Symphony

Overview

Metadata about this workflow.

Name: symphony-wc
Line of business: workers_comp
Version: v1
Status: published
Created: Aug 6, 2026, 3:00 PM
Published: Aug 7, 2026, 1:12 PM
Archived: —
Author: —

Driver brief

How the driver thinks. Role, operating principles, tone, and domain knowledge — combined with the active capabilities below into the brief the agent reads before every action.

Active capabilities · added to the brief

Parse the intake packetCompute AWW / TTD / PD rate correctlyRun three-point contactManage statutory clocksAssemble compensability recommendation

Capabilities

Named routines that pair one or more tools with the instructions for using them. Enabling a capability adds those instructions to the driver's brief.

5 of 5 enabled

Parse the intake packet
active
Classify each attachment, extract the facts, run the completeness check, and surface the two documented gaps (wage-basis and field-26 restrictions).
toolsread_submission_stateclassify_documentupdate_captured_fieldset_claim_field
Compute AWW / TTD / PD rate correctly
active
Run the LC §4453 arithmetic on the certified wage statement, surface the gap against the rate×schedule figure, and persist the recommended AWW.
toolscompute_awwset_claim_field
Run three-point contact
active
Coordinate employee (voice→voicemail→SMS→scheduled callback), employer (email + reply), and provider (records request) with a confirm-not-interrogate opening on each.
toolsinitiate_outbound_callschedule_outbound_callsend_smssend_emailschedule_followupset_claim_field
Manage statutory clocks
active
Post the §5402(c) authorization on DWC-1 receipt, diary the §5402(b) 90-day date, and keep the §4650 TD trigger diaried to onset of disability rather than DOI.
toolsset_claim_fieldschedule_followup
Assemble compensability recommendation
active
When the last checklist item closes, build the 10-line basis and stage the compensability decision for the adjuster.
toolsset_claim_fieldsend_templated_emailescalate_to_human

Enabled Tools

The actions the agent is allowed to take. External tools touch the outside world via an integration — pick which adapter and endpoint each tool calls; the env-specific connection is resolved at runtime.

