Biotechnology AI agent · Clinical operations

Clinical-Trial Support Agent (Study Conduct)

Track sites and subjects against the protocol schedule, find deviations and discrepancies, draft queries and keep the trial master file picture visible — for your medical monitor, CRA and investigator to act on.

Typical delivery: 4–6 weeks
Deployment: Your stack
Data access: Blinded by role
After launch: Agent Care

What this agent does

Prepares what the monitor and the CRA read

01
Take the protocol, its visit schedule and the plans that say what counts as an important deviation.
02
Read the study systems as copies — subject status, visits, forms, queries, monitoring records and the TMF index.

Reason

03
Compare each subject's visits, windows and required forms against the schedule the protocol sets.
04
Look for deviations, missing pages and values that disagree between forms or between two systems.
05
Draft queries in the site's own wording, and assemble the site picture a CRA takes into a visit.

Decide

06
Send a subject-safety signal or a possible serious breach to a named person at once, not to a queue.
07
Hold anything whose record copy, consent version or blinding status cannot be confirmed.

Out

08
Present the deviation list, the drafted queries, the TMF gaps and the visit pack with the record behind each.
09
Retain the reads, the drafts, the reviewer edits and the plan version the run was made under.

Product statement

The agent tracks, compares, drafts and routes inside the approval boundaries agreed during implementation. Eligibility, enrolment, randomisation, unblinding, clinical decisions and query closure stay with the investigator, the medical monitor and the sponsor's named roles.

Example workflow

One site review, end to end

Agent

Human

1Protocol and plans readThe protocol, its visit schedule, the monitoring plan and what the sponsor pre-specified as an important deviation
2Study data copied inSubject status, visits, forms, queries, monitoring records and the TMF index, read as copies with each read logged
3Schedule comparedEach subject's visits, windows, required forms and consent version set against what the protocol states
4Findings and drafts preparedDeviations, discrepancies and TMF gaps listed, queries drafted and the site-visit pack assembled

No human action required

Stages 1 to 4 run without a person in the loop — the medical monitor is not asked to read anything until the comparison and the record checks are done. A subject-safety signal leaves for a named person the moment it is seen.

5DecisionSplits on the record check and the confidence gate

Findings hold up

Goes to the monitor as a drafted review.

A record will not confirm

Held with the unconfirmed findings marked.

Medical monitor, CRA and investigator

Read the deviation list and the drafted queries against what could not be confirmed on a record copy.

Approve · Amend · Escalate

Approved — handed back

6Pack filed to the study systemsWritten as drafts only where access and policy allow; no query is answered, and none is closed
7Outcome evaluatedDeviations missed, queries the monitor rewrote, findings dropped as false and what the CRA found on site

Amendments

Findings the monitor rewrites or drops are counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

Assessing eligibility or enrolling a subject.
- Randomising, withdrawing or discontinuing a subject.
- Any clinical decision or instruction about a subject.
- Unblinding, or releasing treatment-revealing data.
Automation boundary
Agent acts unaided
- ✓Compare visits and forms to the protocol schedule.
- ✓List candidate deviations, discrepancies and TMF gaps.
- ✓Draft queries and assemble the site-visit pack from record copies.
- ✓Check consent version and blinding status, and hold what will not confirm.
Write actions run only inside the approval boundaries agreed during implementation. Closing a query is not one of them.
- Altering source data or a source document.
- Deciding that a deviation is not reportable.
- Answering, signing or closing a query.
Changing the protocol, the plans or the thresholds.

Example output

One flagged visit, annotated

Everything the agent flags is attached to the record copy and the protocol rule it came from.

Deviation review · single subject visit

Illustrative example

Site: Investigator site, blinded view
Subject: Randomised, arm not shown
Visit: V4, ±3 days
Proposed: Visit outside its window
Confidence: 91%
Reportability: Left for the sponsor

As received

The visit and the schedule as the study systems hold them, read as a copy — nothing on this side is edited.

Evidence used

Protocol visit schedule
Visit date on the copy
Consent version in force

Why it is flagged

The visit falls outside the window the protocol sets, and the plan calls that important.

What the score decides

Below the configured threshold the finding is held for the monitor instead of the pack.

Action

ApproveAmendEscalate

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every subject visit and form

From your study systems

03Comparison & drafting

Hold the study to its own schedule

Use the visit schedule, the windows, the monitoring plan and what the sponsor pre-specified as important.

01Approved path

Cut the reading before the judgement

The deviation list, the discrepancies and the drafted queries arrive together, so a medical monitor spends the time on the calls only they can make.

02Human review

Send the CRA to the sites that moved

Out-of-window visits, ageing queries, consent-version gaps and missing trial master file records are named per site, so a visit is planned on what changed.

