create-disposition

Bolna Dispositions (Post-Call Extractions)

Dispositions turn raw transcripts into structured data after every call. Each disposition asks one question; results land in extracted_data on the execution payload and in the webhook delivery.

Hierarchy

Agent
└── Extractions (feature)
    └── Category  (e.g. "Lead Quality")
        └── Disposition  (e.g. "Call Outcome")

Category groups results visually. Disposition is the unit — one question, one answer set.

When to add a disposition

Use case	Field name	Answer type
Call outcome	Call Outcome	Pre-defined: interested / not_interested / follow_up / wrong_number
Lead qualification (BANT)	Budget, Authority, Need, Timeline	Mixed — free text + Yes/No
Appointment booked	Appointment Date, Appointment Time	Free text with timestamp / text validation
Sentiment	Customer Sentiment	Pre-defined: positive / neutral / negative
Agent handover required	Handover Needed	Pre-defined: Yes / No
Compliance	Disclosure Read, Consent Captured	Pre-defined: Yes / No
Contact info captured	Email, Alt Phone	Free text with email / regex validation
Open-ended summary	Main Concern, Reason for Cancellation	Free text

Endpoints

Endpoint	Purpose
GET /dispositions/	List all dispositions in the account (optionally scoped to an agent)
GET /dispositions/{id}	Get a single disposition
POST /dispositions/	Create one disposition, optionally linking to agents
POST /dispositions/bulk	Create N dispositions for an agent atomically
PUT /dispositions/{id}	Update a disposition (copy-on-write aware)
DELETE /dispositions/{id}	Delete a disposition
POST /v2/agent/{agent_id}/dispositions/test	Test all of an agent's dispositions against a transcript

Auth: Authorization: Bearer $BOLNA_API_KEY. JSON body for all POST/PUT.

Disposition object

{
  "name": "Call Outcome",
  "question": "What was the outcome of the call?",
  "system_prompt": "You are analyzing a sales call transcript.",
  "category": "Lead Quality",
  "model": "gpt-4.1-mini",

  "is_subjective": true,
  "is_objective": true,

  "subjective_type": "text",
  "subjective_type_config": null,

  "objective_options": [
    { "value": "interested",     "condition": "Customer expressed genuine interest and agreed to a next step" },
    { "value": "not_interested", "condition": "Customer declined all proposals" },
    { "value": "follow_up",      "condition": "Customer asked to be contacted later" }
  ],

  "agent_ids": ["3c90c3cc-0d44-4b50-8888-8dd25736052a"]
}

Field	Type	Notes
name	string	Display name in extracted_data.
question	string	The prompt the LLM evaluates against the transcript. Be specific.
system_prompt	string | null	Optional system context for the evaluation LLM.
category	string	Groups outputs. Default "General".
model	string	Default gpt-4.1-mini. Use a stronger model for high-stakes extractions.
is_subjective	bool	Enables subjective free-text response.
is_objective	bool	Enables objective pre-defined selection.
subjective_type	string	Free-text constraint. See below.
subjective_type_config	object	Required when subjective_type == "regex".
objective_options	array	Required when is_objective is true.
agent_ids	array	Agents this disposition is linked to.

At least one of is_subjective or is_objective must be true.

Free-text validation types

subjective_type	What it validates	Example value	Notes
text	Anything	"Customer was happy with the demo"	Default. No validation.
timestamp	ISO 8601	"2026-04-08T14:30:00"	Parsed and validated.
numeric	Int or decimal	"42", "3.14"
boolean	Exactly true or false	"true"
email	Valid RFC email	"[email protected]"
regex	Matches subjective_type_config.pattern	"1234567890"	Add pattern + optional description in subjective_type_config.

When validation fails, the original LLM response is preserved in subjective and validation.is_valid is false. No data is lost — the consumer decides whether to use it.

Example with regex (10-digit phone):

{
  "subjective_type": "regex",
  "subjective_type_config": {
    "pattern": "^\\d{10}$",
    "description": "10-digit phone number"
  }
}

Objective options

{
  "value": "interested",
  "condition": "Customer expressed genuine interest and agreed to a next step",
  "sub_options": []
}

sub_options is recursive and supports hierarchical classifications (e.g. interested → enterprise / pro / starter).

