# 20 — Product scope and use-case catalog

**Status:** Product direction confirmed by the user on 2026-11-12. This document defines
intended scope and dependencies. It does not certify implementation or availability.

## 1. What customers buy

Charla lets people and organizations create agents they can delegate work to through
conversation. Each agent combines instructions, knowledge, memory, tools and explicit limits.
The platform provides execution, identity, permissions, approvals, human handoff, channel
routing, versions, evaluations, observability and usage billing.

The audience includes individuals, teams, businesses, agencies, capability builders and
software companies embedding agents in their products. The launch footprint and locales stay
as defined in [03](03-market-and-channels-latam.md). WhatsApp is the first business messaging
integration; Telegram and API are core surfaces too. Customers choose channels suited to the
job and their eligibility. The platform's purpose remains the same across those surfaces.

The reference is Keiki's broad platform model. Its public [product description](https://onkeiki.com/about),
[template catalog](https://onkeiki.com/templates), [capability model](https://onkeiki.com/docs/capabilities)
and [partner deployments](https://onkeiki.com/docs/partner-deployments), reviewed on 2026-11-12,
show personal assistance, learning, internal work, customer-facing agents and infrastructure
for other products. Keiki template names below identify reference examples, not Charla brands
or integrations that we have delivered.

## 2. Product families

| Family | Job and observable outcome | Required capabilities | Delivery dependency |
|---|---|---|---|
| Personal assistance (Orchid reference) | Organize tasks, retrieve personal information and follow up commitments; confirm saved tasks or completed changes | Private memory, account-specific tools, identity and approvals; scheduling for reminders | Basic tool-backed examples on Telegram/API in Phase 1; account linking and schedules in Phase 2; each email/calendar connector needs validation |
| Learning and coaching (Study Coach / Office Hours) | Explain, quiz, maintain a learning plan and follow up progress | Knowledge, memory, evaluated guidance; schedules for recurring review | Basic evaluated examples in Phase 1; proactive schedules in Phase 2 |
| Company knowledge (Company Brain) | Answer from authorized sources with citations and create or follow up work items | Source connectors, retrieval, per-user permissions, tool results | Tool-backed examples on API in Phase 1; connector-specific work and Slack/Teams follow the roadmap |
| Internal operations and back office (Leaf / Night Drop) | Triage documents, inspect billing, propose changes and execute approved operations | Document tools, identity, scoped tools, previews, approvals, audit | API/Telegram examples with ready tools in Phase 1; email intake and schedules in Phase 2; Slack/Teams in Phase 3 |
| Customer service, sales and reception | Qualify a lead, record a booking, collect a payment or resolve a request | Business knowledge, CRM/booking tools, payments, memory, handoff | Existing five SMB verticals are the initial content cohort; actual integrations and payment providers need separate validation |
| Research and web errands (Instinct) | Research options, complete a web task or purchase and report evidence of the result | Browser, scraping, sandbox where needed, personal authentication and purchase controls | Phase 3 or later; account access, spending and result verification must be validated before release |
| Embedded agents for software products | Provision an agent for a product's user and expose it through the product's chosen channels | Partner deployments, SDK, API/MCP, identity and external harness | Direct agent API in core; partner deployments and SDK/external harness in Phase 2 |
| Agency operations and reusable capabilities | Automate the agency's own prospecting and client onboarding; operate agents for clients; distribute functionality | Tenant isolation, versions, capabilities, marketplace and commercial controls | Existing organization boundary remains; marketplace in Phase 3; white-label and revenue share in Phase 4 |

The template catalog (`catalog.family`, T-154) has three families: `customer`, `personal` and
`internal`. Learning and coaching templates belong to `personal`, as [01 §8](01-product-vision.md)
and [11 §2.1](11-dashboard-and-ux.md) group them; the `education` sector tells them apart. There
is no `learning` family. The sectors added for these families are `personal`, `education` and
`operations` (decided by the Lead on 2026-11-15, [18](18-open-questions.md) H1 and H7).

An agent may combine families. An internal billing agent and a personal assistant use the same
execution and governance contracts. Their sources, credentials and authorized users differ.

## 3. Separate job, capability, execution and channel

- **Agent definition:** instructions, model, knowledge, memory settings, tools and policies.
- **Capabilities:** reusable functionality that adds tools, guidance, schedules or stored state.
- **Execution:** cloud by default; SDK and external harness according to delivery phase.
- **Deployment:** channel identity, credentials, channel constraints and entry points.
- **Operation:** versions, evals, traces, actions, handoffs, usage and billing.

An agent template is a starting configuration for a job. A WhatsApp message template is an
approved outbound message used under that channel's rules. They are different product objects.
Choosing WhatsApp, Telegram or API does not by itself supply an account connector, knowledge
source, browser or scheduler.

An organization is the isolation and billing boundary, including for an individual workspace.
Contacts can be personal users, coworkers or customers. Sensitive tools always require the
appropriate identity, scopes and approvals; joining a group does not grant access to internal
or personal data. Credentials stay out of prompts and tools report confirmed outcomes.

Channel eligibility must follow current provider terms. The WhatsApp policy assumptions and
verification gaps are in [03](03-market-and-channels-latam.md). Personal/general-purpose agents
must use eligible channels such as Telegram/API; a channel restriction does not narrow the
whole product to customer service. No automatic cross-channel identity linking is implied.

