AI agents and workflow automation for MedTech, built for the processes where volume and regulatory consequence occur together.
The repetitive, high-volume, genuinely automatable work in a device company is complaint intake and triage, nonconformance handling, CAPA administration, service coordination and field action logistics. Every one of those produces a record inside a quality system, under a defined procedure, with a timeline somebody is accountable for. The automation opportunity is real and the deployment is not a support chatbot, because the output lands somewhere that carries obligations.
Reliable automation is defined by what happens when something goes wrong. And a complaint the agent classified as non-reportable produces no error anywhere, which is exactly the case that needs designing for.
A quality engineer processing complaint intake performs a similar task hundreds of times: read the description, extract what happened, classify the event, decide whether it meets criteria for escalation, and route it. The repetition makes it an obvious automation candidate. The classification decision inside it is a regulated judgement with a timeline attached, and an incorrect non-escalation generates no alert, no exception and no downstream signal. The easy part and the consequential part are the same step.
The second problem is that these processes are described in controlled procedures, and an agent participating in them is participating in a documented process somebody validated.
The output is a quality record
A complaint, a nonconformance or a CAPA entry carries retention, traceability and review obligations from the moment it exists.
A missed escalation produces no signal
An event classified downward disappears, and nothing in the system indicates that a decision was made incorrectly.
The process is in a controlled procedure
An agent participating in it is participating in something documented, approved and subject to change control.
Timelines are counted from receipt
Which means an agent that defers everything to review has not helped, and one that over-escalates consumes the capacity it saved
The agent can select the wrong record
Complaint, product, supplier, batch or service record, and an action executed against the wrong one is both wrong and recorded as correct.
Workflows cross QMS, ERP, PLM, CRM and service systems
Each handoff is a point where information is re-entered, state diverges and nobody owns the end-to-end outcome.
Tune complaint handling toward recall and accept the review cost
In complaint and nonconformance triage, a false escalation costs a quality engineer twenty minutes. A missed escalation is a regulatory event that may surface months later through an entirely different route. Those are not symmetric and the threshold should not be set as though they were. An agent that over-escalates by a measurable margin, flags everything it was uncertain about, and never silently classifies downward is the design that can be deployed, and the review burden it creates is the price of the thing being safe.
The reliable division in quality workflows is an agent that reads, extracts, structures, cross-references and prepares, with the classification and escalation decisions remaining with the qualified person. That removes most of the elapsed time without moving any of the judgement, and it is the version that survives a quality review.
Step 1
Map the controlled procedure first, since the agent is participating in a documented process rather than an informal one.
Step 2
Separate intake and preparation from determination, because one is automatable and the other carries the obligation.
Step 3
Establish where the output lands, since a quality record and a tracking spreadsheet have entirely different consequences.
Step 4
Tune toward recall wherever a miss is the serious failure, and accept the resulting review load deliberately.
Step 5
Design the review interface as the product, since that is where the time is spent and where the control operates.
Step 6
Classify every action by risk rather than by the agent confidence, separating administrative automation, record preparation, quality-impacting updates and safety or regulatory determinations.
Step 7
Define what the agent must escalate, including everything it was uncertain about.
Step 8
Agree the validation position with your quality function before building, because it shapes the architecture.
Step 9
Give the workflow explicit state, so retries do not duplicate records or restart a timeline.
Step 10
Record the complete interaction, since reproduction is unavailable and the record is the only account.
An agent inside the quality system is inside your validated process.
Complaint handling, nonconformance and CAPA run under procedures that were written, approved and are subject to change control. Inserting an automated component into one of them is a change to that process, not an IT deployment alongside it, and the question of what qualification it requires belongs to your quality function before the first prototype rather than after the pilot succeeded. Teams that build first and ask afterwards frequently discover the architecture cannot produce the evidence the process needs.
General agent architecture and orchestration are covered on our agent and copilot pages. These are the capabilities specific to workflows producing quality records and touching devices in the field.
Read and Prepare
Unstructured Intake Processing
Complaint descriptions, service notes and field reports arriving in free form, several languages and variable completeness, extracted into a consistent structure.
Classification Support
Proposed categorization with the evidence behind it, presented for determination rather than applied, where the classification carries regulatory consequence.
Duplicate and Linkage Detection
Identifying that a new report concerns the same device, batch, event or existing record, which is where volume quietly inflates.
Context Assembly
Device history, firmware, service record, prior complaints and manufacturing lot gathered automatically, which is most of the elapsed investigation time.
Route and Act
Workflow Orchestration
State, routing, escalation and exception handling across quality, service and commercial systems with explicit transitions.
Enterprise Integration
Quality management, service, ERP, CRM and device platform systems, with the permissions and controls each requires.
Device-Affecting Action Control
Where an agent can trigger a dispatch, configuration change or field action, the authorization and confirmation covered on our remote management page.
Timeline and Obligation Tracking
Elapsed time against procedure timelines surfaced as work, since a process clock running unnoticed is the operational risk.
