An offline AI assistant drafts and rewrites documents, summarises long files and meetings, answers questions from your own document sets, and carries work between business functions, all with no internet connection. Mickai® runs it on hardware your organisation owns, beside your own email server, meeting infrastructure and notes, so nothing leaves the building.
What tasks can an offline AI assistant do without any internet?
Anything that is words in and words out. Drafting, rewriting, summarising, comparing two versions of a document, pulling figures or obligations out of a file, answering a question from material you already hold, and handing a finished piece of work to the next function that needs it.
The reason this survives an unplugged network cable is that nothing about it is remote. The model weights sit on disk in your building. The index of your documents sits beside them. The runtime doing the work is a process on a machine you bought. When a member of staff asks a question, that question never becomes a request to somebody else's data centre.
What decides how well it runs is memory and a capable graphics card. Those two things set the size of model you can hold and how many people can use it at once. I will not tell you it runs on any hardware, because that is not true of any system. Models that fit comfortably on hardware an organisation owns are sized for the work described above, and for that work the difference matters far less than most buyers expect.
Can it answer questions from our own documents?
Yes, and for most organisations this is the capability that justifies the installation. You point it at document sets you choose: policies, contracts, procedures, past correspondence, technical files. Each indexed set is a private knowledge base.
The assistant searches the knowledge base, reads the passages that matter and answers with the document and section it drew from, so the person asking can open the source and check the work. Two things follow from that design. Access follows the permissions you already operate, so indexing a document does not make it readable by someone who could not read it before. And the index is a live thing: remove a document and it stops informing answers from that moment, with no retraining and nothing left behind inside a model.
Can it handle email, meetings and notes?
It works against your own email server, your own meeting infrastructure and your own notes, rather than a hosted copy of any of them. On email that means triaging an inbox, summarising a long thread before a call, and drafting a reply in the voice of the person who will send it. On meetings it means a transcript, the decisions taken and the actions assigned to names. Notes and team collaboration sit in the same installation, so a summary can become a task without leaving the building.
There is a line I want to be clear about. Reading, summarising and drafting are safe to automate. Sending is not. Anything consequential, and an outbound email to a client is consequential, waits for a named person to approve it before it happens. That is a design decision, not a limitation we are working around.
What can it do inside finance, legal and engineering work?
Each function gets a ready-made studio rather than a blank chat box, because the work in each is different. In finance the assistant drafts the narrative around numbers you supply: variance explanations, reconciliation notes, the written half of a board pack. In legal it compares clauses across contracts already indexed, extracts obligations and dates, and produces a first pass at a review that a qualified person then does properly. In engineering it drafts specifications, reviews configuration and code you hold, and writes the change notes nobody wants to write.
In every case the pattern is the same. The assistant does the reading and the first draft, quickly, against your own material. A person with the authority and the liability decides. I have never met a finance director who wanted a machine to sign off a return, and I did not build one that asks to.
What can an offline assistant not do?
It cannot tell you anything that is not in your building. No live web, no market feeds, no external lookups unless you deliberately bring that data inside and index it. If a workflow genuinely depends on the public internet in real time, cloud is the right tool for that piece and I will say so. Our argument is not with the companies building the compute and cloud layer. It is with the assumption that a regulated organisation must rent its intelligence, ship its data offsite and take a vendor's word for what happened to it.
Two specifics are worth knowing before you plan around them. Scanned documents and paper archives: a local reading runtime has handled scanned PDFs in controlled tests, but extraction and ingestion into the wider system is still being completed, so do not buy on the strength of it yet. And the assistant does not act on its own initiative. It proposes, a named person approves, and the approval is recorded.
Does an offline assistant need us to train it on our documents?
No, and this is the single most common misunderstanding I correct. The model is not trained on your files. Your files are indexed, and the relevant passages are read at the moment the question is asked. There is no fine-tuning project, no weeks of preparation and no document absorbed into a set of weights where nobody can find it again.
The practical consequences are worth spelling out. Indexing a new document set is an afternoon's task, not a quarter's. A correction to a policy takes effect on the next question, not the next training run. And when legal asks you to prove that a withdrawn document is no longer shaping advice, you can remove it from the index and show that it is gone. The ICO's guidance on AI and data protection is worth reading on why that kind of control over personal data in an AI system matters, because retrieval gives you it and training largely does not.
How does an offline assistant record what it did?
Every consequential action is sealed into the Open Audit Record. The entry is signed before the action runs, the log is append-only, and each entry is hash-linked to the one before it, so the order cannot be quietly rearranged afterwards. The signature scheme is ML-DSA-65, which the US National Institute of Standards and Technology published as FIPS 204 in 2024.
