Ship conversational banking on top of TwelveAI
A grounded, action-taking AI layer for fintechs and neobanks. Understand what a customer said, look up their real financial context, and — if they meant to — do something about it. One API, the guardrails of the Business platform underneath.
Waitlist gets early access + 90 days free at launch.
POST /api/business/v1/ai/converse
Authorization: Bearer tw_at_...
{
"customer_id": "cus_9f2a...",
"message": "send 20k to my sister ada at gtbank, then how much do i have left?"
}
# → {
# "reply": "Sent NGN 20,000 to ADA ADEBOLA · GTBank 0123456789.
# Your available balance is now NGN 84,250.",
# "actions_taken": [
# { "type": "transfer", "reference": "TWAI_PO_...", "status": "processing" }
# ],
# "citations": [
# { "type": "recipient", "resolved_from": "recent_recipients" },
# { "type": "balance", "as_of": "2026-07-29T09:14:12Z" }
# ]
# }Four primitives, one API
Use them together (the /converse endpoint above) or independently — parse intent, retrieve financial context, take an action, or generate a reply.
Understand
One endpoint parses a customer message and returns the intent — "check balance", "send NGN 20k to Ada at GTBank", "why did my card get declined last night?" — plus the entities it extracted.
Ground
Every answer is grounded in the account's real, live data — balances, transactions, cards, payouts — pulled through the same permissioned surface as the Business API. No hallucinated numbers.
Act
Turn intent into action. Initiate a transfer, resolve a NUBAN, freeze a card, or open a dispute — with the same scope + approval guardrails the rest of the platform enforces.
Speak
Generate a reply that's accurate, in the customer's language, and safe. Nigerian Pidgin, Yoruba, Igbo, Hausa on the roadmap.
Every fintech surface
If a customer can type or talk to you about money, this is the layer that answers. Same API for chat, in-app support, SMS, WhatsApp, and voice.
Neobanks & wallets
Replace your rules-based chatbot with a real assistant that can answer "how much did I spend on Uber last month?" — and then split the fare with a friend, in the same thread.
Business banking
Give SME customers a 24/7 finance officer: cash-flow summaries, invoice reminders, tax-set-aside nudges, payout status — all in one chat.
Compliance & support
Route customer messages to the right queue, auto-draft the first response with the relevant transaction facts already cited, and reduce ticket volume.
Voice & IVR
Ship a natural-voice caller experience that isn't a phone tree. Same models, same grounding, same actions — behind a phone number instead of a keyboard.
Safe by construction
A conversational banking API only works if it's trustworthy. Ours enforces the same limits the rest of the platform does.
Scoped
Every action runs against the connection's Connect scopes — a payments:read token can't initiate a transfer, no matter how the model interprets the message.
Grounded
Numbers, names, and statuses come from live TwelveAI data, not the model. Every reply carries its citations.
Auditable
Every message, tool call, and generated reply is logged with the customer_id and the connection — searchable in the dashboard and via the API.
Confirmable
High-value actions surface a confirmation step by default; you decide the threshold. The model can't silently move money.
Answers before you ask
Who is this for?
Fintechs, neobanks, wallets, POS operators, and treasury products building on TwelveAI. If you're already using our payments and Connect APIs, these are the AI layer on top of the same permissioned data plane.
How is this different from calling OpenAI myself?
Two things: (1) grounding — we plumb the customer's live TwelveAI financial data into every response, so answers cite real numbers not guesses. (2) action — the same call can initiate a transfer, resolve an account, or freeze a card, gated by the connection's scopes.
What about hallucinations?
Every response is grounded in retrieved facts (balances, transactions, KYC state) and returns the citations alongside the answer. If we don't know, we say so instead of making it up. Numbers never come from the model.
What models will you use?
We route per-task — the intent parser doesn't need Opus, the drafting layer sometimes does. You won't manage that; you send a message, we return a grounded, safe answer.
Nigerian language support?
Nigerian Pidgin at launch. Yoruba, Igbo, and Hausa on the roadmap. You can also just tell it "reply in Pidgin" per request.
How much will it cost?
Metered per successful message, with a free tier for building. Exact pricing at launch — waitlist gets 90 days free.
Be the first fintech shipping this
Waitlist members get early API access and 90 days free at launch. Tell us what you'd build.