Build an evidence trail

Retain the copies read, the protocol rule behind each finding, the drafted query, the confidence, who was told of a safety signal and every reviewer edit — on both paths.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

EDC & clinical data
- Medidata Rave
- Veeva CDMS
- Oracle Clinical One
- data APIs
CTMS & site management
- Veeva CTMS
- Medidata CTMS
- site status
- Visit reports
- issue logs
eTMF & documents
- Veeva Vault eTMF
- Montrium
- SharePoint
- TMF Reference Model index

Agent

Clinical-trial support

Reads study copies
Compares to protocol
Routes to the monitor

Randomisation & safety
- IRT / RTSM blinded views
- safety database
- Deviation and breach logs
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the study record

Each control wraps the one inside it. A finding clears every layer before it reaches the pack a medical monitor reads.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeRestrict automation if evaluations or production signals degrade.Roll back

L5Trail and versionReads, findings, reviewer edits and model, prompt and rule changes are recorded under change control.Record

L4Monitor and PI gateDefine what may be filed as a draft; answering or closing a query waits for the site and the monitor.Gate

L3Safety routingWording on the sponsor's safety and breach lists pages a named person and records when the clock started.Escalate

L2Blinding controlsFields set as treatment-revealing are withheld by role and a read outside it refused; a field nobody set as revealing is not caught here.Withhold

L1Record bindingFindings are checked against the record copy and the protocol rule they name; unbound ones are held.Verify

Model coreFinding proposed — the deviation or discrepancy, the record copy behind it, a drafted query and confidence

L1 – L2: Decide if a finding stands and who sees it
L3: Routes safety wording to a person, records the clock
L4 – L5: Leave the decision to the site, keep the trail
L6: Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the whole review — not only the deviation list at the end.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the review pack the monitor opens

E1Final-output evaluationWas the finding real, and the deviation the protocol actually defines?
E2Comparison evaluationWere visits, windows and forms read against the current protocol version?
E3Blinding evaluationDid a treatment-revealing field reach a role that should not see it?
E4Query-draft evaluationDid the query name its record and ask one answerable thing?
E5Slice evaluationHow does performance change across specific site and subject cohorts?
E6Business outcomeHow much of the pack was rewritten, and how many gaps were already filed?

Floor — the study conduct the sponsor has to defend

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle.

Agent lifecycle

Direction of processing →

01 · Intake / retrieval
2 modes
- TR-01
  Compared against a superseded protocol
  An amendment moved the window; the old schedule still runs.
- TR-02
  Consent version not read at the visit
  A visit is reviewed against a form the subject never signed.
Stage gathers
The protocol, its schedule and the study systems as copies
02 · Comparison
2 modes
- TR-03
  Deviation left off the list
  A visit outside its window reads as compliant and is never raised.
- TR-04
  Query raised against data that is right
  The site is asked to explain a value its own source copy supports.
Stage compares
Subject visits, windows and forms against the schedule
03 · Drafting
1 mode
- TR-05
  Treatment-revealing value in a query
  A drafted query quotes a field that shows the arm.
Stage drafts
Queries, the deviation list and the site-visit pack
04 · Pack / output
1 mode
- TR-06
  Safety signal handled as a data issue
  An event in the data is read as a discrepancy, not a signal.
Stage returns
What the monitor, the CRA and the investigator read
05 · Change / Version
1 mode
- TR-07
  Silent rule change moves the findings
  A check or threshold changes with no record of who approved it.
Stage tracks
Amendments, model, prompt and rule changes

Sev-1 · the study or the blind is put at risk
Sev-2 · the pack misstates what happened
Sev-3 · findings degrade, sites carry the cost

See the complete Biotechnology failure-mode catalogue

Affected slices

A site three weeks old is not a site three years old

Sites are not equivalent and this number says so. It counts findings the monitor dropped or rewrote before the pack went out, highest where staff have not run the schedule yet or where source is on paper.

Slice performance — reported separately, not only in aggregate

Illustrative example

Slice	Failure rate	Lift	Status
Newly activated sites	5.9%	3.4×	Review
Paper-source and scanned visits	4.1%	2.3×	Review
Subjects under an amended protocol	3.0%	1.7×	Watch
Long-running stable sites	1.3%	0.7×	Normal

Bar: dropped-finding lift vs. stable-site baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

One dropped finding changes the check that produced it

A finding the monitor threw out is not fixed by throwing it out. The check behind it moves under version control, and the next pack is read against that.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

