AI Intelligence

Windows desktop automation

Desktop automation

Which products can help turn the way I do a Windows task into a repeatable workflow?

These 19 entries compare individual capabilities. Each row describes what its evidence covers and who published it. Date added is the first appearance in this comparison. Last checked records the official-source review, not a hands-on product test.

All 133 criterion ratings were reviewed against provider documentation on 10 Nov 2026. Yes means explicitly documented by the provider. Partial identifies a narrower capability. No applies to the described product/configuration. Unverified means the reviewed source does not establish the requirement. These are documentation assessments; independent end-to-end testing remains outstanding. Each rating includes its limitation and a benchmark source.

19 of 19

19 tracked products and projects · Last checked 10 Nov 2026

Record ID
Product or project
Benchmark
Organisation
Category
Availability
Compared with ADA
Deployment
Date added
Last checked
Evidence
Video and publisher
What it covers
1Mediar
  1. Observe everyday work – Unverified Deliberate recording; automatic mining unproved
  2. Detect repetition – Unverified No recurring-task detector established
  3. Human approval – Partial Workflow inspection documented; trial approval unproved
  4. AI execution & verification – Partial AI recovery; business-result verification unproved
  5. Reuse microtasks – Partial Reusable workflows; automatic incremental growth unproved
  6. Run unattended – Partial Managed execution; full exception process unproved
  7. Fully self-hosted – Unverified Local VM does not establish local AI
Benchmark source
MediarWorkflow automationContact salesRecords Windows workflows and turns them into automations with validation and recovery when controls move. Learning a new business rule from an operator’s correction remains unverified.Windows applications; hosting terms unconfirmedOfficial sourceAutomation of SAP B1 payments reconciliation · Mediar & FazmProvider-hosted example of payments reconciliation in SAP Business One, embedded on Mediar’s site.
2Terminator — Mediar
  1. Observe everyday work – Unverified Workflow recording is explicit
  2. Detect repetition – Unverified No automatic candidate mining established
  3. Human approval – Unverified Engine-level review and approval loop unproved
  4. AI execution & verification – Partial AI recovery documented; outcome checks unproved
  5. Reuse microtasks – Partial YAML workflows; tested suite growth unproved
  6. Run unattended – Unverified Scheduling and exception handling unproved
  7. Fully self-hosted – Unverified Local engine alone does not establish local models
Benchmark source
MediarAutomation enginePublic repositoryMIT-licensed Windows engine behind Mediar. Supports human workflow recording for deterministic automation and exposes Windows controls to AI agents.Local Windows engineOfficial sourceLouis Beaumont on automating repetitive workflows in legacy Windows software · CohereA project developer presents on an external channel. The speaker’s affiliation matters when assessing independence.
3OpenAdapt
  1. Observe everyday work – Unverified Demonstration capture; everyday mining unproved
  2. Detect repetition – Unverified No automatic repetition proposal established
  3. Human approval – Yes — provider documented Review, canary and admission documented
  4. AI execution & verification – Partial Save verification documented; ordinary replay uses zero model calls
  5. Reuse microtasks – Partial Compiled program reuse; automatic suite growth unproved
  6. Run unattended – Partial Automatic execution and halt; recurring operations unproved
  7. Fully self-hosted – Unverified Own model key does not establish full local AI
Benchmark source
OpenAdaptAIWorkflow automationPublic repositoryTurns a demonstrated Windows task into a local program that can repeat steps, check a save and stop on a mismatch. Ordinary replay runs without model calls.Local replay program; model requirements vary by stageOfficial sourceQuick demo of OpenAdapt · OpenAdaptAIProject demo of desktop automation. The release shown matters when comparing the footage with current replay behaviour.
4Sola
  1. Observe everyday work – Unverified Recorded workflow generation
  2. Detect repetition – Unverified No everyday repetition detector established
  3. Human approval – Partial Visual editor; approval gate unproved
  4. AI execution & verification – Partial Adaptive execution; independent outcome check unproved
  5. Reuse microtasks – Partial Composable workflows; incremental learning unproved
  6. Run unattended – Partial Orchestration documented; exception handoff unproved
  7. Fully self-hosted – Unverified Full local AI and control plane unproved
Benchmark source
SolaWorkflow automationVendor demoTurns a screen recording into a bot with editable logic. The provider describes desktop and browser workflows, adaptation to new events and scheduled work in financial software.Desktop and browser workflows; hosting terms unconfirmedOfficial sourceHow WSI rapidly scales their brokerage with Sola AI · SolaNamed customer example published by Sola. A complete recording-to-replay sequence remains unverified in this video.
5Automation Anywhere — Automation 360 and Automator AI
  1. Observe everyday work – Yes — provider documented Process Discovery captures Windows employee activity
  2. Detect repetition – Yes — provider documented Patterns and automation opportunities identified
  3. Human approval – Partial Admin fallback controls; user trial gate unproved
  4. AI execution & verification – Partial AI fallback; business-result verification unproved
  5. Reuse microtasks – Partial Platform workflows; incremental mined suite unproved
  6. Run unattended – Partial Unattended fallback documented; complete exception flow unproved
  7. Fully self-hosted – Unverified Customer AWS option does not prove all-local cycle
Benchmark source
Automation AnywhereRPA platformCommercial platformProcess Discovery observes work to identify candidates. Generative Recorder builds automations with an AI fallback during execution. Windows Unattended Bot Runner executes workflows.Automation 360 with Windows Bot RunnerOfficial source

