Scribe

Scribe is a sales employee agent that turns a rep's call recap into real CRM work, updated Opportunity fields, follow-up Tasks, logged call notes, and new Contact records for stakeholders who came up on the call but aren't in Salesforce yet. The rep pastes or dictates what happened; Scribe does the 10–15 minutes of post-call data entry that reps otherwise skip or half-do.

Every change Scribe ever makes is written to a dedicated audit trail, reviewable through its own app, tabs, LWCs, and list views. Nothing Scribe does is invisible.

---

Table of contents

• Why Scribe
• Design goals
• The three surfaces
• Architecture and data flow
• The five agent topics
• Data model
• Component inventory
• Guardrails & safety
• Install & deploy
• Configuration reference
• Testing
• Project structure
• Status & known environment notes
• License

---

Why Scribe

Reps hate post-call admin, so it rots: Opportunity fields go stale, verbal commitments ("I'll send pricing," "we'll loop in their CFO") never become Tasks and quietly get forgotten, new stakeholders mentioned in passing never get added as Contacts, and deal notes live in the rep's head instead of the record.

None of that is laziness, turning a conversation into structured CRM fields is tedious work that adds value to the org's data, not to the rep. Scribe's job is to make "give me a two-sentence recap" produce the same CRM outcome as a rep spending fifteen minutes doing it by hand, and to make every automated change fully visible afterward.

Design goals

Goal	How Scribe honors it
Never invent data	Every field update, task, or contact traces to something literally in the recap. The deterministic extraction engine (Scribe_ExtractionService) is the only place text is interpreted, and it flags lowConfidence instead of guessing.
Confirm before writing	Opportunity updates and new Contacts use a two-phase propose → apply flow; the agent only calls the apply action after the rep explicitly says yes. The two risky actions also carry require_user_confirmation.
One extraction, many outputs	The recap is parsed once by Log This Call and cached on Scribe_Call_Log__c.Extraction_JSON__c. Every other topic reuses that single parse via the call-log id, no re-pasting, no inconsistent re-parsing.
No new habits	Extraction works on bullets, stream-of-consciousness paragraphs, or a pasted transcript. No required format.
Fit the deal cycle	Flag Missing Info compares the deal against per-stage requirements and creates reminder tasks for what's still missing, actively catching deal risk, not just recording history.
Full transparency	Every write, no exceptions, funnels through Scribe_ChangeLogService and produces a Scribe_Change_Log__c row with before/after values and a link back to the source recap.

The three surfaces

1. The employee agent (conversational), Scribe_Agent, an Agentforce AiAuthoringBundle with five topics. The rep talks to it; it does the CRM work.
2. The automation engine (headless), five autolaunched Flows, each wrapping an invocable Apex action that performs the write and logs the audit row as its final step.
3. The Scribe app (review & audit), a Lightning app where reps and admins see everything Scribe has ever done, by object, rep, date, or change type.

Architecture and data flow

Rep recap ──▶ Scribe agent (Agentforce)
                     │
                     ▼
        Scribe_ExtractionService  ── one structured parse ──▶ cached on Scribe_Call_Log__c
                     │
   ┌─────────────┬───┴────────┬──────────────┬─────────────────┐
   ▼             ▼            ▼              ▼                 ▼
 Log call   Update Opp   Create Tasks   Flag gaps      Detect Contacts
 (additive) (confirm)    (additive)     (additive)     (confirm)
   │             │            │              │                 │
   └─────────────┴────────────┴──────────────┴─────────────────┘
                     │
                     ▼
             Scribe_Change_Log__c   ← every action writes exactly one audit row

The change log is the single source of truth for "everything Scribe has ever done," independent of whether someone views it via the app, a list view, or the record-page badge.

The five agent topics

#	Topic	Risk	What it does
1	Log This Call	Additive	Extracts stage-relevant facts, objections, and sentiment; logs a Scribe_Call_Log__c; caches the one parse every other topic reuses.
2	Update Opportunity	Confirm	Proposes Stage / Close Date / Amount / Next Step changes field-by-field; writes only after explicit confirmation; audits each field with before/after.
3	Create Follow-Up Tasks	Additive	Turns each commitment ("I'll…", "we will…") into a Task with an inferred owner and due date.
4	Flag Missing Info	Additive	Compares the deal (this recap + all prior call logs) against Scribe_Stage_Requirement__mdt for the current stage and creates reminder Tasks for the gaps.
5	Multi-Contact Detection	Confirm	Drafts Contacts for stakeholders mentioned but not yet on the deal; creates them (linked via OpportunityContactRole) only after confirmation.