18 / 25 enabled

send_smsexternal
Send an outbound SMS to a party on this submission/claim. Use for quick asks (single field, status updates). Avoid for anything requiring an attachment. Always include a rationale explaining why this is the next best step.
used byRun three-point contact
send_emailexternal
Send an outbound email to a party. Use for asks that need detail, document requests, or anything requiring an attachment trail. Threading is automatic when in_reply_to is set to a prior outbound email message_id.
used byRun three-point contact
send_templated_emailexternal
Render a hardcoded letter template to PDF and email it to a party (or to whichever party on the claim holds a given role). Generates a Document row and threads into the comms hub.
used byAssemble compensability recommendation
send_upload_linkexternal
Text a party a private link to upload photos, videos, or documents (receipts, estimates) for a submission or claim. Use after a call when the caller indicated they have evidence to share and their preferred follow-up channel is sms/text (or unset). Skip if their preference is 'email' — the inbound-email parser handles attachments. Pass exactly one of submission_id or claim_id; the link is scoped to that record.
initiate_outbound_callexternal
Place an outbound phone call to a party via Retell. Use when: (a) the party's preferred_channel is 'voice' and the next outreach is conversational rather than a document ask, (b) ≥3 required fields are still missing or the loss narrative is empty/<1 sentence (a call collects ambiguous info faster than SMS), or (c) the most recent prior call's transcript shows it didn't reach its objective (caller hung up, agent didn't get past identity, line dropped). Default to SMS for short asks and email for documents. Provide a focused objective, a short ordered list of talking points, and a voicemail summary in case the recipient doesn't pick up. Always include a rationale.
used byRun three-point contact
schedule_outbound_call
Schedule a future outbound phone call. Use when you've decided a call is right but timing matters — e.g. callback windows the recipient gave you, or backing off a few hours after a missed call attempt. The adjuster can cancel or run-now from the UI. Use initiate_outbound_call when you want to call right now.
used byRun three-point contact
update_captured_fieldautopilot-safe
Merge a single field into submissions.captured_fields. Use when you derive a value from a message (e.g., reporter texts back "Oct 4 morning" → set loss_date). Key must match an existing FieldDefinition. Triggers a re-evaluation of required_fields_status.
used byParse the intake packet
mark_blockerautopilot-safe
Add or remove an entry on submissions.conversion_blockers. Use when something is preventing the submission from being converted that does not show up in required_fields_status (e.g., reporter says policy has lapsed, contradiction between policy address and reported address).
schedule_followup
Schedule a future SMS or email follow-up. Use when you've asked once and want to nudge again if there's no reply by a deadline. Always pick a reasonable due_at (24-48h for first SMS retry, 3-7 days for email). The adjuster can cancel or run-now from the UI.
used byRun three-point contactManage statutory clocks
cancel_followupautopilot-safe
Cancel a scheduled follow-up activity (followup_sms / followup_email) by its activity_id. Use this when the reason for the follow-up has been resolved — e.g., after the holdback release is approved, cancel the 30/60/180-day receipt-reminder follow-ups for that exposure since the holdback is no longer outstanding. Idempotent: cancelling an already-completed or already-cancelled follow-up returns ok=false but does not error. Always include a clear `reason` (the rationale field).
archive_submission
Terminal. Archive the submission (intake_status → abandoned) when the inbound correspondence is not an actionable claim — spam, a sales pitch, a misdirected message, a duplicate of an existing matter, or a question that needs no claim. Use this instead of escalate_to_human when no human follow-up is warranted. Provide a clear reason; the adjuster sees it in the activity log. This action requires human approval.
set_candidate_policyautopilot-safe
Link the submission to a candidate policy by policy number. Use during email intake once the reporter has given (or you've extracted from an attachment) their policy number, so the candidate policy + its coverages appear in your context and the submission can be converted. Returns matched=false if no policy matches — ask the reporter to confirm the number. The candidate policy is required before convert_submission_to_claim.
attach_to_claim
Terminal. Route this submission's correspondence (the email thread + any attachments) onto an existing claim instead of creating a new one, and close the submission as a duplicate. Use when find_claims shows the inbound email is about a claim that already exists. The claim's driver then picks up the new message and documents. This action requires human approval.
convert_submission_to_claim
Terminal. Convert the current submission into a Claim row. The claim driver will then open exposures explicitly via `create_exposures`. `coverage_codes` records your coverage hypothesis as advisory metadata on the conversion activity — pick the coverage codes you believe the loss implicates, but no exposure rows are created here. Only call when (a) required_fields_status is all satisfied AND (b) conversion_blockers is empty AND (c) you can confidently pick the coverage codes implicated. This action requires human approval.
escalate_to_human
Terminal. Flip the submission to escalated and pull the driver out. Use for: contradictory info that you cannot resolve, signs of fraud, policy lapsed, attorney represented from the start, third-party liability beyond personal property scope.
used byAssemble compensability recommendation
verify_policy
Verify whether the claim's policy was in force at the date of loss. The deterministic lookup + plain-language rationale are computed as soon as you call this tool; when human approval is required, the adjuster sees the AI's verdict on the approval card and can confirm or override before it lands on the claim. The verdict is persisted to claim.extended_fields.policy_verified + policy_verification_notes and recorded as a 'policy_verification' work-activity. Call this as the FIRST step on a fresh claim. Call with no arguments — the lookup happens server-side.
create_exposures