Generative Recorder and resilient automation · Automation Anywhere

Vendor presentation and shorter release video cover recording and resilience. Evidence for these separate capabilities does not establish one continuous run.
6UiPath — Studio, ScreenPlay, Healing Agent and Unattended Robot
  1. Observe everyday work – Partial Task Mining captures selected known tasks
  2. Detect repetition – Partial Clusters traces; analyst selects opportunities
  3. Human approval – Partial Collaborative review; full trial approval loop unproved
  4. AI execution & verification – Partial Healing and inspection; outcome validation needs workflow design
  5. Reuse microtasks – Partial Export skeletons; automatic suite growth unproved
  6. Run unattended – Partial Unattended robots; benchmark-specific exception loop unproved
  7. Fully self-hosted – Unverified Cloud AI components; full self-hosting unproved
Benchmark source
UiPathRPA platformCommercial platformTask Capture exports a prototype for Studio. ScreenPlay directs desktop actions with AI, Healing Agent can repair supported interactions, and unattended robots run published workflows.Studio authoring and robot execution; options vary by componentOfficial source

Healing Agent and ScreenPlay · UiPath

Step-by-step self-healing demo · pikorpa

Vendor presentation and creator walkthrough of execution and recovery. Healing coverage depends on the activity.
7Microsoft Power Automate for desktop
  1. Observe everyday work – Partial Task mining uses intentional recordings
  2. Detect repetition – Partial Analysis supports candidates; passive automatic suggestions unproved
  3. Human approval – Partial Flow editing; approval gate unproved
  4. AI execution & verification – Partial AI self-healing; business outcome checks require authoring
  5. Reuse microtasks – Partial Flows and subflows; incremental mining integration unproved
  6. Run unattended – Yes — provider documented Documented unattended cloud-triggered runs
  7. Fully self-hosted – No — described configuration Described AI and orchestration depend on cloud services
Benchmark source
MicrosoftDesktop flow builderReleased; licensing variesRecords and edits desktop flows. Unattended execution requires appropriate licensing. The supplied comparison describes AI self-healing as a preview and flags older Record with Copilot demos as historical.Windows desktop runtime; cloud orchestration for unattended flowsOfficial source

