Switchboard makes real API calls during conversations. Five production tools — search case, verify identity, look up status, update field, create case. Every call schema-validated, permission-bound, and Timeline-logged.
Most AI assistants can read your knowledge base and tell you what your policy says. Switchboard can update your policy, log a claim, schedule the survey, and confirm the appointment — during the same call.
When a customer asks “can you update my email address?” the AI does it (after identity verification). It does not tell the customer to log into the portal. It does not promise to pass it on to a handler.
Every tool invocation is a Timeline entry: actor (agent definition + version), timestamp, resource (case + field), outcome (success/failure/partial), severity. Auditable, exportable, attributed.
Each tool requires the prerequisites it needs. Update operations require a verified session. Sensitive fields require additional confirmation. Out-of-policy requests trigger escalation, not workarounds.
The Five Tools
These are the production tools the AI uses on every regulated conversation. Custom tools live alongside them.
search_caseLook up an existing case by registration plate, phone number, email, postcode, or customer name. The AI fetches the case file and any open workflow state before continuing the conversation.
Inputs
Outputs
Matched case ID, case status, owner, last activity, open tasks. Multiple matches surface as a disambiguation step.
Severity
Low — read-only.
verify_identityTwo-step verification on different fields: typically date of birth + postcode, or memorable info + reference. Issues a verified-session token that lasts 30 minutes. Verification is required before any case detail is shared.
Inputs
Outputs
Verified session (30-min default) bound to channel + case + caller. Failure routes to handler with diagnostic.
Severity
Medium — gates all subsequent tool use on the case.
lookup_statusFetch the current status of a case, sub-case, or task. By task ID, by reference, or by relationship to the active case. Supports the 'where is my claim?' conversation without escalating to a handler.
Inputs
Outputs
Current status, last update timestamp, expected next milestone, any blockers flagged.
Severity
Low — read-only on verified sessions.
update_fieldWrite a new value to a structured field on the case. Natural-language aliasing maps customer phrasing to canonical field names — “car reg” → vehicle_registration, “mobile” → phone_mobile. Validation runs before commit.
Inputs
Outputs
Updated field value, prior value preserved in Timeline, confirmation message rendered to caller.
Severity
Medium — write to case file, attributed to caller via verified session.
create_caseOpen a new case from scratch with workflow discovery — the AI selects the right workflow definition based on case type, customer, and conversation context. Used heavily for FNOL intake.
Inputs
Outputs
New case ID, workflow attached, initial routing applied, follow-up tasks scheduled.
Severity
Medium — creates new resource, full Timeline trail.
Example Call
An anonymised slice of a Switchboard FNOL call. Tool calls render inline so you can see what the AI is doing as the conversation unfolds.
Hi, I had a bump this morning. I want to put in a claim.
I'm sorry to hear that. Are you OK? Let me get you set up. Can you give me your registration plate first?
Yeah, AB12 CDE.
search_case→ Existing case found · Customer: J. Smith · Policy activeThanks. I've got your details. To confirm I'm speaking to John Smith, can I ask for your date of birth and the first line of your address?
12 March 1985. 47 Acacia Avenue.
verify_identity→ Verified · Session valid 30 min · Channel: voiceThanks John, you're verified. I'm going to take you through what happened. When and where did the incident occur?
create_case→ New FNOL case CL-184932 · Workflow: motor-fnol-v3 · Routed to first-notification queueI've opened your claim — the reference is CL-184932. Now, in your own words, what happened?
Every tool call is a Timeline event. Three tool invocations in this slice — search_case, verify_identity, create_case — each with actor, timestamp, resource, outcome, and severity logged. The case file reflects the conversation in real-time; the audit trail is complete before the call ends.
FAQ
Switchboard executes only the tools defined in its agent configuration. Anything outside that scope routes to a human handler with the conversation context. New tools can be added to the configuration during onboarding (or later) — they need an API endpoint, an input/output schema, and a permission policy. Once defined, the AI can use them; until then, it cannot.
Tools are scoped per agent definition. A TL-settlement agent has tools for case lookup and structured response capture; it does not have a create_case tool because TL is not a new-case conversation. Within the agent's allowed toolset, the LLM selects which tool fits the conversation — and the platform validates inputs against the schema before execution. Mismatches fail closed and escalate to a handler.
Each writable field has a canonical name plus a configurable alias list. “Car reg,” “numberplate,” “reg,” and “vehicle registration” all alias to vehicle_registration. The AI resolves aliases automatically. Aliasing rules are configured per tenant — easy to extend for industry-specific phrasing.
Tool failures are logged in Timeline with the diagnostic (timeout, validation failure, downstream error). The AI either retries (for transient failures), asks the caller a clarifying question (for input ambiguity), or escalates to a handler with the partial conversation and the failure reason. The caller is never told “the system is broken” without a path to a human.
Yes. Custom tools are part of agent configuration. Common additions in claims include: book_survey, schedule_callback, link_to_third_party, fetch_policy_document, send_consent_form. Each follows the same pattern — schema, permissions, severity, Timeline-logged. Most custom tools are added during the 30-day pilot.
Function calling is the LLM-side mechanic. Switchboard adds the operational layer around it — schema validation, permission policy, identity-binding, audit logging in Timeline, escalation on failure, multi-LLM routing so the same tool call works across providers. Building this from scratch on top of raw function calling is a 12-month project; we have it as platform infrastructure.
Run Switchboard alongside your existing process for 30 days. Watch the tool-call volume, the verification rate, the case-creation throughput. Decide.