Data model

Custom objects

• Scribe_Call_Log__c, the raw recap/transcript, a link to the Opportunity, extracted sentiment/objections, and Extraction_JSON__c (the cached single parse). The source every change-log row traces back to.
• Scribe_Change_Log__c, one row per action Scribe takes. Change_Type__c (Call Logged / Field Update / Task Created / Stage Gap Task / Contact Created), Related_Record_Id__c + Object_API_Name__c, Field_Name__c / Old_Value__c / New_Value__c (field updates only), Source_Call_Log__c, and Created_By_Rep__c (the human behind the conversation, distinct from the running user).

Custom metadata

• Scribe_Stage_Requirement__mdt, what "complete" means per Opportunity stage (economic buyer, budget, timeline, …), with detection keywords. Tunable per sales process without a code change. Ships seeded for Qualification, Needs Analysis, Value Proposition, Proposal/Price Quote, and Negotiation/Review.

Standard-object additions

• Task.Scribe_Generated__c, Contact.Scribe_Generated__c, mark agent-created records so the "Created by Scribe" list views are trivially reliable (more robust than filtering on the integration user's CreatedById, and deployable without knowing that user's Id).
• Task.Scribe_Source_Call_Log__c, Contact.Scribe_Source_Call_Log__c, trace each generated record back to its recap.

Component inventory

Type	Components
Agent	Scribe_Agent (AiAuthoringBundle, 5 topics)
Apex, core	Scribe_ExtractionService, Scribe_CallParse, Scribe_ChangeLogService, Scribe_Constants, Scribe_Util
Apex, invocables	Scribe_LogCallInvocable, Scribe_UpdateOpportunityInvocable, Scribe_CreateTasksInvocable, Scribe_CheckStageRequirementsInvocable, Scribe_DetectNewContactsInvocable
Apex, controllers	Scribe_ChangeLogConsoleController, Scribe_HomeDashboardController
Apex, tests	Scribe_ExtractionService_Test, Scribe_Invocables_Test, Scribe_Controllers_Test, Scribe_TestUtil
Flows	Scribe_Log_Call, Scribe_Update_Opportunity, Scribe_Create_Followup_Tasks, Scribe_Check_Stage_Requirements, Scribe_Detect_New_Contacts
LWCs	scribeChangeLogConsole, scribeHomeDashboard, scribeRecordHistoryBadge
App / tabs	Scribe app; tabs: Scribe Home, Change Log, Change Log records, Call Logs (+ standard Task/Contact)
List views	Task & Contact "Created by Scribe"; Change Log "This Week / By Object / Field Updates Only"
Permission sets	Scribe_Agent_User (write, for the running/integration user), Scribe_Reviewer (read-only review access)

Guardrails & safety

• Confirmation is enforced two ways: the propose→apply action split (the agent only calls apply after an explicit yes) and require_user_confirmation: True on the two write actions.
• Graceful degradation: when a recap is too vague, Scribe_ExtractionService returns lowConfidence and the agent asks the rep to clarify rather than writing a low-confidence value.
• The audit row is mandatory, not configurable. Every write path calls Scribe_ChangeLogService; there is no code path that writes without auditing.
• Apex runs with sharing; controllers are read-only and use bound variables (no SOQL injection).

Install & deploy

Requires an Agentforce-enabled org for the agent bundle. The data model, Apex, Flows, app, and LWCs deploy to any org.

# 1. Set your target org
sf config set target-org <alias>

# 2. Deploy everything
sf project deploy start --source-dir force-app

# 3. Assign permission sets
sf org assign permset --name Scribe_Agent_User      # running/integration user
sf org assign permset --name Scribe_Reviewer        # reps & managers who review

# 4. Validate, publish and activate the agent (Agentforce org)
sf agent validate authoring-bundle --api-name Scribe_Agent
sf agent publish  authoring-bundle --api-name Scribe_Agent
sf agent activate --api-name Scribe_Agent

Then add the Scribe Record History badge to the Opportunity Lightning record page in App Builder.