How to record actions · Anders Jensen

How to set up unattended robots · Anders Jensen

Separate practitioner tutorials on recording and unattended setup. Check release and licence when comparing capabilities.
8Microsoft Copilot Studio — computer use
  1. Observe everyday work – Unverified Instruction-led operation; everyday mining unproved
  2. Detect repetition – Unverified No automatic repetition proposals established
  3. Human approval – Unverified Reviewable logs; procedure approval gate unproved
  4. AI execution & verification – Partial AI Windows actions; business-result verification unproved
  5. Reuse microtasks – Partial Reusable tool; mined microtask suite unproved
  6. Run unattended – Unverified Background capability alone leaves full exception loop unproved
  7. Fully self-hosted – No — described configuration Copilot Studio service remains cloud-managed
Benchmark source
MicrosoftComputer-use agentDocumented feature; access variesAn agent follows written instructions, examines the desktop and operates apps. Screenshots and execution logs let the user inspect the steps.Copilot Studio with a connected computerOfficial source

Computer use in Copilot Studio · Microsoft Copilot Studio

Step-by-step live demo · Microsoft Techy Tips

Instruction-driven computer use. This evidence addresses desktop operation, while discovery of repetition is a separate question.
9Microsoft Record-and-Run
  1. Observe everyday work – Unverified Intentional demonstration recording
  2. Detect repetition – Unverified No repetitive-action discovery documented
  3. Human approval – Partial Editable playbook; enforceable approval unproved
  4. AI execution & verification – Partial Vision-based task variants; independent result check unproved
  5. Reuse microtasks – Partial Saved playbooks; suite composition testing unproved
  6. Run unattended – Unverified Interactive run; scheduling and exception loop unproved
  7. Fully self-hosted – Unverified Documented model API; full local configuration unproved
Benchmark source
MicrosoftResearch prototypePublic repositoryRecords Windows actions, generates a written playbook and uses a vision model to execute a changed task. The documented setup uses a model API.Windows workstation and model APIOfficial sourcePaired demonstration and execution videos · Microsoft (GitHub)Compare the original demonstration with the changed case side by side. YouTube match unconfirmed.
10Microsoft UFO²
  1. Observe everyday work – Unverified Demo retrieval; passive observation unproved
  2. Detect repetition – Unverified No automatic repetition proposal established
  3. Human approval – Partial Interactive controls; complete procedure approval unproved
  4. AI execution & verification – Partial Windows planning/actions; independent result check unproved
  5. Reuse microtasks – Partial Experience reuse; incremental suite testing unproved
  6. Run unattended – Unverified Run completion differs from tested recurring operations
  7. Fully self-hosted – Unverified Configured API alone does not establish local inference
Benchmark source
MicrosoftComputer-use agentPublic projectRetrieves a previous human demonstration to help plan a similar Windows task. Uses screen information and application controls across apps.Windows agent with configured model backendOfficial source

Automating Windows desktops using AI with Microsoft UFO2 · BusinessBreeze

Demonstration-learning example · Microsoft

The creator video features UFO². The project example is the more specific reference for learning from a demonstration.
11Microsoft UFO³ Galaxy
  1. Observe everyday work – Unverified Device agents operate tasks; passive mining unproved
  2. Detect repetition – Unverified No recurring-action suggestion flow established
  3. Human approval – Unverified Delegation approval workflow unproved
  4. AI execution & verification – Partial Multi-device execution; Windows result checks unproved
  5. Reuse microtasks – Partial Task graph composition; mined suite growth unproved
  6. Run unattended – Unverified Coordination; full unattended exceptions unproved
  7. Fully self-hosted – Unverified API-configured agents; fully local cycle unproved
Benchmark source
MicrosoftMulti-device agent frameworkPublic repositoryCoordinates agents across devices and can use UFO² as its Windows agent. The single-desktop demonstration-learning capability above belongs to UFO².Agents deployed on participating devicesOfficial sourceUFO3: Weaving the Digital Agent Galaxy · Chaoyun ZhangProject-author presentation on coordinating agents across devices.
12ShowUI-Aloha
  1. Observe everyday work – Unverified Intentional human demonstration
  2. Detect repetition – Unverified No repetition detector documented
  3. Human approval – Unverified Human procedure-correction approval gate unproved
  4. AI execution & verification – Partial Task variants; independent result checks unproved
  5. Reuse microtasks – Partial Learned workflow reuse; suite test loop unproved
  6. Run unattended – Unverified Demo execution; recurring exceptions unproved
  7. Fully self-hosted – Unverified Documented API key; full local cycle unproved
Benchmark source
Show Lab and collaboratorsResearch prototypePublic repositoryCaptures screen, mouse and keyboard activity, interprets a demonstration and uses AI to perform variants. Open-source code with a model API key in the documented setup.Local Windows or macOS desktop with model APIOfficial source

