SME finance · AI agents
Khata Agent
An AI bookkeeper that lives in WhatsApp and speaks like the shopkeeper does.
Shopkeepers send a voice note, a photo of a receipt or a forwarded mobile-money SMS. The agent records the sale, tracks who owes what, reminds customers to pay, and turns a year of messy records into a clean ledger that lenders and tax offices accept.
Launch market: Bangladesh · 5 cited sources · Facts reviewed 14 October 2026
- Cash to MVP (est.)Our estimate of the cash needed to reach a working MVP: one-time setup plus monthly running costs for the build period. It excludes founder salaries.
- $2,280
- MVP timelineMinimum viable product: the smallest version real users will use and pay for, built to learn fast.
- 6 weeks
- Margin after AI (est.)Of every 100 you charge, how much is left after paying for the AI models and infrastructure behind it. Higher means a healthier business.
- ~70%
- Editorial scorePlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- 4.5 / 5
At a glance
- Who uses it
- Owners of small shops and the wholesalers who supply them.
- Who pays
- Shop owners, by a low monthly fee; lenders pay per referred loan.
- Who should build it
- A full-stack engineer plus someone who has run, or sold to, small shops.
- The hardest part
- Turning messy voice notes and receipt photos into ledger entries owners trust.
01
Opportunity
Small businesses run on memory, paper notebooks and mobile-money messages. They cannot see their own cash flow, chase debts politely or prove income to a lender, so they stay small and borrow at the worst rates. Bookkeeping apps exist, but they demand typing, forms and time that a one-person shop does not have.
The evidence
90%
Globalof businesses worldwide are SMEs, and they provide more than half of all employment.
16% vs 29%
Globalof working-age people use generative AI in the Global South versus the Global North. The tools have not reached the people who run most businesses.
40%
Globalof enterprise apps are forecast to include task-specific AI agents by the end of 2026, up from under 5% in 2025.
239M
Launch marketregistered mobile-money accounts in Bangladesh by January 2025 — the transaction trail the agent reads.
Why now
- Models now understand voice notes, handwriting and mixed Bangla–English reliably enough to keep books from casual input.
- Mobile money is everywhere, so most transactions already leave a digital trace the agent can read.
- Agents can now take actions — send a reminder, draft an invoice, file a summary — not just answer questions.
- Lenders want alternative data on small businesses; a clean, continuous ledger is exactly that.
02
The AI case
What AI makes possible here, why the big labs won't absorb it, and whether the margins hold.
AI leverage — what is newly possible
- Turns a voice note like “Karim bhai took 2 bags of rice, will pay Friday” into a structured sale and a due-date reminder.
- Reads receipt photos and forwarded mobile-money SMS without the owner typing anything.
- Sends polite, personalised payment reminders in the customer's language and logs replies.
- Answers “how much did I make this month?” in plain words, with the numbers behind it.
Lab-proof moatsA reason competitors, including big AI companies, can't easily copy you: data, distribution, trust or local knowledge.
The ledger itself
Months of a shop's sales, credit and repayment history. Switching means losing the record the lender trusts.
Local rails and rules
Integrations with local mobile-money formats, VAT and trade-licence paperwork. Global tools rarely go this deep.
Credit data network
Aggregated, consented repayment behaviour becomes a credit signal that lenders pay for — a second business no general assistant has.
Inference economicsThe cost of running AI models each time a customer uses the product. It decides whether an AI business makes money. (estimates)
- Unit
- Active shop per month
- AI cost per unit
- ~$0.30 model + ~$0.40 messaging
- Price per unit
- ~$2.50
- Gross margin
- ~70%
How the margin is protected
- →Small models handle most parsing; escalate only ambiguous entries.
- →Batch reminders and summaries instead of real-time calls.
- →Lender referral fees subsidise the shop price.
03
Market & expansion
Launch in Bangladesh. Then expand.
- Start with one dense trade cluster — groceries, pharmacies or wholesale in one city — where word of mouth travels fast.
- Owners pay for money they can see: recovered debts and loans unlocked sell better than “bookkeeping”.
- Organisations are adopting agents fast but few have scaled them in any one function; small, focused agents win first.
- Validate: how much credit a typical shop extends, and how much it loses to unpaid debts each month.
Editorial scorecardPlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- AI leverage
- 5/5
- Moat durability
- 4/5
- Competition gap
- 3/5
- Capital efficiency
- 5/5
- Timing
- 5/5
- Scalability
- 5/5
Where it travels next
India & Pakistan
The same credit-based kirana and dukan economy, the same voice-first owners.
Indonesia & Philippines
Millions of sari-sari and warung shops running on WhatsApp and mobile wallets.
East & West Africa
Mobile-money-first economies with huge informal retail and thin credit data.
04
Competition & edge
Who already plays here, what they leave open, and what could go wrong.
Digital ledger apps (e.g. TallyKhata, Khatabook)
Apps for recording credit sales and sending reminders.
GapThey need the owner to open an app and type; they record, but do not act.
