Years of Scattered SOPs: Why ₹35 Items Were Leaving at ₹5/kg

A worker on the sorting line picks up a small item from the mixed household stream. It looks like scrap. Into the ₹5 per kilogram pile it goes. Pulled and sold on its own, that same item clears around ₹35 a piece. Nobody was careless. The rule that separates those two piles sat in a manual in the supervisor's office, and the three people who knew it without looking were busy.

• Years of SOPs and grading documentation existed only on paper and in the heads of a few veterans
• High-value items in the mixed household stream left at ₹5/kg commodity rate instead of ₹35 per piece
• Floor workers interrupted supervisors constantly for lookups, so supervising became answering
• New hires took about four months to become useful, because the only way to learn was to ask someone

One missed call costs a few rupees. Repeated across a shift, on a stream that runs every day, it becomes the difference between a thin margin and a real one.

A private AI assistant, running on an isolated instance, loaded with the facility's own SOPs and grading documentation. Workers query it by voice, in their own language, from a phone on the floor.

It answers the questions the floor actually asks. Identification tests for material and grade. Processing routes by grade. What can and cannot be mixed. Packing and labeling requirements. And which items to pull for separate sale instead of sending out by weight.

The shape was chosen to fit the work, not the software. Voice matters because hands are busy and gloves are on. First language matters because a translated manual adds a step between the question and the answer. A phone matters because the decision happens at the pile, not at a desk.

The content is theirs. We did not write new procedures. We took the documentation the facility had already built over years and made it answerable in seconds.

What are the boundaries of the system?

It repeats the facility's procedures. It does not replace them. Every answer comes from the documentation the facility wrote, not from a model's general knowledge about scrap. It does not grade material and it does not decide anything.

Why does that boundary matter?

A floor team that expects to be replaced will not use a tool, and will not trust it. The veterans here hold real expertise, built over years of handling material. Nothing about this deployment competes with that. It handles lookups so their judgment goes to the calls that actually need it.

How is that boundary held in practice?

1. 1Answers cite the facility's own manual. Nothing is invented, and nothing arrives from outside their documentation.
2. 2A supervisor's call overrides the system. If a veteran disagrees with an answer, the veteran is right and the document gets corrected.
3. 3No procedure was rewritten during the build. The manual did not change, only how fast the floor can reach it.
4. 4Nothing is measured or scored about a worker. The system logs nothing about who asked what, because that was never the point.

What does an isolated instance mean here?

The deployment runs on an instance dedicated to this facility. Their documents go into that instance and stay there. Nothing is used to train anyone's model. Nothing moves to a shared service.

Why does this matter for a facility like this one?

Grading rules, processing routes, and mixing constraints are competitive information. A facility's margin sits inside exactly those documents. Handing them to a general-purpose tool with unclear retention terms is a real risk, not a theoretical one.

How do you check this on any deployment?

1. 1Ask where the instance runs. Dedicated or shared is the first question, not the last.
2. 2Get training exclusion in writing. A verbal assurance is not a contract term.
3. 3Ask for the export. If you cannot take the structured knowledge base with you, you do not own it.
4. 4Confirm the retention window. Know how long anything is held and who is able to read it.

What happens when the rules change?

The facility adds to the knowledge base whenever they want. A new grade, a revised buyer specification, or a changed packing rule gets loaded, and the assistant answers from it from that point on. No vendor ticket is involved.

Why was self-service updating a requirement?

The paper binders had drifted because updating them meant reprinting them. A manual that is hard to change becomes a manual nobody trusts. Buyer specifications move, and documentation that lags them is worse than no documentation at all.

How does an update reach the floor?

1. 1Someone with the authority loads the change. Usually a supervisor or the owner, not a developer.
2. 2The structured knowledge base absorbs it. The new rule is indexed alongside everything already there.
3. 3The next query answers from the new version. There is no release cycle and no waiting on us.
4. 4The export stays complete. Everything the facility adds remains theirs to take at any time.

The floor stopped queueing at the supervisor's desk, and the household stream stopped being a single commodity line:

The recovery on high-value items is the number the owner watched. It was also the easiest to verify, because it shows up directly in what the separated stock sells for.

The ramp-up change had a second effect nobody had priced in. The knowledge that used to live only in three people's heads now exists in a structured form the facility owns. That was true the day the system went live, and it stays true if those three ever leave.