Reads shared inboxes, drafts replies, files attachments.
mail.search(to:"billing@")Relay: hand the repetitive work to AI agents
Agents for operations teams
Hand the repetitive work to agents that show every step, and wait for your OK when it matters.
01 · The pile
Monday, 9:04. Everything is urgent and none of it is interesting.
Invoices to match against purchase orders. New leads to enrich. Tickets to tag and route. The Friday report, rebuilt by hand again. None of it is hard. There is just always more of it.
- Match invoices to POs
- Enrich new leads
- Tag support tickets
- Rebuild the Friday report
- Chase missing receipts
Act II
Agents claimthe work.
02 · How it works
Describe the job once. An agent takes it from there.
- 01
Brief it like a colleague
“When an invoice arrives, match it to the PO and flag anything that is off by more than 2 %.”
- 02
Give it tools and limits
Choose what it may read, what it may change and when it has to ask. Every agent has one role.
- 03
Read the log, not the pile
Every tool call, input and output is recorded and replayable, so you always see why something happened.
03 · Integrations
Works inside the tools you already pay for.
Agents use the same apps your team does, with only the permissions you grant. No new inbox, no new spreadsheet.
Calendar
Proposes slots, books rooms, moves the meeting nobody needs.
calendar.propose(slots: 3)Spreadsheets
Matches rows, appends results, keeps a tab of every change.
sheets.append(tab:"POs")CRM
Enriches new leads, logs calls, updates the deal stage.
crm.update(stage:"qualified")Help desk
Tags and routes tickets, answers the ones with known fixes.
tickets.tag(id, "billing")Docs
Reads contracts and policies, writes summaries you can cite.
docs.read("contract_v3")Data warehouse
Runs read-only queries and builds the Friday numbers.
sql.query(readonly: true)Team chat
Posts results, asks for approvals, answers “is it done yet?”.
chat.ask(approver, task)04 · Human in the loop
Nothing important leaves without your OK.
Draw the lines an agent may not cross alone. Above them, the task waits at the gate with its reasoning attached, and you approve it in one click from the app, email or chat.
- Refunds over €500ask the finance lead
- First email to a new customerask the account owner
- Anything that deletes dataalways ask
Waiting for you · 00:42
Refund €740 on order 4471. Drafted by the Ledger agent: duplicate charge on 12 Sep, same card, same amount.
05 · A sample week
The pile, done and filed.
Every finished task lands in the outbox with its log attached. Here is what one demo workspace looks like after a week.
- 0tasks handed off to agents
- 0approvals a person actually gave
- 0 minmedian from arrival to done
Illustrative figures from a demo workspace, not customer data.
How teams use it · illustrative
Less chasing, more deciding.
“Month-end used to take me three days of matching. Now I read one log and approve the eleven things that need me.”
Maya · finance lead · illustrative example
“Tickets arrive tagged and routed before anyone is online. The team starts the day on the hard ones.”
Tomás · support operations · illustrative example
06 · Pricing
Start free. Pay when agents do real work.
Starter
€0forever
- 1 agent, 300 runs a month
- Mail, calendar and sheets
- 7-day run log
Team
€29per seat / month
- Unlimited agents, 10,000 runs a month
- Every integration, approval rules
- Full audit log, 1-year retention
- Shared playbooks for the team
Scale
Custom
- Private deployment and data residency
- SSO, SCIM and custom roles
- Named engineer, uptime agreement
Questions
Before you hand anything over
Can an agent do something I did not allow?
No. Each agent only gets the tools you connect and the permissions you tick. Anything above your approval rules stops at the gate and waits for a person.
Where does our data go?
Runs execute in your workspace region. Tool inputs and outputs are stored in your run log for the retention you choose, and are never used to train models.
What happens when an agent is unsure?
It says so. Low-confidence results are routed to a person with the reasoning and the sources attached, instead of being guessed.
Do we need engineers to set it up?
No. Agents are briefed in plain language and connected with sign-in buttons. Engineers can add custom tools through the API if they want to.
What counts as a run?
One task from arrival to done, however many tool calls it takes. Failed runs and runs you reject at the gate are not counted.