Every object with a tab in the Scribe app ships an unfiltered All list view (filterScope: Everything), and that is what each tab should default to - the narrower views ("Created by Scribe", "This Week", "Field Updates Only") are filters a reviewer chooses deliberately, never the landing state.

Configuration reference

Tune Flag Missing Info entirely through Scribe_Stage_Requirement__mdt, no code change:

Field	Purpose
Stage_Name__c	The Opportunity StageName this requirement applies to.
Requirement_Key__c / Requirement_Label__c	Machine key + human label (used in task subjects).
Detection_Keywords__c	Comma-separated, case-insensitive phrases. If any appears across the deal's recaps, the requirement is "covered."
Reminder_Task_Subject__c / Reminder_Due_In_Days__c	Subject and due window for the gap reminder Task.
Sort_Order__c / Is_Active__c	Ordering and on/off.

Testing

Apex tests cover the extraction engine, all five invocables (including propose→apply, dedup, and bad-input paths), both controllers, and the utilities:

sf apex run test --tests Scribe_ExtractionService_Test --tests Scribe_Invocables_Test \
                 --tests Scribe_Controllers_Test --result-format human --code-coverage

Static analysis (SLDS/ESLint) is clean:

sf code-analyzer run --target force-app

Project structure

force-app/main/default/
├── aiAuthoringBundles/Scribe_Agent/   # the employee agent (5 topics)
├── applications/                       # Scribe Lightning app
├── classes/                            # extraction, invocables, controllers, tests
├── customMetadata/                     # seeded stage requirements
├── flexipages/                         # Scribe Home, Change Log app pages
├── flows/                              # 5 autolaunched action flows
├── lwc/                                # console, dashboard, record badge
├── objects/                            # Call Log, Change Log, Stage Requirement + Task/Contact fields
├── permissionsets/                     # Agent User, Reviewer
└── tabs/                               # Scribe Home, Change Log, Call Logs

Status & known environment notes

Deployed and verified to a Developer Edition org: all objects, the stage-requirement type and its 13 seed records, 16 Apex classes, 5 Flows, 3 LWCs, 2 flexipages, 4 tabs, the Scribe app, both permission sets (assigned), and the Scribe_Agent AiAuthoringBundle. All 20 Apex tests pass at 94% coverage of the exercised classes.

Two org-specific deployment notes, already handled in this repo:

• Task fields live on Activity. Scribe_Generated__c and Scribe_Source_Call_Log__c are defined under objects/Activity/ (Task is a sub-type of Activity, so a field placed directly on Task is rejected). They still surface on Task as Task.Scribe_Generated__c etc.
• Seed CMT records deploy via the Apex Metadata API. This org returns an UNKNOWN_EXCEPTION gack when deploying custom-metadata records through the Metadata API (reproducible even with a single minimal record, an org/platform issue, not the files). The 13 rows were loaded with Metadata.Operations.enqueueDeployment instead; on a healthy org sf project deploy start handles them normally.

Remaining, environment-limited:

• Apex PMD/SFGE static analysis was not run here because no Java runtime is installed (sf code-analyzer PMD/SFGE engines require Java 11). SLDS/ESLint analysis passes. Run sf code-analyzer run --target force-app on a machine with Java to exercise the Apex engines.
• Agent publish/activate: the bundle deploys cleanly, but sf agent validate/preview (the afscript/v2/parseandcompile reasoner endpoint) returned 422 across every available org, a service-side issue independent of the bundle. Re-run sf agent validate authoring-bundle --api-name Scribe_Agent when that service is reachable, then sf agent publish / sf agent activate.

Screenshots

Captured from a live org at a 1200px viewport.

Scribe Home. Field values overwritten leads the page, because it is the only change type that can replace what a human entered.	Change Log. The transcript view: what Scribe heard beside what Scribe wrote.
Call Logs. All Call Logs, the tab default.	Change Logs. All Changes, the full audit trail.

License

MIT © 2026 VurtuoLabs