Output format on executions

{
  "execution_id": "abc123",
  "status": "completed",
  "transcript": "...",
  "extracted_data": {
    "Lead Quality": {
      "Call Outcome": {
        "subjective": "Customer expressed strong interest and asked about enterprise pricing.",
        "objective": "interested",
        "confidence": 0.92,
        "confidence_label": "High",
        "reasoning_subjective": "Customer asked about enterprise pricing and requested a demo.",
        "reasoning_objective": "Customer explicitly expressed interest and agreed to a next step.",
        "validation": null
      }
    },
    "Contact Info": {
      "Customer Email": {
        "subjective": "[email protected]",
        "objective": null,
        "confidence": 0.95,
        "confidence_label": "High",
        "reasoning_subjective": "Customer clearly provided their email address during the call.",
        "reasoning_objective": null,
        "validation": { "is_valid": true, "expected_type": "email" }
      }
    }
  }
}

Field	Notes
subjective	Free-text response. "" when nothing found, null when is_subjective is false.
objective	Selected pre-defined value, or null.
confidence	0.0 – 1.0.
confidence_label	"High" (≥0.8), "Medium" (≥0.5), "Low" (<0.5).
reasoning_subjective / reasoning_objective	LLM's explanation per answer type. Useful for audit + debugging.
validation	Present only when subjective_type is typed (not plain text).

Same payload arrives via webhook. See https://raw.githubusercontent.com/bolna-ai/skills/main/references/execution-payload.md.

Create one disposition

curl --request POST https://api.bolna.ai/dispositions/ \
  --header "Authorization: Bearer $BOLNA_API_KEY" \
  --header "Content-Type: application/json" \
  --data @- <<'JSON'
{
  "name": "Call Outcome",
  "question": "What was the outcome of the call?",
  "category": "Lead Quality",
  "is_subjective": true,
  "is_objective": true,
  "subjective_type": "text",
  "objective_options": [
    { "value": "interested",     "condition": "Customer expressed genuine interest and agreed to a next step" },
    { "value": "not_interested", "condition": "Customer declined all proposals" },
    { "value": "follow_up",      "condition": "Customer asked to be contacted later" }
  ],
  "agent_ids": ["3c90c3cc-0d44-4b50-8888-8dd25736052a"]
}
JSON

Returns the full disposition object with id. Save it if you intend to update later.

Bulk create (recommended for new agents)

Add a full analytics suite atomically — either all succeed or none do:

curl --request POST https://api.bolna.ai/dispositions/bulk \
  --header "Authorization: Bearer $BOLNA_API_KEY" \
  --header "Content-Type: application/json" \
  --data @assets/sample_bulk_dispositions.json

The asset file is a complete sales-call analytics pack: Call Outcome, Customer Sentiment, Budget mentioned, Decision Maker, Handover Needed, Customer Email.

Test before going live

Run all of an agent's dispositions against a transcript and inspect what comes out:

curl --request POST "https://api.bolna.ai/v2/agent/$AGENT_ID/dispositions/test" \
  --header "Authorization: Bearer $BOLNA_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "transcript": "Agent: Hi, how can I help? Customer: I am interested in your enterprise plan. Agent: Great — what email should I send the proposal to? Customer: priya at example dot com.",
    "call_date": "2026-05-19T10:00:00Z"
  }'

Response is the same shape as a real execution's extracted_data. Iterate on dispositions until results match what you'd want from production calls.

Updates and the copy-on-write rule

Scoped update (pass agent_id in the body or URL):

Disposition state	Behaviour	Status
Exclusive to this agent	Edited in place	200 OK (id unchanged)
Shared with other agents	New private copy created, agent re-linked to copy	201 Created (new id)

If you store disposition IDs in your own database, update the stored ID when you see 201. The old ID still exists — it just no longer belongs to your scoped agent.

