Book a discovery call

What a project looks like, before you sign anything.

Six worked examples of the problems we take on, how we would approach them and how we would measure success.

Every scenario on this page is illustrative. None of them is a client, and none of them describes a result. Envisstack is a new company and we will not show you client work we have not done. These examples show how we think: the questions we ask, the metric we would agree with you in writing, and the risks we would look for early. Targets shown are the kind we would propose, not outcomes we have achieved.

Six illustrative engagements

Illustrative

An NBFC’s collections calls in Marathi and Hindi

Timeline
about 10 to 12 weeks to go-live
Indicative budget
₹10 lakh to ₹40 lakh, excluding GST, plus pass-through telephony and model costs

The situation

A vehicle and small-business lender in Maharashtra makes several thousand early-bucket reminder calls a month. Calls run from a script in Marathi and Hindi. Outcomes are typed into a spreadsheet at the end of each shift.

The problem

Callers spend most of their time on short, routine calls and too little on the hard conversations. Promise-to-pay data is incomplete, so follow-ups are missed.

Our approach

  1. Discover

    Sit with the calling team for two days, listen to recorded calls, map the dispositions and agree with compliance which calls are in scope (reminders before and just after the due date, not hardship or dispute cases).

  2. Prove

    A 6-week pilot on a sample of live accounts with the agent disclosing that it is automated, offering a key press to reach a person, and writing outcomes to the LMS.

  3. Build

    Integration with the dialler and LMS, call-time rules, consent checks, redaction of account details in logs, a dashboard for the collections head.

  4. Run

    Weekly review of a sample of call recordings and dispositions against the evaluation set.

Success metric

Target agreed in writing: share of in-scope calls completed without a person, with disposition accuracy of at least the level agreed on a reviewed sample, and zero calls outside permitted hours.

What could make it fail

Poor call audio, customers dropping on hearing an automated voice, or LMS data that is out of date. We would see all three in the first two weeks of the pilot from completion and drop-off rates.

Illustrative

A diagnostics chain answering report-status questions on WhatsApp

Industry
Healthcare
Timeline
about 8 to 10 weeks to go-live
Indicative budget
₹8 lakh to ₹35 lakh, excluding GST, plus pass-through WhatsApp and model costs

The situation

A diagnostics chain with collection centres across two states gets most of its customer messages on WhatsApp. The same questions repeat: is my report ready, when can someone collect a sample from home, do I need to fast.

The problem

The support team is stretched at peak hours, response times slip, and staff answer from memory, so instructions are not always consistent.

Our approach

  1. Discover

    Read three months of exported chats, cluster the questions and agree which are operational (in scope) and which are clinical (always to a person).

  2. Prove

    A 6-week pilot on one city’s WhatsApp number, connected read-only to the LIS for report status and to the booking system for slots.

  3. Build

    Booking writes, escalation rules for urgent words, Hindi and English with Hinglish tested, an audit log, and hand-off into the existing helpdesk.

  4. Run

    Weekly sample review and monthly update of approved answers.

Success metric

Target agreed in writing: share of in-scope conversations resolved without a person, answer accuracy on a reviewed sample, and every clinical or urgent message handed to a person within the agreed time.

What could make it fail

LIS status fields that lag reality, or customers asking clinical questions more often than the chat history suggested. Both show up in week one of the pilot.

Illustrative

A general insurer reading health-claim documents

Timeline
about 8 to 10 weeks to go-live
Indicative budget
₹6 lakh to ₹30 lakh, excluding GST, plus pass-through model and hosting costs

The situation

A general insurer’s health claims team receives reimbursement files as scans and phone photos: claim form, discharge summary, hospital bills, pharmacy bills, investigation reports.

The problem

Assessors spend much of their time finding and keying information. Missing documents are often noticed days after the file arrives, which delays the customer and creates repeat requests.

Our approach

  1. Discover

    Review a sample of anonymised files with assessors and agree the fields that matter, the document checklist per claim type, and what “correct” means for each field.

  2. Prove

    A 6-week pilot on historical files with known correct values, deployed in the insurer’s environment in an Indian region.

  3. Build

    Integration with the claims system, a review screen where assessors confirm low-confidence fields, a missing-document check on arrival, and access controls aligned to the insurer’s security guidelines.

  4. Run

    Weekly field-level accuracy checks and retraining of extraction rules as hospitals change formats.

Success metric

Target agreed in writing: field-level accuracy on the agreed evaluation set, share of files with a correct missing-document flag on day one, and assessor time per file. The claim decision stays with the assessor.