Open one or more Exposures on the claim. Each exposure is the unit money attaches to — one Exposure per CoverageGrant the loss actually implicates. Use `policy_snapshot.coverageGrants` as the source of truth for `coverage_grant_id` (do not invent IDs); typical burst-pipe pattern is Coverage A (dwelling) + Coverage C (contents), plus D (ALE) if the home is uninhabitable. `claimant_party_id` is normally the insured (reporter on the submission); use a different party only when a third party is the claimant for that coverage line. `valuation_method` defaults to the CoverageGrant's own valuation_method — only override when you have a specific reason. Existing exposures on the same coverage_grant_id are silently skipped (idempotent on retry). Requires human approval when autopilot is off.
set_reserves
Set reserve amounts on one or more Exposures on this claim. Reserves are the insurer's running internal estimate of what each exposure is likely to cost — split into `reserve_indemnity` (paid to/for the insured for the loss itself: repairs, contents, ALE) and `reserve_expense` (fees, vendor costs, independent adjuster, etc.). At intake/exposure opening, base the number on cause-of-loss severity and the CoverageGrant's limit. Once an inspection or repair estimate is on file, re-call this tool to true up reserves to the estimate's per-coverage RCV — reserves are documentation and should track the latest evidence. Per-exposure `rationale` explains how you arrived at the amount (e.g., 'mid-size burst pipe, drywall + flooring two rooms ≈ $20k indemnity' or 'trued up from inspection report: Coverage A RCV $7,171.32'). Requires human approval when autopilot is off; auto-applied when autopilot is on.
compute_payoutautopilot-safe
Compute the deterministic payout breakdown for each exposure once the inspection report's per-coverage RCV and recoverable depreciation are known. The tool takes RCV and recoverable_depreciation per exposure (you read these off the verified inspection report fields and split per coverage), then deterministically: (1) computes ACV = RCV − depreciation, (2) applies the policy deductible exactly once — to the exposure with the highest RCV — using the deductibleAmount from that coverage's grant (typical: Coverage A's $1,000), (3) computes net_acv_payable_now = ACV − deductible (or just ACV for non-bearer exposures), (4) writes `depreciationAmount`, `appliedDeductibleAmount`, and `recoverableDepreciationOutstanding` to each Exposure. Returns the per-exposure breakdown including net_acv_payable used as the input to `propose_payment`. Auto-applied — pure math + DB write, no human gate.
compute_business_incomeautopilot-safe
Compute the Business Income (time-element) loss for a commercial-property claim, once the period of restoration is known from the engineering/restoration estimate. Deterministically computes gross BI = rooms_down × nights × daily_rate, subtracts the waiting-period (time) deductible, adds continuing expenses to get the BI loss, and reports Extra Expense alongside. Writes the BI loss to the exposure's indemnity reserve and stamps the full breakdown into extended_fields. Pass `reasonable_restoration_nights` so the tool flags when actual downtime exceeds the reasonable restoration period. ONLY run this after a restoration/engineering estimate has produced the period of restoration. Large payouts still gate at `propose_payment` for human/forensic approval. Auto-applied — pure math + DB write.
forfeit_recoverable_depreciationautopilot-safe
Zero out the outstanding recoverable depreciation (RCV holdback) on an exposure. Use this when the receipt-submission window has expired without receipts being provided (default: 180 days from the initial payment). The Config carries the timeline as a rule; the driver calls this tool only after the timer fires. Per-call `reason` is required — typically cites the cumulative timer (e.g., '180 days elapsed without receipts; per Config rule the holdback is forfeited.'). Auto-applied — no human gate; clearing an already-zero holdback is a no-op.
propose_paymentautopilot-safe
Propose a payment for a SINGLE exposure. ⚠️ ONE call per exposure — never combine multiple exposures into one payment. The tool validates that `amount` matches the exposure's own computed payout (reserve_indemnity − depreciation_amount − applied_deductible_amount for non-holdback indemnity; recoverable_depreciation_outstanding for holdback release); a mismatch is a hard error, usually meaning you lumped values across exposures or skipped compute_payout. Creates a `payments` row at `status='proposed'` and a `pending_approval` activity that the adjuster sees on the claim's Overview tab. The adjuster can edit the amount before approving, or reject with a reason. Once approved, the payment row flips to `issued` and the driver re-ticks. Payee logic is deterministic: insured-only when amount < $5,000 OR the exposure is not Coverage A/B (mortgagee has insurable interest in dwelling only); joint with the highest-priority mortgagee on the policy otherwise. Pass the per-exposure breakdown for the initial post-inspection payment — it surfaces in the approval UI as a calculation explanation. For holdback-release payments, set is_holdback_release=true and amount=outstanding holdback; breakdown can be null. Per-call `rationale` is shown alongside the proposal — cite the calculation.
add_party_to_claimautopilot-safe
Attach a party to this claim with a role (e.g., 'inspector', 'contractor', 'public_adjuster'). Creates a party_on_claims row so the party appears in the stakeholders list and can be messaged via send_email / send_sms. Idempotent on (claim, party, role) — re-calls return alreadyLinked=true without duplicating. Use for stakeholders the driver brings into the claim itself (a default inspector for an inspection, a contractor the insured engages, an attorney). The insured/reporter is already linked on claim creation — don't re-add them.
set_claim_fieldautopilot-safe
Merge a single key/value into claims.extended_fields. Key must match an existing claim-scope FieldDefinition (rendered in your context under field_definitions). Use this to record decisions like `inspection_status='scheduled'`, `inspection_scheduled_at=<ISO 8601>`, or `inspector_party_id=<uuid>`. Value type should match the FieldDefinition.type (ISO 8601 for datetime, enum string for enum, etc.). Internal bookkeeping — not gated by autopilot.
used byParse the intake packetCompute AWW / TTD / PD rate correctlyRun three-point contactManage statutory clocksAssemble compensability recommendation
classify_documentautopilot-safe
Override or set the classification of a document (e.g., correct a misclassified estimate that should be an inspection_report). Used when the automated classifier got it wrong or for manually-uploaded documents. Internal bookkeeping — runs without approval even when autopilot is off.
used byParse the intake packet