Govern and Evidence
Human Review Engineering
The determination interface built as the product, with evidence proximity, easy disagreement and captured reasoning.
Validation Support
Qualification evidence, versioned configuration and change control to the position your quality function defines.
Complete Interaction Recording
Input, retrieved context, versions, proposed output and every human decision, retained because reproduction is unavailable.
Evaluation and Monitoring
Recall on escalation-worthy events, over-escalation rate, review time, correction patterns and whether review remains genuine.
What CaliberFocus does, and does not do?
The objective is not maximum autonomy. It is reliable automation with the right level of control, and an agent should know what it may do, what it may not, and when to stop. We build agents that prepare determinations rather than make them, because in a quality process the determination carries an obligation a system cannot hold. We tune toward recall where a miss is the serious failure and we state the review cost that creates rather than hiding it in an efficiency figure. Classification criteria, reportability determinations, validation strategy and what the procedure permits remain with your quality and regulatory functions.
These differ mainly in whether the output becomes a quality record and whether a miss generates a signal. The third column determines the autonomy granted and the review required, and it should be settled before anything is built.
| Workflow | What the Agent Does | Where the Output Lands |
|---|---|---|
| Customer Support Triage | Classifies and routes inbound contacts | A support queue. Low consequence and the right first deployment. |
| Service Request Handling | Prepares dispatch, parts and scheduling | Service system. Operational, with cost rather than clinical consequence. |
| Complaint Intake and Extraction | Structures the report and assembles device context | A quality record, with a procedure timeline starting at receipt. |
| Complaint Classification Support | Proposes categorization and escalation | A regulated determination. Prepared by the agent, decided by a person. |
| Nonconformance Processing | Extracts, categorizes and links related records | A quality record feeding trend analysis and potentially CAPA. |
| CAPA Administration | Tracks actions, deadlines, evidence and closure readiness | A quality record where the administrative burden is genuinely large. |
| Field Action Logistics | Coordinates notification, tracking and effectiveness checks | A regulated action affecting devices in use. Highest consequence here. |
CAPA Administration Is the Largest Unglamorous Win Available
CAPA generates a substantial, continuous administrative load: tracking actions against owners, chasing evidence, monitoring deadlines, assembling closure packages and flagging overdue items. Almost none of that involves a regulated determination.Â
Human oversight appears in every agent design and is implemented at very different strengths. The third column is whether the model functions as a control in practice, which depends on whether the reviewer can realistically do what the design assumes.
| Model | How It Works | Whether It Is a Control |
|---|---|---|
| Agent Prepares, Person Determines | The agent assembles and the qualified person decides | Yes. The determination never left the accountable person. |
| Person Verifies Against Evidence | Reviewer checks the proposal against presented source material | Yes, where evidence is presented. No, where it must be located. |
| Person Approves with Full Context | Reviewer sees inputs and reasoning before deciding | Yes, if time allows and disagreeing is as easy as agreeing. |
| Person Approves the Summary | Reviewer sees the proposal but not what produced it | Weak. They are assessing plausibility rather than correctness. |
| Sampled Review | A proportion checked, the remainder passes through | Only with a measured error rate that justifies it, and monitoring. |
| Intervention Available | Nobody reviews unless something is noticed | No. This is autonomy with a stop button and should be named that. |
Give the workflow explicit state
Received, extracted, proposed, awaiting determination, determined, routed, closed, with exception paths and retry safety.
Never let a retry restart a timeline
A procedure clock counted from receipt must not reset because a job ran twice, which is easy to introduce and hard to notice.
Present evidence beside the proposal
Review speed is governed by evidence proximity, and a reviewer who has to open the source is doing the original job.
Make disagreeing as easy as agreeing
If accepting is one click and overriding takes five minutes, the interface has determined the outcome.
Scope agent permissions to the task
Not inherited from the invoking user, since an agent spanning quality, service and device systems otherwise aggregates authority.
Record the determination and its basis
Who decided, on what evidence, under what authority, since that is what a later review will ask for.
Watch the override rate, because a decline toward zero is ambiguous
change a meaningful proportion of proposed classifications, which is the control working. As confidence grows the rate falls, and some of that is genuine improvement while some is reviewers accepting proposals they would previously have questioned. Those two causes are indistinguishable in the data and have opposite implications. Periodically re-reviewing accepted determinations independently is the only way to tell, and in a quality process the difference matters more than it does almost anywhere else.
Quality processes in a device company run under a documented management system. Inserting an automated component into complaint handling, nonconformance or CAPA changes how a controlled procedure operates, which makes it a process change subject to whatever assessment, qualification and documentation your quality system requires. That position determines the architecture, and it should be established before the prototype rather than discovered during an audit.
