ADA Shortlist
Candidates worth investigating against our own desktop-agent evaluation framework.
The ADA benchmark
The ADA benchmark is our own evaluation framework for desktop agents, not an industry standard.
An autonomous desktop agent observes a person’s everyday work and suggests repetitive actions to automate. The user reviews and corrects each proposal before a trial.
- Observe ordinary Windows work
- Find repetition and propose a microtask
- Let the user correct the procedure and approve a trial
- Use AI on a changed case and verify the result
- Reuse approved microtasks and test combined workflows
- Run unattended and report exceptions
- Support fully self-hosted data, AI and execution
Price is a separate consideration. An end-to-end match needs evidence across the same workflow, including a successful run on a changed case.
Candidates below can cover different parts of the workflow. This is not a scored ranking. Full self-hosting remains unverified for every candidate below. Price is a separate consideration.
| Candidate | Why we shortlisted it | Evidence still needed |
|---|---|---|
| Automation Anywhere Process Discovery, Automation 360 and Automator AI | Provider documentation covers discovering repetition in desktop activity and developing reusable automation with AI fallback. | One mined task taken through user correction and trial approval, successful AI adaptation to a changed case, verified business results, combined microtasks and unattended exception reporting. |
| Mimica Miner, Mapper and Maker | Provider describes discovery, process mapping and agent generation with approval and unfamiliar-case alerts. See all seven criterion assessments. | A completed Windows workflow showing procedure correction, changed-case AI execution and independent result verification, reuse of approved microtasks, combined-workflow tests and unattended exceptions. |
| Mediar | Its provider-described pilot centres on reviewing and correcting recorded actions, replaying workflows and checking outputs across input cases. | Discovery and proposals from ordinary work, rather than a chosen recording; completed changed-case evidence, verified outcomes, tested microtask composition and unattended exception handling. |
| OpenAdapt | Its repository documents approval of compiled procedures, reusable execution, saved-record checks and stops for uncertainty. | Automatic repetition discovery and proposals; Windows evidence connecting changed-case AI repair to approval and verified results; tested combined workflows and sustained unattended operation. Local replay alone does not establish local AI for learning and repair. |
| UiPath Task Mining and execution components | Documentation covers task capture, collaborative trace correction and automation export; our comparison also considers its authoring and robot components. | The reviewed Task Mining configuration captures known tasks. We need discovery-to-execution evidence across the same approved workflow, including changed-case adaptation, outcome verification, microtask reuse and exception reporting. |
What would establish a full match?
One traceable workflow must cover all seven ADA criteria: ordinary Windows observation; repetition discovery and a microtask proposal; user correction and trial approval; AI execution on a changed case with a verified result; reuse and combined-workflow tests; unattended execution and exception reporting; and fully self-hosted data, AI and execution.
A successful run on a changed case is required. Separate feature demonstrations or provider claims do not establish an end-to-end match.
Evidence status: linked source statements have been reviewed. Product behaviour and complete workflow results have not been verified by us. The comparison’s source reviews are dated 10–18 November 2026; product capabilities can change.