AI that does the work, and asks before it acts.
We map what is worth automating, then build it into the systems you already run. Reads and drafts happen on their own. Anything with consequences waits for a person. Try it below.
I'm a demo of the assistant we build. Ask me about yesterday's sales, the inbox, stock counts, or say clean up. The data is invented; the rules are real.
stores reporting through systems we built, every week.
users on a ticketing app with no per-seat licence.
unaudited writes. Nothing leaves without a name on it.
every morning, a brief with the replies already drafted.
One rule,
every system.
The model may propose. Only a deterministic policy layer may authorise. Only registered tools may execute. Scroll to watch one action go through it.
- 01
Read
Search, summarise, classify. No approval, because nothing changes.
- 02
Propose
Draft the email, the ticket, the order. Show exactly what would happen.
- 03
Approve
A person confirms the exact action. Sends, deletions and payments never skip this.
- 04
Audit
Every step logged with who approved it, so you can answer for it later.
- 31 unread across 4 mailboxes
- 3 need you
- 9 can wait
- 19 noise
Hi Nadia, confirming payment goes out Friday. Could you send the remittance address for the new entity? Thanks, Drew
Where
to start.
Most start with the plan, because four weeks is the cheapest way to find out whether the rest is worth doing. If you already know what you want built, start there instead. Either way the plan is written by the people who will build it, so it never promises what cannot ship.
Strategy
Use-case discovery and roadmap
Four weeks with leadership and the people doing the work. A ranked roadmap, the first pilot scoped and priced, and a list of things we would leave alone.
Governance and compliance
A risk assessment, a framework people will follow, and the POPIA, GDPR and ISO material you need. Policies people read, not binders.
Team enablement
Role-specific sessions: what the tools do for finance, ops or support, what must never be pasted into one, and how to check the answers.
Build
Ask your data a question
Plain-language questions over point of sale, ERP, reports and spreadsheets, with every answer traced to the source rows.
Operations automation
Reporting, reconciliation, stock counts, ticketing, document handling. Real software with logins, audit trails and no seat fees.
Assistants that act
Connected to mail, files and tasks. Proposes actions, executes only what a named person approves. Provider-neutral, so no vendor lock-in.
Recent work
Anonymised. Counts are the client's own; figures inside the visuals are illustrative.
Delivery reconciliation for a 78-store QSR group
Weekly exports from two delivery platforms turned into order-level reporting: late, missed or misreported orders, and what the platforms paid versus what they claimed.
Point of sale you can talk to
The group's POS connected to an assistant, so the ops team asks about yesterday's sales at any store in plain English.
Ticketing freed from per-user licences
Repairs and maintenance rebuilt as an installable web app for 250 users across three email domains, with photo capture and offline queueing.
Stock counts on phones
Paper and spreadsheets replaced by mobile forms that know which stores belong to which region, feeding the inventory system directly.
Azani.
The assistant we built for ourselves and use every day.
07:00. It reads.
One consolidated mailbox, sorted into what needs a reply, what can wait, and noise.
It drafts.
Each reply written in the owner's voice, with the thread it answers underneath.
You tap.
Nothing sends without approval. The tap is the approval.
It is logged.
Who approved what, and when. Kept for 90 days, then purged.
Built like software,
not a chatbot.
Six things that are true of it, and of everything we build.
Nothing sends without a tap.
Send, save, delete: every write waits for a named person. Reads and drafts happen on their own.
Everything is logged.
Every step, with who approved it and when, so you can answer for it later.
Passkeys, no passwords.
Nothing to phish, nothing to reuse.
Model swappable.
Anthropic and OpenAI pass the same tests. No rebuild to change.
Your own cloud.
Google Cloud, Johannesburg, in a project you own. No stored model key; identity federation instead.
Kept 90 days, then purged.
Voice is discarded the moment it is transcribed.
Four of our own.
Products we built and run.