It does the work. You keep the last word.
Oarlock agents take on tickets and routine tasks inside the tools you already use. Every run is written down step by step, and anything that moves money stops for a person.
- Read-only until you say otherwise
- Live on one queue in the first week
- Pay per resolved task, not per seat
- Read the ticket0.6s09:14:02 helpdesk.read
Damaged on arrival, wants a refund. Photo attached, no order number given.
- Look up the order1.2s09:14:02 orders.find
Found by email. Order 88214, one kettle, $64.00, delivered 3 days ago.
- Draft the refund1.4s09:14:03 policy.search, payments.draft
Damaged items are refunded in full with no return. Draft for $64.00, no money moved.
- Ask a person to approve2m 41s09:14:05 approval.request
Refunds over $25 need sign-off, so the run stops here and waits.
Renata Oyelaran approved from the team queue$64.00 - Send the reply0.9s09:16:46 payments.refund, helpdesk.reply
Refund issued, reply sent, ticket closed.
Repeat work, with a clear place to stop
Customer support
Checks tracking and replies with a real date.
Applies your returns policy and drafts the refund.
Updates the record and confirms with the customer.
Finds invoices and explains a charge.
Back office
Matches supplier invoices to purchase orders.
Sends polite follow-ups on late deliveries.
Finds duplicate customers and proposes merges.
Checks returned stock against refunds paid.
Permissions live outside the model
The agent can only reach the tools a policy gives it. Rules are checked on every call, however the request is worded.
| Tool | Can read | Can write | Needs approval |
|---|---|---|---|
| Your helpdesk | Tickets, history | Replies, tags | None |
| Your orders database | Orders, tracking | No | None |
| Your payments system | Charges | Refund drafts | Over $25 |
| Your shipping system | Parcels | Replacement labels | Always |
| Your customer records | Contact, plan | Address, notes | After dispatch |
| Your knowledge base | All articles | No | None |
Spend and action limits, run 4182
Always needs a person
Audit trail Every call, every approval, every policy change. Export any time.
It works where your team works
No new inbox and no migration. The agent signs in to your tools with its own account, so its work shows up beside everyone else’s.
- Its own login, scoped to what the policy allows
- Keys you can revoke from your side at any time
When it stops, it hands over properly
Your team never starts from a blank ticket. The customer is told a person has it, and the note arrives with the work already done.
- why i stopped
- The customer mentions going to their bank about this charge.
- what i found
- Order 88560, two parcels. One delivered, one still in transit.
- what i did
- Nothing sent except “a colleague will reply today”.
- i would suggest
- Confirm the second parcel’s date, then offer to refund it.
Trust goes up one step at a time
You stay on each step until you are happy with the runs. Most teams loosen approvals over the following weeks.
- Day 1Connect, read-only
Link your tools with read access. Nothing can be changed yet.
can read - Day 2Load your rules
Add your policies and saved replies, and set the first limits.
can read - Day 3Replay old tickets
It works through last month’s closed tickets. Your team marks each run.
can draft - Days 4 to 5Shadow live tickets
It drafts on real tickets and your team decides what gets sent.
can suggest - Days 6 to 7One queue live
One simple queue goes live, with every action behind approval.
can act, with approval
Test a change on last month first
Open any past run, step through it, and re-run it from any step. Before a policy goes live, replay real tickets against it and read what would have changed. Replays never touch your live tools.
Your team marks runs as good, wrong or needs a rewrite. Wrong ones join a test set that every future change has to pass.
| Run | Live, v13 | Draft, v14 | Result |
|---|---|---|---|
| Kettle lid, refund $64Run 4182 | Stopped for approval | Stopped for approval | Same |
| Mug set, refund $32Run 4097 | Stopped for approval | Refunded without asking | Looser |
| Late parcel, refund $38Run 4110 | Stopped for approval | Refunded without asking | Looser |
| Damaged lamp, replyRun 4151 | Offered a replacement | Offered a refund first | Reply changed |
| Address after dispatchRun 4033 | Stopped for approval | Stopped for approval | Same |
You pay when a task gets finished
per resolved task, with no seat fees and no setup fee.
- Unlimited team seats and approvers
- Every tool connection included
- Full run logs and replay
- First 100 resolved tasks free
| Resolved tasks a month | Price per task |
|---|---|
| First 1,000 | $1.20 |
| 1,001 to 5,000 | $0.95 |
| Over 5,000 | $0.75 |
What counts as resolved
The task is finished and the ticket stays closed for 72 hours. If it reopens, the charge comes off.
What you never pay for
Handoffs to your team, failed runs, shadow drafts and replays.
Monthly cap
Set a ceiling. When it is reached, new tickets go straight to your team.
Leaving
Monthly billing, cancel any time, and take your run logs with you.
What teams ask before switching it on
What happens when it isn’t sure?
It stops and hands the ticket to your team with a note on what it found and what it would do next. New queues start with every action behind approval.
What if a customer tries to talk it into something?
Text in a ticket is treated as information, never as an instruction. Permissions sit outside the model, so a persuasive message can’t unlock a tool or raise a limit.
Which AI models do you use?
Each step goes to a model that suits it, and the run log records which model and version did the work. You can pin a version per queue.
Is our data used to train anything?
No. Tickets and run logs stay in your workspace, and you choose how long they are kept.
How do we switch it off?
Anyone with admin rights can pause a queue or the whole workspace with one button. Open runs hand back to your team.
Who is responsible if it gets one wrong?
You set what it may do, and everything it did is on the record. If a run broke your policy, we credit the task and help you trace why.
Does it replace our support team?
It takes the repeat work. People keep the judgement calls, the upset customers and everything it isn’t allowed to do alone.
What comes back
- 1A full run log, like the one at the top of this page
- 2The reply it would have sent, and where it would have stopped
- 3A short list of what we would need to connect
It runs in a sandbox with nothing connected, so nothing reaches your customer. We reply within one working day.