ShowUI ALOHA Agent: Automate anything by video input · WHA Coder

Demo gallery · project authors

Creator walkthrough and researcher examples using Excel and PowerPoint.
13Tencent UI-Mate
  1. Observe everyday work – Unverified Native demonstration recording
  2. Detect repetition – Unverified No everyday task mining documented
  3. Human approval – Partial Human editing documented; approval gate unproved
  4. AI execution & verification – Partial Demo-guided actions; Windows independent result checks unproved
  5. Reuse microtasks – Partial Demo library; incremental composition testing unproved
  6. Run unattended – Unverified Benchmark rollouts; operational exception loop unproved
  7. Fully self-hosted – Unverified Open weights do not establish full locally hosted cycle
Benchmark source
Tencent and collaboratorsComputer-use modelPublic research and weightsOpen-weight computer-use model that accepts a demonstration as context. Researchers report improved performance on later tasks. The Windows benchmark and demonstration-assisted results are separate experiments.Model inference with a separate desktop-control environmentOfficial sourceUI-Mate demo gallery · Tencent researchersLook for DemoCUA for the demonstration-guided example. YouTube match unconfirmed.
14ServiceNow AI Desktop Actions
  1. Observe everyday work – Unverified Recorder mode; passive task discovery unproved
  2. Detect repetition – Unverified No automatic repetition proposals established
  3. Human approval – Partial Action authoring; trial approval gate unproved
  4. AI execution & verification – Partial Adaptive browser mode; changed native Windows case unproved
  5. Reuse microtasks – Partial Reusable desktop actions; mined suite growth unproved
  6. Run unattended – Unverified Execution documented; complete unattended exceptions unproved
  7. Fully self-hosted – No — described configuration ServiceNow-managed platform dependencies
Benchmark source
ServiceNowWorkflow automationDocumented platform featureA recorded sequence can run against native Windows apps. The documented adaptive mode chooses steps dynamically on Chrome websites.ServiceNow platform with desktop action runnerOfficial sourceRecorded desktop-action walkthrough · ServiceNowWorked recorder example. Evidence for browser adaptation should be assessed separately from native Windows replay. Exact YouTube match unconfirmed.
15OpenRPA
  1. Observe everyday work – Unverified Recorder and workflow builder
  2. Detect repetition – Unverified No automatic repetition detector established
  3. Human approval – Partial Workflow editing; approve-trial gate unproved
  4. AI execution & verification – Unverified AI adaptation and independent outcome checks unproved
  5. Reuse microtasks – Partial Workflow reuse; incremental mining integration unproved
  6. Run unattended – Unverified Full recurring exception process unproved in reviewed README
  7. Fully self-hosted – Unverified Open source client alone leaves full cycle unproved
Benchmark source
OpenRPA projectRPA platformPublic repositoryOpen-source Windows workflow builder for conventional recording and replay. The supplied evidence leaves AI adaptation to a changed case unverified.Windows desktop clientOfficial source

