# 01 — Product vision

**Status:** Draft for team review. Written 2026-11-07; product scope updated 2026-11-12.

## 1. One sentence

Charla lets people and organizations create AI agents they can delegate work to through
conversation, with knowledge, memory, connected tools, permissions, approvals and human
supervision, accessible through the channels and applications suited to each job.

## 2. The platform thesis and regional adaptation

Conversation is an interface for using software and delegating work. An agent can serve an
individual, a team, a business's customers, or users of another software product. Charla runs
the shared operating layer around it: execution, channel delivery, memory, identity, tools,
approvals, handoffs, versions, evaluations, traces and billing.

We adopt Keiki's broad platform model. Personal assistance, learning, company knowledge,
internal operations, back-office work, customer service and sales are all product families.
Templates are starting configurations; reusable capabilities extend what agents can do.
The SDK, API, MCP and partner deployments let other products build on this infrastructure.
See [20-product-scope.md](20-product-scope.md) for jobs, dependencies and delivery stages.

WhatsApp is our first business messaging integration in the launch markets. Telegram and API
are core surfaces too. An agent's job, permissions and capabilities determine where it can
run; a channel's policies apply only to deployments using that channel.

Regional adaptation covers Spanish, Portuguese and Italian variants, local payments,
consent, privacy, data residency and channel-native interaction. These support the platform
across Latin America, Brazil, Spain, Portugal and Italy. See
[03-market-and-channels-latam.md](03-market-and-channels-latam.md).

The [blog review](research/keiki-blog-product-lessons.md) adds two adoption paths: build an
agent hosted by Charla, or connect an agent that a customer already runs. In either case,
conversation can expose the customer's product operations as well as answer questions. The
[validation plan](21-product-validation.md) defines complete personal, internal and embedded
journeys; infrastructure availability alone does not validate these products.

## 3. Who it is for

| Segment | Job to be done | What they buy | Suitable surfaces (subject to roadmap) |
|---|---|---|---|
| **Individuals** | Delegate personal organization, reminders, research and learning | Configurable assistants, memory, account connections, scheduled tasks | Telegram, API; email and voice later |
| **Teams and operators** | Retrieve company knowledge, process documents, manage billing and follow up work | Connected tools, per-user permissions, approvals, audit and human handoff | API, eligible Telegram conversations; Slack/Teams and email later |
| **Businesses serving customers** | Answer, qualify, book, collect payments and follow up | Template or custom agents connected to the business's systems | WhatsApp, Telegram, API; Instagram, email and voice later |
| **SaaS / fintech / marketplaces** | Embed agents for their users and customers | Partner deployments, SDK, identity, MCP, governed actions and external harness | API and channels selected per deployment |
| **Agencies and capability builders** | Build and operate agents for clients; distribute reusable functionality | Multi-org operations, templates, capabilities; white-label and marketplace later | Channels selected for each client's job |

SMB front desks remain an initial pilot cohort. This is a sequencing choice, not a boundary
on the product. Broader use cases ship when their required integrations, permissions and
evaluations are ready. Existing vertical templates do not establish completion of the catalog.

## 4. Principles (inherited from Keiki, kept)

1. **Describe once, run everywhere.** Channel is deployment config, not agent config.
2. **The turn is the atomic unit.** Scheduling, credentials, tracing, billing, idempotency and
   cancellation hang off a turn id.
3. **The platform owns ingress, delivery and credentials.** Tool servers and external
   harnesses get short-lived turn-bound credentials; the model never sees secrets.
4. **Fail closed, visibly.** Identity webhook down → scoped tools refused and the model says
   so. Never default-allow, never silent drop.
5. **Governed actions are records, not prompts.** Args hashed at proposal; approval replays the
   stored call; freshness re-check supersedes; the agent reads action state every turn.
6. **Trust the result, not the intent.** Tools return what actually happened; the agent never
   claims a carousel, payment or booking succeeded without confirmation.
7. **Evals and replay before change.** A prompt change is tested against saved cases and real
   tapes; replay is a veto.
8. **Agent-first developer experience.** `llms.txt`, Markdown docs, OAuth MCP server, idempotent
   `deploy_agent` by `external_ref`, paste-in prompts for coding agents.