Accounting software (e.g. Zoho Books, QuickBooks)
Full accounting for formal businesses.
GapBuilt for accountants and forms, priced for the Global North, and not voice-first.
Messaging platforms' own business AI
Generic AI replies inside business messaging.
GapAnswers customers but keeps no ledger, knows no local tax rules and builds no credit history.
Where the edge comes from
- Zero new app: it works inside the chat the owner already uses all day.
- Speaks the owner's language — including the mix of Bangla, English and trade slang.
- Pays for itself through recovered debts, which is a sale owners understand instantly.
Risks and how to handle them
Wrong entries erode trust in the books.
Confirm each parsed entry with a one-tap reply; flag low-confidence items for review.
Messaging platform policy or pricing changes.
Keep a lightweight web ledger and SMS fallback; never depend on one channel.
Handling financial data invites regulation.
Consent-first data sharing, local data residency where required, and partner with licensed lenders rather than lending directly.
05
Business model
Shop subscription
Low monthly fee after a free month; priced below one unpaid debt.
Lender referrals
Fee per loan originated from consented ledger data.
Supplier tools
Wholesalers pay to manage credit across their retailer network.
06
Stack & data
Modern, proven tools a small team can run without a DevOps hire.
Channel
WhatsApp Business Cloud API + SMS fallback
Meet owners in the app they already live in.
Agent
Claude API with tool use — small model for parsing, large model for reasoning and summaries
Reliable structured extraction and actions at low cost.
Speech
Speech-to-text tuned for Bangla voice notes
Most input arrives as voice.
Web ledger
Next.js 16 + Tailwind CSS v4
A clean ledger view and lender export, from the same codebase as the API.
Database
PostgreSQL (Neon) + Drizzle ORM
Double-entry ledgers need relational integrity and audit trails.
Jobs
Inngest
Scheduled reminders, retries and monthly summaries without managing queues.
Evals
Golden set of real (anonymised) messages with expected ledger entries
Parsing accuracy is the product; test every prompt change.
Hosting & ops
Vercel + Sentry + PostHog
Simple, observable operations.
Data model6 core tables, for the engineer starting the first migration.
shops
A business and its owner.
- id
- owner_phone
- name
- trade
- district
- plan
contacts
Customers and suppliers.
- id
- shop_id
- name
- phone
- type
entries
Ledger lines (double-entry).
- id
- shop_id
- contact_id
- amount
- direction
- category
- occurred_at
- confidence
inbound_messages
Raw input before parsing.
- id
- shop_id
- kind (voice/photo/sms/text)
- media_url
- transcript
- parsed_entry_id
reminders
Payment follow-ups.
- id
- entry_id
- send_at
- status
- reply
consents
Data sharing with lenders.
- id
- shop_id
- lender_id
- scope
- granted_at
- revoked_at
07
Costing
Directional monthly estimates in USD. Your numbers will differ — the structure is what matters.
| Item | MVP / monthWhat the product costs to run each month (hosting, AI models, tools) at MVP stage. | Scale / month |
|---|---|---|
| Hosting (Vercel Pro) | $20 | $200 |
| Database (Neon) | $20 | $200 |
| Model inference | $60 | $1,500 |
| Speech-to-text | $30 | $600 |
| WhatsApp + SMSPer-conversation pricing | $60 | $2,000 |
| Field onboarding (part-time) | $250 | $2,000 |
| Total | $440 | $6,500 |
One-time setup
08
MVP & roadmap
Shippable in about 6 weeks, then a path to production.
Build
- Record sales and credit from voice notes, photos and text in WhatsApp.
- One-tap confirmation of each entry.
- Automatic payment reminders to customers.
- Daily and monthly summaries on request.
- Simple web ledger with export.
Not yet
- Lender integrations
- Inventory
- Tax filing
- Multi-staff shops
Roadmap to production
Validate
Weeks 0–3- Keep books by hand over WhatsApp for 15 shops
- Measure debts recovered
- Test willingness to pay
MVP
Weeks 4–10- Ship the agent
- 300 active shops
- Parsing accuracy above 97%
Traction
Months 3–8- First lender partnership
- Supplier plan
- Second city
Production
Months 9–18- Inventory and tax summaries
- Second country
- Seed round
09
Start this week
Your first steps. Tick them off as you go.
Metrics that prove it is working
- Weekly active shops and entries per shop
- Debt recovered per shop per month
- Entry accuracy (confirmed without edits)
- AI and messaging cost per active shop
10
Sources
Every figure on this page links to its source. Facts reviewed 14 October 2026. Costs and market notes without a citation are our estimates.
- 1.World Bank — SME finance and job creation: insights from a new report (2022)
- 2.Microsoft — The continued state of global AI diffusion in 2026 (2026)
- 3.Gartner — 40% of enterprise apps will feature task-specific AI agents by 2026 (2025)
- 4.The Financial Express (Bangladesh Bank data) — Bangladesh MFS accounts surge by 20 million in a year (2025)
- 5.McKinsey — The state of AI in 2025: Agents, innovation, and transformation (2025)