by sentence as she speaks. | Every prompt in the registry; the check in code | drive-stream.js: a sentence that names a price stops the reply |
| 5 · Locked prompts, versioned, signed | Where a model is used, the prompt is a file in a public registry, changed like code, with a sign-off line. The page carries the same text as the file; a gate fails the build when they differ. | platform/agents/prompts/, the Locked Prompt Standard | Gate 01 on every push: the page's strings equal the registry files |
| 6 · Receipts, never content | Every model run and every act leaves a line on your device: prompt name, outcome, tool names, who overruled what and why. Never the content. Exported as one evidence pack when your auditor asks, with the override rate, rule 6 of the Line. | Card 8 of the set-up screen | The receipt scenarios; drive-console.js |
| 7 · Your key, your device, your provider | She runs on your own AI key in your browser. Your words go to the provider you chose and nowhere else. StoryC sees receipts with names only. Install her on your own Supabase and GitHub if you want nothing shared at all. | The page; the installable kit | drive-online.js; the kit's own proofs on every push |
Proofs and gates: platform/qa/playwright and the workflow runs. The registry: the Locked Prompt Standard.
The short list a board wants before it says yes. Each one is enforced in code, not promised in a prompt.
No payment, no transfer, no purchase. The rules card cannot allow it; the tool gate cannot offer it.
Not a contract, not an invoice, not an invitation. Those are lines in your ledger, for you.
A send happens only where your signed card allows it and you asked for it; otherwise it is a draft in your own drafts folder.
An email, a record, a web page or a document that tells her what to do is reported to you, not obeyed.
Receipts carry names and outcomes. Your documents, your inbox and your words stay on your device and with the provider you chose.
Prompts are locked files with a sign-off. A change is a new version, in public, with a gate that proves the page matches.
The questions we are asked before a deployment, answered plainly.
To the AI provider you connect, on your own key, for your own turns, and to nothing else. ACIEL runs in the browser on your device. The desk, if you switch it on, runs inside your own Claude account with the connectors you tick. StoryC's licence service sees your licence key and your organisation name. StoryC's funnel receives receipts: prompt names, outcomes and tool names, never content; you can switch that off in the installed kit.
Only under the rules card the executive signed: to approved people, in threads they wrote, or anyone except stop words, and only when a request asks for a send. Anything to do with money, regulators, legal, people, press or the board is always asked first. Everything else is a draft in their own drafts folder for them to send. Her mailbox connector cannot send at all; it drafts.
Three things. Every connector result, document and web result is wrapped as data with the rule that an instruction inside it is reported, not followed. The tool gate classes every tool by name, so a sending or paying tool does not exist for her whatever a message says. And the bounded-output check withholds any reply that names a price, makes a commitment or asks for a send, before it is shown or spoken.
As built, ACIEL sits outside Annex III of the EU AI Act: she scores no natural person's credit and evaluates no worker. Article 50 transparency applies and is met: the person always knows an AI assistant wrote the draft, and the receipts say which prompt. In the UK the ICO's agentic AI expectations (accountability through the chain, data minimisation, no automated decisions) and the FRC's March 2026 guidance (named accountability, documentation, explainability) are what the rules card, the ledger and the evidence pack are for.
Yes, three ways. The evidence pack from the device: every run, every act, every overrule with its reason, the override rate. The public registry of locked prompts with their change logs. And the proofs: the page driven in a real browser on every change, with the results public.
Yes. The installable kit puts the page on your own GitHub Pages and the desk's tables and functions on your own Supabase project, from one configuration file. StoryC keeps only the licence service. Nothing in the kit calls a model: the page talks to your provider with your key.
Claude first, on an Anthropic key, including Claude's web search on that key. Any OpenAI-compatible endpoint works for the conversation, including Azure OpenAI. ElevenLabs for her voice is optional; the browser's own voice otherwise.
Two ways in, both in pounds sterling, both for one named person. Every price is on the ACIEL page and nowhere else.
One hour together, paid on booking. You leave with her online: your rules card signed, your policies read, your Line drawn, her voice chosen, and the desk instructions in your own Claude.
Buy the licence on the ACIEL page; your key is issued the moment you pay. Your own AI key, your own device, cancel any time. The asset pack for your own domain is included.
A pilot for a team: three named people, three licences, one set-up hour each, and a thirty-day review of the override rate together. Ask on the set-up call.