What could make it fail

Very poor image quality, or handwritten bills from small hospitals. We would measure accuracy by document type in the first pilot week and narrow scope if needed.

Illustrative

A D2C brand’s marketplace and GST reconciliation

Timeline
about 5 to 8 weeks to go-live
Indicative budget
₹4 lakh to ₹20 lakh, excluding GST, plus pass-through model costs

The situation

A personal-care brand sells on its own site and three marketplaces. Each marketplace sends settlement reports in a different format, with commissions, fees, returns and TCS deducted.

The problem

The finance team spends the first week of every month matching settlements to orders, invoices and gateway reports in spreadsheets. Unreconciled amounts carry forward, and GST and TCS credits are hard to trace.

Our approach

  1. Discover

    Walk through one full month’s reconciliation with the finance team and their CA, and agree the matching rules and the exceptions a person must decide.

  2. Prove

    A 4-week pilot rerunning two past months, where the correct answer is already known.

  3. Build

    Scheduled ingestion of settlement and gateway files, matching against Tally or the ERP, a weekly exceptions list, and an audit trail of every match.

  4. Run

    Monthly review of exceptions and rule changes when marketplaces change their report formats.

Success metric

Target agreed in writing: share of settlement lines matched automatically with zero incorrect matches on the reviewed sample, and days taken to close the month’s reconciliation.

What could make it fail

Missing order IDs in marketplace reports, or rules that live only in one person’s head. Both surface during discovery, before any build.

Illustrative

A manufacturer’s SOP assistant for shop-floor supervisors in Hindi

Timeline
about 6 to 8 weeks to go-live
Indicative budget
₹5 lakh to ₹20 lakh, excluding GST, plus pass-through model and hosting costs

The situation

An auto-components manufacturer runs two plants. SOPs, work instructions and machine manuals are in English, and many supervisors are more comfortable in Hindi.

The problem

On night shifts, supervisors cannot find the right instruction quickly, so they call someone or work from memory. Quality and safety teams worry about both.

Our approach

  1. Discover

    Spend a shift on the floor, collect the 50 questions supervisors ask most, and agree with the quality head which documents are the controlled source.

  2. Prove

    A 5-week pilot on one line, answering in Hindi or English with the SOP and page cited, and saying “I do not know” when the answer is not in the documents.

  3. Build

    Document control integration so only current revisions are used, access on shared tablets and phones, usage logs for the quality team.

  4. Run

    Monthly review of unanswered and corrected questions, and reloading of documents on every revision.

Success metric

Target agreed in writing: answer accuracy and citation accuracy on an evaluation set approved by the quality head, and zero answers given from outside the approved documents.

What could make it fail

Scanned or outdated SOPs, or supervisors not trusting the tool. We would test document quality in week one and track daily usage from the start of the pilot.

Illustrative

A logistics company’s leaders asking questions of their own data

Timeline
about 8 to 10 weeks to go-live
Indicative budget
₹8 lakh to ₹30 lakh, excluding GST, plus pass-through model and hosting costs

The situation

A part-truckload logistics operator has data in a TMS, a warehouse system and Tally. Leadership reviews run on slides an analyst builds over two days.

The problem

Leaders cannot ask a follow-up question without waiting for another report. Numbers in different reports do not always agree.

Our approach

  1. Discover

    Collect the 40 questions leaders actually ask, and agree one definition for each metric, such as on-time delivery and freight cost per tonne.

  2. Prove

    A 6-week pilot on a read-only copy of the data, where the agent answers in plain language and shows the query and source tables behind every number.

  3. Build

    A governed metrics layer, role-based access, and the agent inside the tool leaders already use.

  4. Run

    Weekly checks of answers against the analyst’s numbers and new questions added to the evaluation set.

Success metric

Target agreed in writing: share of evaluation questions answered correctly, matching the agreed metric definitions, and the agent declining questions it cannot answer reliably rather than guessing.

What could make it fail

Inconsistent data across systems. That is often the real problem, and discovery will show it before we build anything.

How we pick the success metric

One number the business cares about

Calls completed, files processed, days to close. Not “AI accuracy” on its own.

One quality bar

Measured on an evaluation set built from your real data and checked by your people.

One safety line

The thing that must never happen, such as a clinical answer, a call outside permitted hours or an incorrect match.

See how we work

Your problem is probably not on this page. Tell us about it.

Thirty minutes with a founder. A one-page note within three working days.

Book a discovery callSee how we work