## 5. Principles added for this market

9. **Channel-native delivery.** Keep job instructions independent of channel configuration.
   On WhatsApp, use buttons, lists, Flows, templates, voice notes and supported payments.
   On other surfaces, use their own capabilities and constraints.
10. **WhatsApp's 24-hour window is a first-class channel state.** The agent knows whether
    it is open and chooses free-form vs template. On WhatsApp, alerts and notices outside
    the window go as templates.
11. **Spanish, Portuguese and Italian are not translations.** Variant-aware prompts (es-MX,
    es-AR, es-CO, es-CL, es-PE, es-419, es-ES, pt-BR, pt-PT, it-IT), locale formatting, STT tuned on real voice notes, judge models and mood classifier
    validated in the supported languages.
12. **Compliance is a product feature.** AI disclosure on first contact, consent ledger with
    local keywords (BAJA, CANCELAR, PARAR, SAIR, STOP), data residency choice, DSAR tooling,
    configurable retention.
13. **Teams work where they already are.** Handoffs to the dashboard, to Slack, *and* to a
    second WhatsApp number or group for the team.
14. **Local money.** Multi-currency from day one (MXN, BRL, COP, ARS, CLP, PEN, UYU, EUR, USD);
    Mercado Pago and Stripe; owner-approved limits and refunds.

## 6. Scope of the clone (parity targets)

Parity is measured against the Keiki docs snapshot of 2026-11-07
([02-competitive-teardown-keiki.md](02-competitive-teardown-keiki.md)). Three bands:

- **Core (must ship for GA):** cloud agents for the whole launch footprint (all of Latin
  America, Brazil, Spain, Portugal and Italy, decided 2026-11-07; locales es-MX, es-AR, es-CO,
  es-CL, es-PE, es-419, es-ES, pt-BR, pt-PT, it-IT with depth tiers per
  [03 §8](03-market-and-channels-latam.md)), WhatsApp (through a BSP) + Telegram + API
  channels, configurable agents for the supported jobs (not limited to front desks), tools
  (webhook, MCP, builtin), identity webhook and scopes, governed actions with text and button
  approvals, handoffs, memory, steering, media handling (voice notes!), traces, eval cases,
  event webhooks, developer REST API, OpenAI-compatible endpoint, dashboard, credits billing.
- **Parity (post-GA):** Instagram, Messenger, email, voice, partner deployments at scale,
  external harness + SDK, capabilities marketplace, replay, signals, BYOK, SSO, A2A, swarms,
  alerts/loops, agent wallets, Code Mode MCP.
- **Dropped:** iMessage lines, Keiki Vault iOS app and Face ID approvals, Apple Messages for
  Business, contact cards, Satori-rendered image carousels, US phone-line pools, Alchemy
  integration, Eve/Flue adapters (replace with a generic AI SDK adapter), AgentID sign-in.

## 6a. Platform policy constraint (Meta)

The WhatsApp policy assumptions and verification gaps are recorded in
[03 §2.3](03-market-and-channels-latam.md). Deployment eligibility must be checked against
current provider terms before launch. Personal assistants and other general-purpose agents
use eligible channels such as Telegram and the API. Every agent belongs to an organization
for isolation and billing; an organization can represent an individual, a team or a business.
That tenancy boundary does not make every agent a customer-service agent.

## 7. Non-goals (for the first year)

- Being a general chatbot builder with drag-and-drop flows. We are an agent runtime with tools
  and governance; flows are WhatsApp Flows, not our own designer.
- Building our own LLM or fine-tuning. Model-agnostic via the Mastra model router.
- Being a BSP ourselves for SMS/voice. We integrate Twilio and regional carriers.
- Replacing the tenant's CRM or helpdesk. We hand off to them and call them as tools.

## 8. Success criteria for the MVP

| Metric | Target |
|---|---|
| Time from signup to a working agent on one eligible core channel (WhatsApp, Telegram or API) | < 15 minutes; report results separately by channel |
| Product-family coverage | Validate a tool-backed job in each of personal assistance/learning, internal operations/knowledge and customer-facing work; record dependencies and eval results |
| Median first-reply latency (inbound to first outbound) | < 4 s free-form; < 8 s with one tool call |
| Voice-note transcription WER on our es/pt WhatsApp sample | < 10% |
| Eval pass rate on template suites across es-MX, es-AR, es-CO, es-CL, es-PE, es-ES, pt-BR, pt-PT, it-IT | ≥ 90% |
| Countries with at least one paying customer in the first quarter after GA | 12 of 22 |
| Unsupported-claim signal rate | < 1% of turns |
| WhatsApp pilot customers live on their own WABA (channel cohort) | 5 |
| Non-WhatsApp pilot coverage | At least one personal/learning pilot and one internal-work pilot on Telegram or API, with observed task outcomes |

Report task outcomes by job cadence. Recurring assistants need useful repeat usage; an
isolated reservation or support task can succeed without a return visit. Measure accepted
request to verified result, human intervention and cost per completion alongside turn metrics.
Definitions and pilot gates are in [21 §4](21-product-validation.md). Published Keiki customer
results inform hypotheses; they are not Charla benchmarks or promised outcomes.
