Exception Queue Sprint
Automate the work your bots hand back.
Your bots handle the clean, rules-based part of the process. In two weeks, we put a live computer-use agent on one of your exception queues, in your sandbox, and report its completion rate.
- 85.2% on OSWorld with Holo4 27B, at $0.08 per task
- SOC 2 Type II
- Read-only by default, outbound-only connection
Your exception queue, worked by an agent
Odd formats, portal changes and Citrix screens still land on people. The Sprint shows how much of that queue an agent can clear, on your systems, with your data.
01
A live agent on one queue
Working real exceptions in your sandbox: portals, legacy screens and documents your bots can't read.
02
Completion-rate report
How many items the agent closed, how many it routed to a person, and why.
03
Human review where it matters
Confidence thresholds you set, so judgment calls always reach a reviewer.
04
A launch plan for that queue and two more
Scope, volumes and KPIs for a production launch, agreed with your process owner.
One queue. Two weeks. A number you can take to your sponsor.
Days 1-2
Pick the queue
Choose one exception queue with clear success criteria, ideally 1,000 or more items a month. We agree the KPIs with your process owner.
Days 3-9
Agent works the queue
A forward deployed engineer sets the agent up in your sandbox with a user-level account. Read-only first; write actions only once you approve them.
Days 10-14
Results and next step
Completion rate, failure reasons and a launch proposal for that queue plus two more.
This isn't rip-and-replace
Your bots keep running. Agents take the items that bots hand back today, so your team spends its time on the decisions that need them.
| Situation | Bots today | With an H agent |
|---|---|---|
| Clean, structured items | Handled by bots | Still handled by bots |
| Odd formats and documents | Exception queue | Read and processed |
| Portal or screen changes | Bot fails, item queued | Agent adapts and continues |
| Judgment calls | Person decides | Person decides, with the work prepared |
The Sprint is built for teams that
- 01
Have a measurable exception queue
Roughly 1,000 or more items a month that bots route to people.
- 02
Can provide a sandbox within two weeks
A test environment that mirrors the systems the queue touches.
- 03
Have a named sponsor and a process owner
Someone to agree success criteria and someone who knows the work.
Book a 20-minute scoping call
We look at your candidate queues, confirm sandbox access and pick the one with the clearest success criteria. Then we schedule the two weeks.
Useful to have on the call
The exception queue you'd pick, and its monthly volume
The systems it touches (portals, Citrix, desktop apps)
Who owns the process day to day
What teams ask before they start
We scope it on the first call. For qualified teams the Sprint is free or low-cost; the goal is a completion rate your sponsor can act on.