Illustration: a plant’s paperwork, from a challan photo and a purchase order to a breakdown note and a customer’s WhatsApp message, travels along two converging lines into the full stop of the headline. One approved record leaves it.
Yucti Works is factory workflow software and supervised AI agents for owner-led plants in Waluj, Chikalthana, Shendra and AURIC. It is a work in progress: we are listening to plant owners before we build anything to sell.
Most plants already run on Tally, Excel, registers and WhatsApp. What costs the time is the chasing and re-typing between them. Yucti Works is light software and supervised AI agents on top of the systems a plant already uses.
Owner-led manufacturers with roughly 20 to 250 people, on Tally or a basic ERP, with Excel, paper and WhatsApp around it.
What we hear so far. Each one is a signal to test in interviews, not a finding.
Gate register · page 118
WhatsApp · Supplier C
Purchase voucherNo. 3187
Typed again from the register, 19:40
Illustrations on synthetic data. Photographs from real floors will replace them only with written consent.
A proposed first set from desk research: work one builder can deliver well, with no hardware and no machine connections. Nothing here is for sale yet. Plant-owner interviews decide what stays.
Supplier paperwork
An agent reads invoices, challans, purchase orders and test certificates, matches them to the PO and GRN, and drafts the entry. Your accountant approves every one.
Gate register · page 118
12 drafts ready. 2 need your review: one quantity, one new supplier.
Example workflow, synthetic
Dispatch status
Approved WhatsApp templates remind the team before loading, tell customers where their order is and pass every delay to a person.
Dispatch team
Yucti Dispatch Agent: the vehicle for order 5512 is due to load at 16:30 today. Please confirm with 1 (on time) or 2 (delayed).
1
Yucti Dispatch Agent: thank you. Customer updated: order 5512 loads at 16:30.
Example messages, synthetic. Agent messages always begin with the agent’s name.
Breakdowns
A breakdown reported on WhatsApp, by text or voice, becomes a ticket. Preventive maintenance comes up as a reminder. The maintenance head assigns and closes each one.
| When | Machine | What happened | State |
|---|---|---|---|
| 07:52 | Press 3 | Breakdown, voice note from the floor | Assigned |
| 09:10 | Press 3 | Hydraulic hose replaced | Closed |
| Tomorrow | CNC 2 | PM due: coolant and way oil | Reminder set |
Example log, synthetic
Proposed in September 2026 from desk research. Reviewed after 12 to 15 plant-owner interviews.
Proposed first
Later, with care
Not for us
Step through one check: a challan that doesn’t match its purchase order. The agent finds the difference and stops. What happens next is yours.
The agent reads challan 2231 from a WhatsApp photo and finds its purchase order, PO 4471, in Tally.
Interactive illustration on synthetic data. Our agents are in development: this demo does not read documents or change any records.
The method we would bring to a plant: start with one workflow, measure it, and grow only on a result you can check.
Diagnose then pilot then prove then scale
One workflow, mapped on your floor, with a baseline in writing.
Fix that one workflow on top of your Tally, registers and WhatsApp.
Measure again against the baseline, at 30 and 90 days.
Roll out what works, one workflow at a time, with local support.
Many workflows. One method.
Each workflow is its own junction. Fixed one at a time, they add up to how a plant runs.
One workflow, mapped on your floor, with a baseline in writing.
Fix that one workflow on top of your Tally, registers and WhatsApp.
Measure again against the baseline, at 30 and 90 days.
Roll out what works, one workflow at a time, with local support.
Many workflows. One method.
Each workflow is its own junction. Fixed one at a time, they add up to how a plant runs.
Every agent would be supervised: one job, your approval before anything important, and a full log.
Every agent has a written job, named data sources and a named owner in your team.
Anything that changes records or messages outsiders needs approval, unless you pre-approve a template.
See what the agent read, drafted and did, and when.
Least-privilege access; no confidential data in public AI tools without your written permission.
If an agent doesn’t save time against the baseline, we turn it off.
We would rather show you where we are than where we hope to be.
Stage as of October 2026.
No pitch. We are learning how plants here handle dispatch, supplier paperwork and breakdowns, in Marathi, Hindi or English. Thirty minutes on your floor helps us build the right thing.
Also from Yucti: Yucti Studio, websites, apps and AI automation for local businesses.