01
Detect
Dropped findings or missed deviations rise at one kind of site.
02
Diagnose
Either the rule was wrong or the copy it read was, and both happen.
03
Improve
The check or the schedule map is changed under change control, approver named.
04
Verify
The affected subjects and sites are compared again from the copies.
05
Learn
The dropped finding stays in the test set, with what the monitor wrote instead.
Learn → Detect
The return edge. A change to a check or a threshold in a study system is itself a change — approved, recorded and version-linked before the next comparison runs.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, access and copies, comparison and drafting, evaluation, integration, then production validation and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

01Workflow discovery and boundary definition.Week 1
02Study-system access and blinded-role design.Week 1, Week 2
03Protocol, schedule and deviation-plan mapping.Week 1, Week 2
04Record copying, read logging and source rules.Week 2
05Visit, window and required-form comparison.Week 2, Week 3
06Deviation and discrepancy detection logic.Week 3
07Query drafting in the site's own wording.Week 3, Week 4
08Trial master file gap and index checks.Week 3, Week 4
09Safety-signal and serious-breach escalation routing.Week 4
10Monitor, CRA and investigator review workflow.Week 4, Week 5
11Evaluation suite and regression subjects.Week 4, Week 5
12Integration, deployment and Agent Care handover.Week 5, Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier	PilotOne study, one site group	ProductionProduction integration	AdvancedSeveral studies / programmes
Introduced at Pilot
Schedule comparison and deviation findings	✓	✓	✓
Trial master file gap list	✓	✓	✓
Blinded-role access and source copies	✓	✓	✓
Consent-version and status checks	✓	✓	✓
Safety and serious-breach routing	✓	✓	✓
Query drafting and review pack	✓	✓	✓
Baseline evaluation	✓	✓	✓
Introduced at Production
Study-system and eTMF integration	—	✓	✓
Observability and evaluation	—	✓	✓
Introduced at Advanced
Several studies, sites and languages	—	—	✓
Multi-programme and enterprise controls	—	—	✓
Build price	From $5,000	From $8,000	Custom quote

Final build price: Confirmed after discovery based on the studies and systems in scope, site and country coverage, subject volume, monitoring and review controls and deployment requirements.
Separate from build: Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bring

We build with it

01The protocol, its visit schedule and the deviation plan
→
Protocol, schedule and deviation-plan mapping
Week 1
02Who may see unblinded data, and who must not
→
Study-system access and blinded-role design
Week 1
03Read access to the EDC, CTMS and eTMF you run
→
Record copying, read logging and source rules
Week 2
04A monitoring visit report your CRA has already written
→
Visit, window and required-form comparison
Week 3
05Queries your data managers wrote — the good and the wasted
→
Query drafting in the site's own wording
Week 3
06Findings that went wrong — a missed deviation, a bad query
→
Evaluation suite, regression subjects and failure-mode testing
Week 4
07Named medical monitor, CRAs and investigator contacts
→
Review workflow, then pilot sites and production validation
Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases are drawn over the weeks they actually occupy. Week 5 puts the first drafted findings in front of a monitor while the cohort slices are still running.

Phase

DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6

Week focus

W1Protocol, schedule, deviation plan and the blinded-role map

W2Study-system access, record copies and read logging

W3Visit and window comparison, deviation and discrepancy logic

W4Query drafting, safety-routing tests and site cohort slices

W5Study-system integration, first monitor-reviewed packs and edits

W6Production validation, live sites with your monitor and Agent Care handover

Reading the band: A bar covers the weeks its work is named in, and nothing either side. Roles and copies are settled in week 1; a study system read under the wrong role cannot be unread.
At the end of W6: The monitor and the CRA decide whether the packs saved them a day. Agent Care picks up the checks, the role map and the drift in both.
Duration: Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Biotechnology AI agent

Build a trial-support agent around the study you are already running.

Show us one protocol, the systems your study data sits in and how a deviation reaches your medical monitor today. We do not read a study system until your roles are mapped onto its copies. That map is week one, and it is the part clients most often find they do not have.

Typical delivery: 4–6 weeks
Deployment: Your clinical stack
Data access: Blinded by role, on copies
After launch: Agent Care available

Nestack Agents · Clinical-trial support

AGT-BT-03 · Agent Care available after launch

Clinical-Trial Support Agent (Study Conduct)

Hold the study to its own schedule

Cut the reading before the judgement

Send the CRA to the sites that moved

EDC & clinical data

CTMS & site management

eTMF & documents

Clinical-trial support

Randomisation & safety

Observability & evaluation

Compared against a superseded protocol

Consent version not read at the visit

Deviation left off the list

Query raised against data that is right

Treatment-revealing value in a query

Safety signal handled as a data issue

Silent rule change moves the findings

Detect

Diagnose

Improve

Verify

Learn

Build a trial-support agent around the study you are already running.