Unscoped update (no agent_id):

• Admins: edits any disposition in place.
• Non-admins: only edits dispositions they own.

Authorization quirks

• Regular users can DELETE only dispositions they created (created_by matches their user ID).
• Admins can delete any disposition.
• Historical executions that already contain a disposition's output are unaffected by edits or deletion — only future calls change.

Going deeper

File	Contents
references/answer-types.md	Free Text vs Pre-defined, validation patterns, when to use each.
references/copy-on-write.md	Worked walkthrough of the 200 vs 201 update flow, with code.
assets/sample_bulk_dispositions.json	Drop-in sales-call analytics pack — 6 dispositions, two categories.
assets/sample_transcripts.json	Three transcripts (sales, support, appointment) for testing.
scripts/create_disposition.py	Wraps POST /dispositions/ with flags.
scripts/bulk_create_dispositions.py	Wraps POST /dispositions/bulk from a JSON file.
scripts/test_dispositions.py	Wraps POST /v2/agent/{id}/dispositions/test for sample transcripts.

Legacy: gpt_assistants.custom_questions

Older agents (and some agents migrated from the v1 dashboard) carry their post-call extraction logic inline on the agent itself, not as separate disposition objects. You'll see this when GET /v2/agent/{id} returns:

{
  "dispositions": null,
  "gpt_assistants": [
    {
      "assistant_id": "asst_f4sCpG4zASXIls0FMl7PkMMt",
      "custom_questions": "1. interested:\n[- What to Extract: User interest in becoming an expert\n- Expected Response: Yes or No]\n\n2. reason_not_interested:\n[- What to Extract: Reason stated by the user...\n- Expected Response: Short reason]\n\n3. full_name:\n..."
    }
  ]
}

Aspect	Legacy gpt_assistants	Modern dispositions
Where it lives	Inline on the agent, single free-text string	Separate disposition objects linked via agent_ids
Format	Numbered list, free-text "what to extract" + "expected response" hint	Structured: name, question, objective_options[], subjective_type
Validation	None — the LLM returns whatever it interprets	Typed (timestamp, email, regex, etc.) with validation.is_valid flag
Confidence	No	Yes — confidence + confidence_label
Reusable across agents	No — copied per agent	Yes — one disposition can link to N agents
Where the result lands	Same extracted_data field on the execution	Same extracted_data field on the execution

Both systems coexist. If dispositions is set, it runs. If dispositions is null but gpt_assistants[0].custom_questions has content, that runs instead. The output structure on the execution is the same shape either way.

Migration: when you ask the user to "add dispositions" to a legacy agent, you typically:

1. Parse the existing custom_questions string — one numbered item per intended disposition.
2. Convert each item into a structured disposition (POST /dispositions/).
3. Optionally clear gpt_assistants once the new dispositions are linked, so it's unambiguous which system is producing the output.

You don't need to migrate to use modern features — but you can't get confidence scores, typed validation, or shared dispositions without it.

Common gotchas

• Vague questions yield low confidence. "How was the call?" → noisy. "Did the customer agree to a follow-up meeting? If yes, on what date?" → clear.
• objective_options.condition is the LLM's selection rubric. Write it as if instructing a human reviewer.
• Don't pile many concepts into one disposition. Split "lead quality and next step" into two dispositions — outcome + action.
• Pick the right answer type. Categorical = is_objective. Open-ended = is_subjective. Both = both flags true (lots of dispositions benefit from both).
• Confidence ≥ 0.8 is generally trustworthy. Below 0.5 → review manually. Build CRM routing around confidence_label, not raw transcript heuristics.
• agent_ids is the linkage. A disposition without agent_ids exists but doesn't run on any agent. Bulk-create is the easiest way to ensure every new disposition is linked.

See also

• get-executions — fetching extracted_data after calls.
• setup-webhook — receiving extractions in real time.
• create-agent — top-level agent config (Extractions is a feature of the Analytics tab).
• https://raw.githubusercontent.com/bolna-ai/skills/main/references/execution-payload.md — shape that contains extracted_data.