Product
Operations Directors and Office Managers at mid-market professional firms, plus adjacent desks with the same leak fear — bookkeeping, property management, and recruitment ops. If clerks already paste client work into ChatGPT, this is the private upgrade: ChatGPT speed. Private UK node. Named unlock before anything client-facing leaves. Not a chatbot wrapper. A2Zai is administrative desk software — private intake, drafts, and named unlock. Not legal AI. Not regulated professional advice.
Long files finish without truncation. Client-facing drafts wait for a named unlock. Inference stays on your private UK node — not in a public chat model.
If your team already uses ChatGPT for client admin, start with Boutique (£449) — ChatGPT speed. Private UK node. Named unlock before anything client-facing leaves. Book a 45-min desk demo; paid pilot £299 credits to setup.
Messy enquiry → structured fields + holding reply
Clerks copy/send it. A2Zai does not transmit mail.
Inbound lead
Website form · new enquiry
Hi — please confirm receipt. We need a holding reply on the lease query for 14 High Street (typos in the mobile form). Access notes in the thread below. Thanks.
Structured extract · draft desk
Holding acknowledgement drafted for Admin Review. Silence over invention.
Give one goal — the worker decomposes it, calls tools in parallel, observes results, and self-validates before it stops. Replans when the outcome misses the objective. Feels like a digital colleague, not a single-turn chat box.
Large emails, PDFs, and multi-doc dumps are chunked, extracted, and consolidated under a VRAM-safe map-reduce path before they hit the context ceiling — so the desk model finishes the file instead of silently cutting mid-document.
Messy emails and forms become clerk-ready fields and non-binding acknowledgement drafts. Background jobs keep long work off the UI thread so desks stay responsive.
Internal categorisation and extraction can run fully automated. Client-facing sends, financial drafts, and issue unlocks route to Admin Review — fail-closed by code, not by hope.
Episodic session state tracks the live workflow; long-term firm memory injects preferences and closed-task summaries when you enable durable retention. Default is RAM-first privacy, segmented per organisation.
Start on Boutique (up to 5 seats) on Staffordshire hardware with fair-queue concurrency. Scale to Regional (£521–£1,728/mo, 6–24 × £72) or Enterprise (£1,800–£3,600/mo, 25–50 × £72) on a private cloud GPU when the office grows — your training box stays separate. Never “all seats interactive at once.”
Selling now: Accountancy practices · Estate agencies · Bookkeeping desks · Property management · Recruitment ops · Professional service offices. Same paste-into-ChatGPT leak fear across the list — not “all SMEs,” not legal AI.
Industry landing pages later (packs already in code): Logistics desks · Commercial interiors & trades · Retail & dealerships. Channel next: accountant networks / MSPs with a referral fee — see Partners.
Isolated multi-tenant node architecture. By default, client matter text is processed inside an isolated private data container on United Kingdom hardware in Staffordshire and does not train public models. Public-cloud inference overflow stays off unless you explicitly enable a named overflow path. Tenant borders are enforced in code. Edge login and card payments use named processors (Cloudflare Access; Stripe) — they are not the inference model.
Where De-ID is enabled for your organisation, UK-focused pattern scrubbing runs before the model (emails, phones, NI numbers, postcodes, money, titled/labelled names, common org suffixes, street lines) with a session-only reverse map for admins — not full NER. Residual identifiers can remain; treat output as still confidential.
Documents Stay-on-Page is extractive: fill from the uploaded source only (no free-form invention of parties or figures). Chat Stay-on-Page is generative but buffered — invented assertions are stripped or refused; it is not the same as Documents extractive mode.
In-progress drafts default to session/RAM and are not kept on host disk across reboot. Durable on-disk retention is optional and only when enabled for that engagement. Account, billing, and organisation records are retained to run the Service.
De-ID events, AI drafts, completed amber checks, sole-approver acknowledgements, and admin unlocks are hash-chained and exportable as an operator audit pack PDF — an operational trail for Practice Managers, not a training certificate.
Upload internal templates; the node builds a structure/style brief (headings, openings, placeholders) for drafting prompts. That is not per-organisation model fine-tuning and is never used to train public models.
Where Admin Review is enabled for a workspace, drafts stay locked until a named human checks amber flags and unlocks issue. Cool-down applies before unlock. Autonomy handles the work; humans remain responsible for regulated or client-facing send. Bookkeeping / general admin desks may use named self-unlock where configured.
Boutique Team (£449 / month, up to 5 seats) is the hero SKU: desk work on your private UK node, named unlock before client issue. Scale to Regional or Enterprise when the office grows. One leaked client file usually costs more than Boutique Team (£449 / month). A2Zai is the private desk that keeps the speed and removes the paste-into-ChatGPT habit.