## 4. Delivery and acceptance

The [roadmap](17-roadmap-and-milestones.md) retains its infrastructure stages. The WhatsApp
walking skeleton and initial SMB pilots are useful delivery slices. They do not validate all
product families or complete the platform catalog.

Existing work and planned gaps:

| Area | Existing task or contract | Remaining product work |
|---|---|---|
| Initial template content | [T-094](../tasks/T-094-templates-content.md), five verticals × ten locales | Personal/learning and internal-work content, integrations and evaluation suites need separately scoped implementation tasks before delivery; T-094's existing cohort stays intact |
| Template catalog | [T-154](../tasks/T-154-template-catalog-metadata.md): each template file declares its use (`catalog.family`), sector, icon and order, and the gallery, its filters and the onboarding are built from the files; installing sends the template's WhatsApp messages to Meta's review, once a number is connected | A new vertical or family needs only its files and its sector's translations to appear; its content, integrations and evals are still separate tasks. Personal and internal templates get no WhatsApp messages (channel eligibility) |
| Job-first onboarding | [T-132](../tasks/T-132-onboarding.md) | Implement and validate WhatsApp, Telegram and API branches, eligibility and capability requirements; no WABA required on Telegram/API |
| Agent operations | [Runtime](07-agent-runtime.md), [governance](09-tools-identity-governance.md), [evals](12-evals-observability.md) | Verify each new job against these contracts; a generic tool dispatcher does not establish a completed connector or product |
| Proactive and embedded work | Phase 2 of the roadmap | Deliver scheduling, account linking, partner provisioning and SDK/external harness before dependent use cases are advertised |
| Browser, sandbox, marketplace and coordinated agents | Phases 3–4 of the roadmap | Implement, govern and evaluate each capability before releasing dependent catalog entries |

Each released template or capability needs an explicit job, eligible channels, required
accounts and permissions, human escalation behavior and automated evals. Evals must check the
expected result, refused access, tool failure and relevant limits. Purchases and writes must
be confirmed by tool results; personal data and team sources must remain isolated.

Locales per family (decided by the Lead on 2026-11-15, [18](18-open-questions.md) H3):

- `customer` templates ship in every locale of `TEMPLATE_LOCALES`, as T-094 does.
- `personal` and `internal` templates ship in at least one Spanish locale (`es-419` unless the
  job targets a country) and `pt-BR`, and may add more. They do not wait for the base + country
  packs of T-171 (#200). The template schema and `templates:lint` enforce this rule (#386,
`requiredTemplateLocales`), together with the channel rule: a family that may not use
WhatsApp ships no WhatsApp messages and runs its eval cases on Telegram or the API.

The native review of each locale is done by a reviewer agent (LLM) with a rubric per locale,
recorded in the template's `review`. There is no human native reviewer. This is an accepted
risk: a model can miss vocabulary, formality or legal wording that a native speaker would
catch, so a locale reviewed this way is not claimed as natively reviewed by a person
([18](18-open-questions.md) H6). The file records it as `review.status: "agent-reviewed"`, with
the `reviewer` agent, the locale's `rubric`, the date and the changes in `notes`, and leaves
`nativeReviewer` empty; `approved` stays for a native speaker's signature.

Without an identity webhook, an owner or admin can grant scopes to a specific contact from the
dashboard, within the scopes the agent's tools declare; the identity webhook wins when both
exist ([ADR-0018](adr/0018-dashboard-scope-grants.md)).

The MVP validates tool-backed examples in three families: personal assistance/learning,
internal work/knowledge and customer-facing work. Measure time to a working agent and task
outcomes by family and channel. Keep WhatsApp-specific pilot, latency and transcription
metrics for that cohort. Targets are in [01 §8](01-product-vision.md); a requirement added here
does not change an existing task's status or mark any acceptance criterion as passed.

## 5. Commercial model and product language

Support two adoption paths: hosted agents, and connecting an existing agent while its owner
retains the model and execution. The second uses the external harness planned for Phase 2.
The [Vercel case and other blog lessons](research/keiki-blog-product-lessons.md) motivate this
path; they do not imply that Charla ships a Vercel connector.

Sell the ability to create and operate agents with connected tools and control. Subscription
and metered execution/capabilities follow [14](14-billing-and-plans.md); prices are provisional.
Number rental and provider message fees belong to channel cost lines. Marketplace revenue
share and agency white-label retain their later phases.

Public explanations should start with the delegated job and its outcome, show what the agent
can access and when a person decides, then show how to reach it. Describe channels as entry
points. Use examples from personal, internal and customer-facing work, with availability
stated per example. Keep this product promise distinct from the currently shipped demo and
initial SMB template content.

## 6. Complete journeys and bounded pilots

The [pilot plan](21-product-validation.md) covers personal account-linked assistance, internal
operations with verified artifacts and embedded product operations, including an existing
external agent. Start Phase 1 with ready tools; complete the dependent Phase 2 gates before
advertising account linking, durable follow-up or external execution as available products.

Candidate pilot jobs include meeting preparation from existing data, capture of voice notes
and next steps, sourced sales opportunities and preparation of a new client's workspace.
Observe one bounded job first, then extend it based on actual user requests. Demo examples
must show a verifiable outcome in a labeled example account, with explicit simulation labels
where appropriate. Metrics distinguish recurring use, one-off task completion and conversion.