Automations

Event-driven actions. When the trigger fires (and the optional condition matches), the agent runs the steps in order. Use sparingly — the agent already handles most cases on its own.

No automations. The agent acts only on its own ongoing judgment.

Checkpoints

Human approval gates. When a checkpoint's condition matches an action the agent is about to take, that action is queued for an adjuster to approve before it runs. First match wins.

Approve claim setup from intake packet
tool_id = "convert_submission_to_claim"
Approve outbound email to a party
tool_id = "send_email"
Approve outbound call to a party
tool_id = "initiate_outbound_call"
Approve AWW / TTD / PD computation
tool_id = "compute_aww"
Approve compensability decision
tool_id = "set_claim_field" AND args.key = "compensability_status" AND args.value in ["accepted", "denied", "delayed"]
Approve §9812 acceptance / benefit notice
tool_id = "send_templated_email" AND args.template = "wc_9812_acceptance_notice"
Approve escalation to a human
tool_id = "escalate_to_human"

Deadlines

Time-based escalations. When a deadline passes without the expected outcome, the agent runs the breach action — usually an escalation, reminder, or notification.

Name
From
Hours
On breach (tool)
Name
From
Hours
On breach (tool)

Configuration overlays

Conditional changes that layer on top of the base workflow as a claim's situation evolves — jurisdiction, severity, claim phase, or anything else. When a claim matches an overlay's condition, that overlay's edits are merged into the configuration the agent uses.

01
Wage basis on the employer report is deficient
inactive· order 1 of 4
When field 39 was left blank on a class code that typically works overtime, chasing the certified 52-week wage statement is the top priority.
when wage_basis_flag = true
patchdriver_prompt: operating_principles
02
Form 5021 field 26 restrictions blank
inactive· order 2 of 4
When the treating physician released the worker to modified duty without specifying restrictions, the modified-duty assignment cannot be validated.
when field_26_restrictions_missing_flag = true
patchdriver_prompt: operating_principles
03
Employee is represented
inactive· order 3 of 4
When a representation letter is on file, block all direct outbound to the employee and reroute to counsel.
when represented_flag = true
patchdriver_prompt: operating_principles
04
LC §5402(b) 90-day decision imminent
inactive· order 4 of 4
Inside 14 days to the 90-day compensability decision, the recommendation outranks everything else on the file.
when days_to_5402b_decision ≤ 14
patchdriver_prompt: operating_principles

Playground · preview overlays against a hypothetical claim

Pretend a claim has these attributes. As you change them, watch which overlays start applying and how the workflow the agent uses shifts in real time.

0 of 4 overlays active

Claim attributes

The same attributes overlays match against.

state

cause_of_loss

phase

severity

total_exposure

$

days_open

mitigation_status

siu_flagged

attorney_represented

Effective config · diff vs base

live

No overlays matched. The agent sees the base workflow exactly as authored.

symphony-wc

Overview

Driver brief

Role & objective

Operating principles

Capabilities preamble

Tone & defaults

Domain knowledge

Capabilities

Parse the intake packet

Compute AWW / TTD / PD rate correctly

Run three-point contact

Manage statutory clocks

Assemble compensability recommendation

Enabled Tools

Automations

Checkpoints

Approve claim setup from intake packet

Approve outbound email to a party

Approve outbound call to a party

Approve AWW / TTD / PD computation

Approve compensability decision

Approve §9812 acceptance / benefit notice

Approve escalation to a human

Deadlines

Configuration overlays

Wage basis on the employer report is deficient

Form 5021 field 26 restrictions blank

Employee is represented

LC §5402(b) 90-day decision imminent