Internal agents for jewellery and D2C operations: where the hours actually go.

Look at a jewellery or fine-goods D2C business from the outside and it is a product page, a checkout, and a campaign calendar. Look at it from the owner's chair and it is something else: a back office that absorbs more hours than the storefront, most of them spent rebuilding numbers that already exist somewhere. This post is about that back office, and where an internal agent belongs in it.

The shape is specific to the sector. The catalog is configurable. One ring is a metal, a purity, a stone, a size, and an engraving, and each combination is its own SKU in the ERP and its own listing on the site. Much of the order book is made-to-order. An order is a job with a karigar or a workshop, a due date, and a chance of slipping. Returns and exchanges carry rules. Resizing, hallmark checks, buyback rates against the day's metal price. Stock lives in four places. The site, one or two marketplaces, the stores, and the warehouse, and each holds its own count.

Then the supplier side. Slips arrive on paper or as a photograph on WhatsApp, often in a second language, keyed in by hand, and one rushed entry breaks reconciliation weeks later. Purchase orders, goods receipts, and invoices are matched by someone who also has three other jobs.

And over all of it, the cycle. A festival season or a campaign window compresses a quarter's volume into three weeks. Every process that is manageable in a quiet month breaks in a loud one, at exactly the moment the owner has the least time to notice.

Which brings the owner. Every morning, before the first store opens, someone rebuilds yesterday: revenue by channel, what moved, what did not, what is stuck in production, what a customer is waiting on. It comes from a ring-around, three exports, and a spreadsheet. The picture lands late enough that half of it is history by the time it is read.

Hubzoid builds agents in three shapes, and the sector uses all three.

01. Tool. A tool is pressed for an outcome. Supplier-slip verification is the one that pays first: photograph a slip, it is read, translated, checked against your purchase records, and staged for approval, with mismatches flagged by line. The invoice-to-PO check is the same shape one step later: every supplier invoice matched against its order and its goods receipt, exceptions listed with the reason. A return eligibility check reads the order, the policy, and today's metal rate and hands the desk a decision to confirm rather than a rule to remember.

02. Q&A. A question in plain language, answered against live data with the source shown. Which SKUs sold in the last seven days across every channel, and which of them are below reorder. Which made-to-order jobs are past their promised date. How this campaign is running against the last one at the same point. What one store did yesterday against the same day last year. Each of those is a call, an export, or a wait today. Each becomes a message in Telegram or Slack that comes back in seconds.

03. Background. No one asks. The agent runs on a schedule and lands the answer before anyone thinks to look. The owner's briefing, assembled overnight and delivered before the first store opens: revenue, movers, anomalies, jobs at risk, and the one thing that needs a decision today. Overnight stock reconciliation across site, marketplace, store, and warehouse, published as a list of differences rather than a spreadsheet to build. A supplier ledger digest of what was received, what was invoiced, and what does not agree.

During a festival or campaign window the Background agents matter most. The briefing runs at the same time whether the day was quiet or the busiest of the year. The reconciliation runs at the same depth. The team's attention goes to customers and the floor, and the numbers keep arriving anyway.

The first month looks like this. Discovery maps how the business actually decides: who approves a return above a threshold, who signs off a slip, what counts as an anomaly worth waking the owner for. A catalog of candidate agents comes out of it, larger than what will be built. One agent ships first, usually the owner's briefing or the slip check, on the surface the team already uses, with tools gated by team group and every allow or deny recorded. By the end of the month the owner reads a briefing that took no one's morning to build, and the number of hours it recovered is written down next to the number that was written down before.

Now the point that separates an agent that works in this sector from one that fails quietly. The dangerous bug is silent. An export where a stock count reads zero, not because the shelf is empty but because the sync failed at two in the morning. A per-line total that gets counted once per line and once again for the order, so revenue reads high and nobody notices until the accountant does. A model handed either of those will summarise it confidently, because summarising confidently is what models do. An agent built for this business has to carry deterministic checks in code: a count that dropped to zero from a healthy figure overnight is a question, not a fact. A total that does not match the sum of its lines does not get published. The agent flags, holds, and asks. Policy on code, judgment on the model, and the checks written down where a reviewer can read them.

None of this touches a customer. No shopper talks to any of these agents. The user is the owner, the accounts desk, the store manager, the person who reconciles stock. That is the scope by design: the work that already has a number attached, done by the people who already know what right looks like.

The measure is hours. Count what the briefing, the slip entry, the reconciliation, and the exception-chasing absorb each week, and who absorbs it. Multiply by what those people cost. That is the figure an internal agent is priced against, and the figure it is judged on a quarter later.