Service · AI and ML engineering

AI assistants for operations

Answer the questions your staff answer all day, from your own documents, with guardrails.

Relevant to Service businesses, Education, Growing businesses

The problem

Staff answer the same questions all day: admissions queries, customer status, SOP lookups, HR policy.

The answers already exist in manuals, FAQs, past tickets and policies. An assistant grounded in those documents, with firm guardrails, takes the repetitive load.

What it costs today

Typical: two to four full-time equivalents per department answering questions a well-built assistant resolves, and slow responses that cost you enquiries.

This is for you if
  • A team that answers the same fifty questions, by phone, WhatsApp or email, all day.
  • The answers exist in documents you are willing to approve and keep current.
  • Someone on your side can review a sample of answers every week.

Probably not for you if

  • A general chatbot that talks about anything. We build assistants that answer from your documents and refuse the rest.
What we build
  • An assistant that answers from your approved documents only, and says when it does not know.
  • Guardrails: topic limits, prompt-injection protection, no comparisons or claims you have not approved.
  • Cost control: cheap models for routine queries, escalation for complex ones, caching for repeats.
  • Deployment on your website, WhatsApp or internal tools, with a review log.
How we build itIn this order
  1. 01

    Collect and approve the documents

    SOPs, policies, FAQs, brochures. Each one is versioned and signed off by its owner, because the assistant will only ever say what they say.

  2. 02

    Test retrieval before any model talks

    A few hundred real questions from past tickets, with the right answer and its source. We measure whether the right passage is found before we measure how well it is phrased.

  3. 03

    Set the guardrails and attack them

    Topic limits, refusal behaviour, injection attempts, requests for comparisons or promises. The adversarial test report is shared before launch.

  4. 04

    Supervised pilot on one channel

    Your reviewer reads a sample of conversations daily, and every correction goes back into the documents or the test set.

  5. 05

    Measure cost and deflection

    Cost per conversation, questions resolved without a person, and handover rate, reported monthly.

Data you already haveThis is usually enough to start
  • SOPs, manuals, policy documents
  • Past tickets, emails and chat transcripts
  • FAQ pages and brochures
In practice

Illustrative example

An admissions assistant that answers 70% of routine queries without a counsellor

Hostel fees, deadlines, eligibility, scholarships: the same few hundred questions, asked thousands of times a season, by phone, email and WhatsApp.

Read the case study
Typical engagement
Duration6 to 8 weeks to a supervised pilot
TeamEngineering lead, ML engineer, domain reviewer from your side
You receiveLive assistant on one channel, guardrail test report, cost per conversation
PriceFixed price, quoted after the readiness assessment
What you own at the end
  • The assistant on your accounts and channels, with the document pipeline.
  • The question set, the guardrail tests and the adversarial report.
  • The review log and the cost dashboard.

Everything runs in your environment and is handed over with documentation. Nothing stops working if we part ways.

Questions we are usually asked

Will it make things up?

It is restricted to answering from your documents and refuses outside that. We test it with adversarial questions before launch and share the report.

What does it cost to run?

Measured per conversation and reported monthly. Routing and caching usually keep it to a few rupees per conversation.

Can it work on WhatsApp?

Yes, through the WhatsApp Business API, which we have run at large scale.

Who reviews the answers?

Your domain reviewer, during the supervised pilot. After launch, a sampled log is reviewed on a schedule you set.

Can it take actions, like booking a visit?

Yes, as narrow, approved actions with a confirmation step. Once the assistant is taking several actions in a row, it is an agent, and we build it with a harness.

Often combined with

Not sure if your data is ready?

Take the free 20-minute readiness check. You get a short scorecard and a plain answer about what, if anything, is worth doing next.

Start the readiness checkTalk to us