Human fallback queue1 item from a failed check. Owner: Wendell Haas.
DriftNo rule has fallen back to suggestion only since go-live, Sep 14. Owner: Naveen Prakash.
Checks that stopped or waited this cycle
Checks that stopped or waited this cycle, Business review, US field
When
Result
Run and step
Action
Check
Source read
Nov 17, 06:00
FailedStopped. Sent to the human fallback queue.
BRV-0418Track actions to closure
Update the F1 call countReview follow-up agent
Source freshness
Veeva CRM day close for Mon Nov 16: not loaded
Oct 20, 11:52
WaitedReleased Oct 26, 06:00 ET.
TAL-0153Cover SE-07, Oct 26 to Nov 15
4 coverage writesTerritory coverage agent
Process calendar
Q3 incentive compensation close, Oct 12 to Oct 23
Checks
1 of 42 failed
1 wait: 4 writes queued by the Q3 IC close
Overrides
0 of 27
Cards overridden since Sep 14. Evidence opened before approval on 26 of 27.
Question to answer
25 s
Halvane median before go-live: 19 days
Question to first corrective action under way
3 h 58 min
Median before go-live: 31 days
The morning view for one process.
Trusted by 2 out of the 5 largest life sciences companies
One screen for each process.
Every run, queue and SLA for the process sits on one screen. The process owner opens it each morning and acts on what is late or blocked.
See which check stopped a run, and why.
Each failed check shows the action, the source it read and what it found. Nothing is written, and the item moves to a person’s queue.
Pause or stop any agent from the same screen.
Every agent has a version and a kill switch. Pause, quarantine, rollback and incident runbooks sit beside the work, and a circuit breaker pauses an agent when errors pass a threshold.
Rules fall back when the world changes.
A rule drops to suggestion only when something it depends on changes, live performance slips below its band or overrides cluster around a new reason.
Measure quality, never people.
Override rates sit beside quality signals and a weekly random audit of the real miss rate. Everything is reported per process, never per person.
Value against your own baseline.
Cycle time is compared with a baseline measured before go-live. A review question that took 19 days now takes 25 seconds, and the report notes what else may explain it.
“What won over our compliance team wasn’t the speed. It was that the agent shows its work, stops at Draft, and tells you when it doesn’t know.”
Head of Marketing OperationsOne of the top five US biopharma companies
Who watches and who acts
Oversight shows the work. Named people decide what to do about it.
Oversight does
Shows each run, queue, failed check, override, drift, cost and cycle time for one process.
Moves an agent’s items to a human fallback queue when the agent or a source is down.
Shows Govern’s quality signals beside the override rate, for the process, never per person.
Prints its caveat when value is measured without a concurrent comparison.
People decide
The process owner opens the screen each morning and acts on what is late or blocked.
Admins set identities, roles and access reviews in Govern’s admin console, which Oversight links to.
A weekly random human audit, stratified by risk tier, measures the real miss rate with its interval.
The customer’s QA and compliance teams judge the evidence and sign it.
What Oversight does not do
Commercial BI.Prescriptions, share and call attainment stay in the customer’s BI tools. Oversight reports on agents and processes only.
Certify compliance or validated state.The customer’s QA and compliance teams sign the evidence.
Rate individuals.Quality signals are reported at process level only. Ontelis does not score employees or publish leaderboards.
Ontelis Oversight is where you monitor AI agents once they're in production, with one screen for each business process. It shows what's running and at which step, what's waiting and on whom, and which actions a check stopped and why. The process owner opens it each morning and acts on what's late or blocked, with pause and the kill switch on the same screen. Evaluations tests an agent version before it may run. Oversight shows whether it keeps working once it does.
Observability tools show engineers what an agent's code did: calls, tokens, latency and errors. Ontelis Oversight shows the process owner what happened to the work. That means which territory coverage request is waiting and on whom, which write a check stopped and what the source said, and how long each run takes against your baseline from before go-live. Oversight reports on agents and processes only. Prescriptions, share and call attainment stay in your BI tools.
Your admins decide who can pause or stop an Ontelis agent: they set those roles in Govern's admin console, which Oversight links to. Every agent has a version and a kill switch. Pause holds agents at their last checkpoint. The kill switch stops one agent, and its items move to a person's queue. Quarantine holds only the affected item while the rest of the run continues. A circuit breaker also pauses an agent automatically when its error rate passes a set threshold.
The work goes to people instead of going ahead. When an Ontelis agent or a source system is down, the agent's items move to a human fallback queue. A check that can't read fresh data, such as a Veeva CRM day close that hasn't loaded, stops the write, and Oversight shows which check failed, which source it read and what it found. If Ontelis itself is down, agents pause at their last checkpoint and your teams keep working in Veeva and SAP.
A weekly human audit tells you. Each week people check a random sample of the agents' work, drawn across risk tiers, and Ontelis Oversight reports the real miss rate beside the override rate. A low override rate alone isn't taken as proof, because it can also mean approvals without a real look. Drift is handled rule by rule: when something a rule depends on changes, such as a territory realignment, or its live results slip, the rule drops back to a suggestion and people decide each case.
No. Oversight reports quality for each process, never for a person. The signals behind it, such as time spent on a card and whether the evidence was opened, are worked out per item and aren't stored or added up by person. Ontelis doesn't score employees or publish leaderboards.
Oversight measures time saved against your own baseline, captured before agents go live: cycle time, human minutes per unit, and error and rework rates. Where it can, Ontelis also compares the same process running without agents in another brand or region over the same period. Where agents already run everywhere, it compares the trend before go-live with the trend after, and the report states in print that other changes in the period may explain part of the difference.
Oversight reports on agents and processes only; prescriptions, share and call attainment stay in the customer’s BI tools.
Vendor documents
Ontelis provides: a SOC 2 Type II report, an ISO 27001 certificate, a HIPAA business associate agreement wherever Ontelis processes PHI, a data processing agreement with transfer clauses, and a sub-processor list.
Models
Ontelis builds no foundation model. No model is trained on customer data, except an opt-in adapter that never leaves the customer’s data plane.
Your controls
Ontelis’s posture as a vendor is kept separate from Govern, which holds your own controls.
Know what every agent is doing, and what is waiting