What we build
AI agents
Agents that carry out routine work, with a person approving the exceptions. Agents that read, decide and act across your systems, such as booking appointments on WhatsApp or matching invoices to purchase orders. Each one is accepted against test cases and approval rules you agree in advance.
The situation
The same steps, done by hand, every day
In most offices a few people spend their day on work that follows clear rules: reading a WhatsApp message and booking a slot, checking a supplier invoice against the purchase order, chasing a customer for a missing document. The rules are known, yet the work still needs someone to open three screens and type. An agent can do much of this if the job is narrow and well defined. The hard parts are deciding exactly what it may do on its own, what it must hand to a person, and how you can see what it did and why.
Support reply assistant
Customer asks why order 4471 is late.
Courier shows a delay at the Nagpur hub. Draft reply ready with the new date. Please check before sending.
Looks right. Send it.
Reply sent by a person · logged
What we build
AI agents, in practice.
WhatsApp and chat agents
Agents on the WhatsApp Business API or your website that answer questions, book slots and take orders, handing over to staff when needed.
Back office agents
Agents that match invoices to purchase orders, check delivery notes and prepare entries for review before anything reaches Tally.
Tool use and system actions
Agents that call your own APIs to look up records, create bookings or update a CRM, limited to the actions you have approved.
Approval steps and handover
Clear points where a person approves, edits or rejects what the agent proposes, with the reasons recorded.
Logs, monitoring and audit trails
A record of every message, decision and action, so your team can review what happened and correct it.
Evaluation and guardrails
Test sets drawn from your real cases, limits on cost and actions, and agreed behaviour when the agent is unsure.
Typical sizes
Roughly how many hours.
Planning sizes only. The real number comes from a signed scope after the first conversation.
150TO200HOURS
One agent for one narrow job
A single agent such as appointment booking on WhatsApp, with an approval step and logs. Roughly INR 29,850 to INR 39,800 before tax at INR 199 an hour.
230TO450HOURS
An agent connected to several systems
An agent that reads and writes across your CRM, accounts and messaging, with roles, monitoring and an evaluation set. Roughly INR 45,770 to INR 89,550 before tax.
500TO1,000HOURS
A set of agents across a business process
Several agents sharing one model layer, approval screens and audit trail. Roughly INR 99,500 to INR 1,99,000 before tax.
An example
Illustrative: an Ahmedabad clinic chain books appointments on WhatsApp
Picture a chain of clinics in Ahmedabad where front desk staff answer hundreds of WhatsApp messages asking for appointments. The specification describes an agent on the WhatsApp Business API that understands requests in English, Gujarati and Hindi, offers open slots from the clinic calendar, confirms bookings and sends reminders. Anything about symptoms or prescriptions goes straight to a staff member. Acceptance is tested against a set of real, anonymised conversations with agreed correct outcomes, and message and model costs are agreed in writing before work starts.
Technology we often use
- Anthropic Claude
- Google Gemini
- OpenAI
- Python
- TypeScript
- LangGraph
- WhatsApp Business API
- PostgreSQL
- Redis
- AWS
Chosen to fit the project and your team, and written into the specification.
Questions
About ai agents.
What stops an agent doing something it should not?
The specification lists the actions it may take, the limits on each, and the points where a person must approve. The agent only gets access to the systems and records those actions need, and every action is logged.
Can the agent be built on a platform we can manage ourselves?
Yes. Agents can be built on AksharAI (aksharai.com), a Code Colonies platform for AI apps and agents, or as custom code in your own systems. The choice is made during scoping and written into the specification.
How is an agent accepted?
Against a test set of real cases agreed before the build, with a clear rule for what counts as a pass and how the agent behaves when it is unsure. You run the tests yourself before you accept.
Related
Related work and reading
Start with a conversation
Tell us what you want built. We will tell you plainly whether it fits.
The first conversation is free. You get a rough size in hours, the tier that suits the work, and what a first 150 hour block would cover.
Or get a ballpark proposal in two minutes