Recording a Windows application · Allan Zimmermann

Working within an application · Anders Jensen

Windows recording and application interaction provide a baseline for comparing AI-assisted replay.
16Nintex Kryon RPA
  1. Observe everyday work – Unverified Recording defined actions
  2. Detect repetition – Unverified No automatic repetitive-work discovery established
  3. Human approval – Partial Edited procedures; approval gate unproved
  4. AI execution & verification – Unverified AI adaptation plus independent result checks unproved
  5. Reuse microtasks – Partial Defined workflows; mined suite growth unproved
  6. Run unattended – Yes — provider documented VM robot executes schedules and triggers
  7. Fully self-hosted – Unverified Local robot/server does not establish whole AI cycle
Benchmark source
NintexRPA platformVersioned product documentationRecords and edits actions across applications. An unattended robot can execute the defined procedure on a virtual machine.Unattended robot on a VM connected to the RPA serverOfficial source

Upload data from Excel into an online portal · ranjit nair

Full-cycle automation demo · Naviant

Example crossing from a desktop spreadsheet into a web portal.
17SS&C Blue Prism Capture
  1. Observe everyday work – Unverified Process-owner demonstration capture
  2. Detect repetition – Unverified No everyday repetition mining documented
  3. Human approval – Partial Review and editing; trial approval unproved
  4. AI execution & verification – No — described configuration Capture exports documentation and skeletons; execution is external
  5. Reuse microtasks – Partial Process outline exports; incremental suite growth unproved
  6. Run unattended – Unverified Unattended execution requires a separate product
  7. Fully self-hosted – Unverified Capture installation does not supply complete cycle
Benchmark source
SS&C Blue PrismProcess captureDownload via Blue Prism PortalCaptures a demonstrated task and produces a process definition document, screenshots and a process map for review.Installed desktop applicationOfficial sourceBluePrism Capture 2.x demo · Benoît CaylaVersion-labelled demonstration of process capture and documentation. Subsequent execution would require separate evidence.
18Mimica Mapper and Miner
  1. Observe everyday work – Yes — provider documented Miner captures everyday desktop activity
  2. Detect repetition – Yes — provider documented Miner ranks opportunities; Mapper shows variants
  3. Human approval – Partial Mapper review; Maker approval is a separate product
  4. AI execution & verification – Unverified Maker claims execution; Mapper/Miner execution unproved
  5. Reuse microtasks – Unverified Maps/blueprints exported; reusable microtask suite unproved
  6. Run unattended – Unverified Maker agent runs differ from Mapper/Miner capabilities
  7. Fully self-hosted – Unverified Complete local AI hosting unproved
Benchmark source
MimicaTask miningVendor demoObserve everyday work, identify variations and produce procedure documents and bot blueprints. Particularly relevant to finding repetition without asking someone to stage a demonstration.Employee desktop activity capture; hosting terms unconfirmedOfficial source

Automated Workflow ID – Mimica Automation · Finnish Center for Artificial Intelligence

Merck and Mimica on task mining · Mimica

Workflow discovery and a named customer example. Coverage of both named modules in the footage remains unverified.
19XYZ by Decision Foundry
  1. Observe everyday work – Partial User-initiated recordings of real work
  2. Detect repetition – Partial Consultants analyse recordings; automatic suggestions unproved
  3. Human approval – Partial SOPs and analyst handoff; approval gate unproved
  4. AI execution & verification – Partial Deterministic guarded replay; AI adaptation unclear
  5. Reuse microtasks – Partial Consultants build workflows from captured knowledge
  6. Run unattended – Partial Completed repeat runs; full exception policy unproved
  7. Fully self-hosted – No — described configuration Customer cloud uses Anthropic models for analysis
Benchmark source
Decision FoundryClient solutionCase study; standalone availability unconfirmedA provider-described deployment connecting process capture with automation. The case study leaves standalone product availability and independent verification open.Client BPO deployment; hosting terms unconfirmedOfficial sourceIllustrated XYZ deployment · Decision FoundryProvider case study with product screens and reported runs. Exact YouTube match unconfirmed.
The ADA benchmark

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.

  1. Observe ordinary Windows work
  2. Find repetition and propose a microtask
  3. Let the user correct the procedure and approve a trial
  4. Use AI on a changed case and verify the result
  5. Reuse approved microtasks and test combined workflows
  6. Run unattended and report exceptions
  7. 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.