Process and validation
- Treated as a change to a controlled procedure rather than as a parallel tool, with the assessment your quality system requires
- Engineering controls capable of supporting the frameworks that apply to your organization, which may include FDA regulation, ISO 13485, GxP expectations and electronic record requirements
- Applicability, interpretation and validation strategy determined by your quality and regulatory functions rather than by us
- Versioned configuration across model, prompt, rules and integrations, since each changes behaviour and only some are releases
Records
- Complete interaction retained: input, assembled context, proposal, evidence, human determination and basis
- Attribution to the person who determined, since the obligation rests with them rather than the system
- Reproduction unavailable for generative components, which makes the record the only account of what happened
- Retention matched to the obligation on the quality record the output became part of
Security
- Retrieved and submitted content treated as data rather than instruction, since complaint text arrives from outside
- Agent permissions scoped to the task rather than inherited, particularly where device-affecting actions are reachable
- Patient and customer information in complaint content handled to the obligations your counsel defines
- Override and escalation patterns retained and analysed, since they are the first indicator of a criteria or extraction problem
Oversight
- Determination authority explicitly retained by qualified personnel for anything with regulatory consequence
- Escalation path where the agent is uncertain or the reviewer disagrees, routing rather than ending
- Override rate monitored with periodic independent re-review of accepted determinations
- A named owner accountable for whether the agent is still performing, distinct from whoever built it
Decide what the agent may never do, and enforce it in code.
A prompt instruction not to classify a complaint as non-reportable is a request. A rule in the application that prevents that classification from being applied without a qualified determination is a control. The distinction matters because the agent will encounter cases its instructions did not anticipate, and the boundary that holds under those conditions is the one implemented in software. Writing the list of prohibited actions before the capability list is the cheapest safety work available and the most frequently skipped.
Agent evaluation defaults to accuracy across all cases, which in a quality workflow is dominated by routine events and tells you almost nothing about the failure that matters. The number worth knowing is recall on the subset that should have escalated, measured against a reference, and it requires somebody to establish what the right answer was rather than reading the output and finding it plausible.
Recall on escalation-worthy cases
Measured against an independently determined reference, since this is the failure that generates no signal.
Over-escalation rate
The cost of the recall tuning, which should be known and accepted rather than discovered as review load.
Extraction completeness
What the agent failed to pull from the source, since an incomplete record looks complete.
Duplicate and linkage accuracy
Both directions, since missed links inflate volume and false links merge distinct events.
Review time and correction patterns
Which indicate whether the problem is extraction, classification or presentation.
Performance after any version change
Model, prompt, rules or integrations, any of which moves behaviour without a visible release.
Build the evaluation set from real cases
Including the difficult ones, the multilingual ones, the incomplete ones and the ones that were escalated late.
Keep a failure library
Every meaningful miss added to the set so a future version proves it did not reintroduce the problem.
Re-review accepted determinations periodically
Independently, since this is the only way to distinguish improving proposals from declining scrutiny.
Measure the whole workflow
From receipt to determination, because intake time is the easy half and the timeline is counted from receipt.
Automation in quality workflows is usually reported as cases processed and time saved at intake. The outcomes that matter are whether the full process got faster, whether recall held, and whether the determinations are still genuinely being made.
| Category | What We Measure | Why It Matters |
|---|---|---|
| Controlled Workflow Completion | Workflows completed correctly with every required approval, permission and record intact, rather than actions automated | An agent completing fewer tasks while handling exceptions correctly is worth more than one executing broadly with gaps. |
| Recall on Escalation-Worthy Cases | Events that should have escalated and did, against a reference | The failure that generates no signal, and the reason the deployment is defensible. |
| End-to-End Elapsed Time | Receipt to determination including review, not intake alone | The procedure timeline runs from receipt, so this is the measure that counts. |
| Over-Escalation Cost | Additional review load created by recall tuning | A known, accepted cost rather than an unpleasant discovery. |
| Override Trajectory | Correction rate over time with periodic independent re-review | Distinguishes improving proposals from declining scrutiny. |
| Context Assembly Time | Time to gather device history, service record and prior events | Where most of the investigation elapsed time actually sits. |
Honest Expectation Setting
An assessment may reveal fragmented workflows across systems, excessive handoffs, duplicate data entry, unclear process ownership, missing exception handling, poor integration between QMS, ERP, PLM and CRM, inconsistent business rules, incomplete audit trails, pilots without action-level evaluation, automation without defined approval boundaries, or processes that need redesign before automation.
We will map the controlled procedure, separate preparation from determination, establish where output lands and what that requires, measure recall on escalation-worthy cases against a reference, and agree the validation position with your quality function before anything is built.
Bring us a specialty or clinical setting where clinicians are spending too much time creating notes. We will assess where ambient documentation fits, what must remain clinician controlled, how it should integrate with your EHR, and how to measure whether it is actually reducing burden.
One conversation with people who have run these deployments, and a written readiness view you can use with or without us.
- AI Agents and Workflow Automation
- Voice and Conversational AI
- Document AI and Intelligent Processing
- Generative AI and Enterprise Copilots
- AI Strategy and Governance
- HCC and Risk Adjustment Analytics
Security & Compliance
