OVIOO AIAsk 
: your company’s memory, on demand.
AI in OVIOO is infrastructure behind four ordinary screens. It proposes; a person confirms. It reads the little that carries work and leaves the rest indexed. And email never depends on it.
- Site access time not yet confirmed with Castellane SAOpen email →
- Vantage asked for the requirements list again this morningOpen email →
- Signed contract still with the customer — requested 24 AugOpen task →
Every line links back to the email, file or task behind it — nothing to take on faith.
Four places you meet it.
Named as they are on screen.
Ask OVIOO
A page and a rail. “Ask about a job, a supplier, a file, a date…” The answer cites the emails and files it was built from, and searches only what you are allowed to see. On the Business tier and above it reaches across every mailbox you may see; on Starter, your own.
Waiting for you
The rail beside the reading pane lists what OVIOO proposed from the mail you are looking at: a reference, a task, a link, a file to attach. Each one is a suggestion until you accept it.
Review
One screen for every proposal, with its evidence: the sentence it read, the code it matched, the rule that fired. Approve, change or dismiss. “Why?” on any suggestion shows exactly what it was based on.
The allowance
Employees never see tokens or prices — the header shows a plain percentage and when the allowance renews. Owners see usage by person and by kind of action, and set caps per person or per mailbox.
Three gates before a model reads anything.
In ascending order of cost. This is why the allowances on the pricing page can be as large as they are.
Free filters
Bulk headers, sender history and the owner’s blocklist stop newsletters, notifications and known noise before any model sees them. Most mail ends here at no cost.
Triage, small model
What passes is read in batches by a small, cheap model that only decides one thing: does this message carry work? A thread that does not is indexed, searchable, and left alone.
Full read, only where it counts
Messages that carry work get the full extraction: tasks, dates, amounts, the reference they belong to. This is where credits are spent, and it is a minority of the mail.
Credits, not surprises.
One credit is a fixed slice of provider cost. Triage of one message is about one credit; a full read with extraction one to three; a question with retrieval five to forty. Credits come with every seat and pool across the organisation.
Three independent safety layers: a monthly allowance with warnings at 80% and 95%; a daily cost ceiling per organisation, after which incoming mail simply queues for tomorrow; and the rule that email itself never depends on AI. The model behind each kind of task is configuration, not code — the vendor can change without a release.
What every employee sees. No tokens, no prices.
| Tier | Credits / seat / month | Ask scope |
|---|---|---|
| Starter | 1,500 | Own mail |
| Business | 4,000 | Whole workspace |
| Scale | 10,000 | Whole workspace |
Owner console: usage by person and by action; caps per person and per mailbox; top-up packs, never auto-bought.
What it never does.
Written down, because the fear is reasonable.
- Sends, replies, forwards or deletes anything.
- Links an email to a reference without a code or your confirmation.
- Reads history older than the analysis window unless you ask it to.
- Spends past the daily ceiling — mail waits for tomorrow instead.
- Sees a mailbox you do not have access to.
Where the intelligence stops
The limits are the product. An assistant that acts on its own is one you have to supervise, which is not help.
It proposes; a person decides
Tasks, links, extracted amounts and drafted replies all arrive as proposals in one review screen, each carrying the sentence it came from. Accept, edit or reject. There is no path by which a proposal becomes a sent message, a deleted record or a linked job without somebody choosing it.
Deterministic first, always
A message carrying a reference is linked by an indexed lookup on the number, not by a model. The intelligence is spent on what is genuinely ambiguous, which is both more accurate and dramatically cheaper than reading everything.
Three gates before a model sees anything
Free filters first — bulk headers, sender history, the owner’s blocklist. Then triage on a small model to decide whether a message carries work. Only what survives both is read in full. Most of a mailbox never reaches the expensive step, because most of a mailbox is not work.
A hard ceiling you set
There is a spend limit per day, and over it mail queues and is picked up tomorrow rather than costing more. The owner sees what has been spent, on what, and can raise or lower the limit. AI spend is owner-reserved and granted to an admin one person at a time.
Answers cite what they were built from
Every answer lists the messages, documents and activity behind it, and each one opens. It only ever searches what the person asking may already open — you cannot learn the contents of a job you have no access to by asking about it a different way.
When it is gone, email is not
If the allowance runs out, the key is missing or the vendor is down, the product says so plainly and everything else carries on: reading, writing, sending, searching, references, tasks, documents. Degrading quietly would be worse than stopping.
Fair questions
No. Your content is not used to train models, and the vendor is a configuration choice rather than a rewrite — OpenAI is the baseline and Anthropic is supported, because a change of vendor should be a setting, not a project.
Link a message to a job when the message carries that job’s reference number — which is a lookup, not a decision. Everything else waits: tasks, extracted amounts, suggested links, drafted replies. In Review-everything mode even the deterministic matches wait.
Spend is metered per call against published model rates and shown to the owner by day and by task. The rates are checked against the provider’s own dashboard and rounded up, so the ledger errs high rather than flattering itself.
You see the proposal and the sentence it came from before anything happens, and every rule action is logged with an undo. The design assumes it will be wrong sometimes; that is why nothing it decides is final.
See it around your own mailbox.
Connect the email you already have and watch the structure appear. Trial for the whole organisation.