Orchestrator is green. The robots are running. The invoices are still in the queue.— The Reveille Bugle · Quick Answer
Quick answer: UiPath Document Understanding extraction fails for one of five reasons: the ML Skill or endpoint cannot be reached or licensed, a document exceeds a hard platform limit, more than 10 parallel requests hit one deployed project version, AI Units are exhausted so a redeployed or resumed skill will not start, or the model and taxonomy no longer match what is arriving. Start with the error text in the job log. If there is no error and work is simply not moving, start with the queue.
Why this one hurts
Four of the five causes produce no faulted job. The robot runs, the activity returns, the document lands in Validation Station or back in the queue, and the first person to notice is accounts payable.
01 — The Symptom
What you’re seeing
Any of these, usually after a change you did not connect to it.
- “Invalid response: content = ML Skill Unavailable code = unauthorized” from the ML Extractor or Extract Document Data activity.
- “Your license could not be validated. Please make sure that the API key parameter is correctly configured.”
- The same handful of documents fail every run while everything else processes.
- Failures cluster at the start of a large batch, then stop; retries succeed.
- Extraction stopped the day after a skill was updated, paused, or redeployed.
- Nothing fails, but every document routes to Validation Station and the human queue is days deep.
None of this is corruption and none of it is the workflow logic. Document Understanding has a set of published limits and a licensing model with sharp edges, and most extraction incidents are one of them. (Queue, job, robot and licensing health are among the things Reveille checks continuously on UiPath — the full test list is published here.)
02 — The Causes
Five causes, ranked by likelihood
Start at the top. The first two account for most tickets.
| # | Cause | How to confirm |
|---|---|---|
| 1 | The ML Skill or endpoint is unreachable or unlicensed. Skill not deployed or not public, wrong API key or endpoint in the activity, or the robot runs in a different Orchestrator folder from the skill. | The job log carries ML Skill Unavailable or license could not be validated. In AI Center the skill shows anything other than Available, or is not marked public. |
| 2 | The document exceeds a hard limit. Over 500 pages (100 at run time for Helix Extractor models, 20 for out-of-the-box models), over 160 MB, over 10,000 × 10,000 pixels, or a format other than PNG, JPG, PDF or TIFF. | The failures are the same documents every time. Check page count and file size on the failed items against UiPath’s known limitations. |
| 3 | Too many parallel requests. One deployed project version accepts a maximum of 10 parallel requests. A burst of robots on the same version gets refused. | Failures arrive together at the top of a batch and clear on retry. Count robots triggered against the same project version at the same minute. |
| 4 | AI Units are exhausted. An already deployed skill keeps serving as overconsumption, but a new deployment or training pipeline will not start — and pausing then resuming a skill counts as a new deployment. | Extraction died right after a skill pause, update or redeploy. Administration → Licenses shows the consumable at zero. |
| 5 | Model or taxonomy mismatch. A model trained on one template overfits (a known issue on 24.4 and 24.10), the classifier meets non-Latin text, a taxonomy field type changed, or the confidence threshold sends everything to a human. | Nothing faults. Confidence scores collapse, Validation Station fills, and the Document Understanding Monitor shows accuracy falling for one document type. |
03 — The Fix
Repair in this order
Each fix is complete on its own. Stop when extraction resumes.
1. Restore the skill and the key
In AI Center, open the skill and confirm it is deployed and public (Modify current deployment). In Studio, open the extractor activity and check the API key argument is a variable, not a string — a widely reported cause is the key wrapped in quotes so Studio sends the literal text. Confirm the Orchestrator folder selected in Studio is the folder the skill is published to. If you use an endpoint rather than a skill, paste the endpoint URL fresh from AI Center. Re-run one document.
2. Split or reject what the platform cannot take
Put a page-count and file-size check before extraction and route anything over the limit to an exception queue instead of the extractor. For long documents, split before the extractor so no single call exceeds 500 pages (100 for Helix models). Convert anything that is not PNG, JPG, PDF or TIFF before digitization. Out-of-the-box models stop at 20 pages; use a trained project for anything longer.
3. Throttle the burst
Cap the number of robots that can run against one deployed project version at ten, stagger the triggers, and add a retry with backoff around the extractor. A second deployed version of the same project doubles the ceiling if the volume is real rather than a scheduling spike.
4. Top up before you touch the skill
Check Administration → Licenses in Automation Cloud for the AI Units balance. If it is zero, do not pause, update or redeploy the skill until units are added — a deployed skill keeps serving, a redeployed one may not. UiPath’s own metering page sets out what consumes units; confirm overconsumption terms with your account team.
5. Retrain, or lower the bar deliberately
If one document type collapsed, retrain it with more than one template. If the classifier is failing on non-Latin documents, skip classification for that type and extract directly. If a taxonomy field type changed, revert it and re-run before investigating further. If Validation Station is full, the confidence threshold is the dial — lowering it is a business decision about accuracy, not a technical fix, so make it with the process owner in the room.
04 — Next Time
Why nothing warned you
Platform SLA is not workflow SLA. The platform was up the whole time.
Orchestrator reported healthy through all five. The services were running, the robots had sessions, the licenses were valid. What nobody was watching was the thing the business measures: whether documents left the queue extracted. That is a content-layer signal, and native dashboards do not raise it — the Document Understanding Monitor shows accuracy after the fact, and Orchestrator shows jobs, not outcomes.
The check a patch window or a skill redeploy needs is simple: are queue items aging past their SLA, are robots throwing errors, did the license and the access token still validate, and is the execution log clean. Run those before and after the change and cause 1, 3 and 4 show up in minutes.
Reveille for UiPath
Reveille runs over 65 out-of-the-box UiPath tests — including Check UiPath Queue Items, Check UiPath Robots Errors, Check UiPath Robots Run Time, Check UiPath License, Check UiPath Get Access Token, Check UiPath Health Status and Check UiPath Execution Log File Errors — with over 90 dashboard metrics and self-healing actions that restart UiPath services and web server application pools on a threshold breach. Reveille does not read the model; it watches the queue, the jobs, the robots and the licenses the model depends on, which is where every one of these failures surfaces. See Reveille for UiPath →
Supporting client estates as an MSP, SI, or ISV? That’s what Reveille SENTRY is built for — multi-tenant observability for the estates you manage.
Quick answers