The part that matters to an auditor is the export. You hand over a record, and your auditor verifies the signatures offline with a public key using standard tooling that we did not write and do not control. That is the difference between an audit trail and a vendor's assurance.
It is tamper-evident, not tamper-proof, and the distinction is not marketing hedging. Tamper-proof would mean nobody can alter the file, which no honest engineer claims about bytes on a disk. Tamper-evident means that if anyone alters an entry, reorders the chain or removes a line, verification fails and the failure points at what broke. You do not get a promise that nothing happened. You get proof when something did. The NCSC's guidelines for secure AI system development and the UK government's Code of Practice for the Cyber Security of AI both treat logging and traceability as part of securing an AI system rather than an optional extra, and I agree with them.
How do we test the capability before we commit?
Run your own work through it. The closed beta is open and one regulated company is onboarding as a design partner, so there is a real installation to compare notes against, though I will not name the partner or its sector.
My advice on how to test is unglamorous. Take ten tasks from last week that actually consumed time: a contract review, a board-pack section, a long thread somebody had to summarise, a policy question three people asked separately. Run them on your own documents, on your own hardware, offline. Then count how many outputs a senior person would put their name to without rewriting. That number is your business case, and it is a better number than anything in a brochure. While you are there, export an Open Audit Record from the session and have somebody who does not work for us verify it.
If it helps to read further, what an offline assistant is and how internal knowledge search works cover the ground underneath this, one assistant across every department covers the cross-function question, and is air-gapped AI less capable than cloud AI takes the capability worry head on.
Frequently asked questions
Does an offline AI assistant still work if the building loses its internet connection?
Yes. The model, the index and the runtime all sit on hardware you own, so a dropped line changes nothing about drafting, summarising or searching your own documents. The only things that stop are anything you deliberately connected to an outside source. Staff carry on working, and the Open Audit Record keeps sealing actions as normal.
Can it draft in our own house style and templates?
Yes. Index your style guide, your approved templates and a set of documents you were happy with, and the assistant drafts against those rather than generic prose. Tone instructions can be set per function, so a legal letter and an internal briefing do not come out sounding the same. A named person still approves anything that leaves the organisation.
Will it take actions in our systems, or only suggest them?
It can take actions, but consequential ones wait for a named person to approve them first. Low-risk steps like preparing a draft, assembling a summary or indexing a file run on their own. Sending an email, filing a submission or changing a record needs a human decision, and that decision is signed into the Open Audit Record before it executes.
Delete this FAQ entry and its answer entirely, leaving five self-contained questions. The H2 "Does an offline assistant need us to train it on our documents?" already covers it.
No. Your documents are indexed, not used to train the model, and the relevant passages are read at the moment somebody asks a question. That means no fine-tuning project and no waiting. It also means a document you withdraw from the index stops informing answers immediately, which is far harder to demonstrate with a model trained on your data.
Can staff in different departments see different answers?
Yes, and they should. Each department works against the private knowledge bases it is entitled to, and access follows the permissions your organisation already operates. Indexing a document does not widen who can read it. Somebody in engineering asking about a contract gets only what their existing rights allow, and the request is recorded either way.
How do we know what the assistant did on our behalf?
Every consequential action is sealed into the Open Audit Record: signed before it runs, append-only, and hash-linked so entries cannot be quietly reordered. Signatures use ML-DSA-65, published by NIST as FIPS 204 in 2024. Your auditor verifies an exported record offline with a public key, using tooling we neither wrote nor control. It is tamper-evident: alter it and verification fails.
Related briefings
Deploying private AI
- Private AI for Business: What It Is and How It Works
- Private AI Deployment Options: Workstation to Air Gap
- On-Premise AI Server Requirements for a Normal Office
- How Sovereign AI Integrates With Your Existing Systems
- From Data to Answers, Without the Data Leaving
Documents and retrieval
- What Is RAG? Retrieval-Augmented Generation Explained
- How to Prepare Documents for a Private AI Knowledge Base
Part of a series of 60 briefings on deploying and governing AI in UK regulated organisations, archived with a DOI at 10.5281/zenodo.22975756.
Evaluating AI for a regulated organisation? Mickai runs on hardware you own, offline. Consequential actions wait for a named person to approve them, and what the AI did is sealed into a signed record an auditor can check without us. Applications for the invitation-only closed beta are open. Apply for the closed beta.
Written by Micky Irons, founder and chief executive of Mickai LTD.
Top comments (0)