NORQ AI

About this series. Four pieces on what AI agents can and cannot automate in retail. Two of the jobs below are systems we build and operate ourselves, and they are marked as such — the rest describe what the technology does reliably today. Where we quote a number, it is measured on our own systems.

Most writing about AI agents explains what an agent is. A retailer deciding whether to spend money needs something different: a list of work that currently costs you people, with an honest note on what each one replaces, what it needs from you first, and how you would prove it works before trusting it.

Here are nine, sorted by how much can go wrong rather than by how impressive they sound.

Nine retail jobs grouped into three tiers: agents that read and tell someone, agents that change something, and agents that watch and raise a human. — Grouped by what the agent is allowed to do. The first group can be scored against known answers before a customer ever sees it.

It reads and tells someone

Nothing in your business changes as a result, so the worst case is a wrong answer — and a wrong answer is testable before launch. This is where to spend first.

1. Answering customer questions readswe run this

On the website and on WhatsApp: where is my order, what does the returns policy say, does this run small, do you have it in store.

Replaces: first-line tickets and the out-of-hours gap.
Needs from you: your policies and product information in one place the agent may read, and a route to a human.
How you prove it: a set of real questions with known answers, scored before launch and on every change. Ours refuses rather than guesses when it finds nothing relevant — and that refusal path never calls the model at all.

2. Answering the phone when nobody can readswe run this

The call at 17:55, or while the one person on the shop floor is serving someone. It answers from the same content as the website agent, takes a callback number when it cannot help, and can text a link if the caller agrees out loud.

Replaces: the missed call, which leaves no trace in any dashboard and is the one channel where failure is invisible.
Needs from you: call forwarding, and a decision about when it picks up.
Measured on ours: 2.26 seconds from the caller finishing to the first word back, with speech and model on EU endpoints.

3. Asking your own data a question readswe run this

“Which categories are declining?” “What share of first-time buyers come back?” Answered in about twenty seconds, with a chart and the query behind every figure.

Replaces: the BI request queue — and more importantly the questions nobody bothers to ask because the answer would arrive after the decision.
Needs from you: agreed definitions of your own metrics. This is the real prerequisite and most retailers have never written them down.
Measured on ours: one right answer in ten became nine in ten, with no change of model. 85% on a question set it had never been tuned against.

4. Writing product copy at catalogue scale reads

Descriptions, category pages, size and care notes, translations, variants for each marketplace — drafted from your attributes, published by a person.

Replaces: agency copy fees and the backlog that leaves new lines listed with one line of text.
Needs from you: decent attribute data, a tone-of-voice sample, and someone who approves. Keep the agent out of the live catalogue.
How you prove it: have a merchandiser grade fifty drafts blind against fifty human ones.

5. Telling you what your reviews say reads

What customers actually report about a line, a store or a delivery partner, summarised by theme and tracked over time, instead of a star rating nobody can act on.

Replaces: nothing — which is the point. Most retailers read none of it beyond the one-star reviews.
Needs from you: access to the review text, and a person who owns the resulting actions.
How you prove it: check its themes against a week of reviews read by hand.

It changes something

Now a mistake is an action rather than a sentence, and the architecture has to carry the safety: the agent should pick from options your systems supply, never invent the values. Each of these wants an approval threshold and a reversible path.

6. Cleaning up the catalogue changes

Attributes and categories extracted from messy supplier feeds; duplicates matched; missing fields filled from the product text you already have.

Replaces: manual keying, and the silent revenue loss from products that cannot be filtered or found.
Needs from you: a labelled sample to measure against, and a rule for what goes live automatically versus into a review queue.
Worth knowing: in Google's own public store data we found one sweatshirt carrying nine different product IDs. Catalogue mess is normal, and it quietly breaks search, recommendations and reporting alike.

7. Processing supplier documents changes

Invoices, order confirmations and delivery notes read into structured records, with discrepancies against the purchase order flagged to a person.

Replaces: manual keying in finance and goods-in.
Needs from you: a sample of each document type, and a threshold above which a human always checks.
How you prove it: run it in parallel with the current process for a month and compare, before it writes anything.

8. Triaging returns against your policy changes

Reason classified from what the customer wrote, routed to refund, exchange, repair or manual review, with the borderline cases deliberately sent to a person.

Replaces: a queue that is pure cost and visibly slow to the customer.
Needs from you: a written policy precise enough to test against — writing it is half the project.
How you prove it: shadow mode. It decides, a human decides, you compare, and nothing it decides takes effect until the two agree often enough.

It watches and raises a human

9. Watching stock, prices, margin and data quality watches

Stock cover against sales rate, competitor price gaps on your top lines, margin drift by category, and whether the numbers feeding all of it are still correct.

Replaces: somebody remembering to look.
Needs from you: a threshold and a named person who receives the alert. An alert nobody owns is noise.
Worth knowing: this is the cheapest item on the list and the most often skipped, because its failure mode is silence. In our own warehouse a scheduled quality check caught broken source data on day one of operations, before a wrong number reached anyone.

How we would choose between them

Ask	Why it decides the answer
How many times a week does this happen?	Agents earn their keep on volume. A task done twice a month is not worth the governance around it.
Is somebody currently waiting?	Work that blocks a person — a queue, an out-of-hours gap, a BI backlog — returns value immediately rather than theoretically.
What does a wrong answer cost?	A wrong product description is embarrassing. A wrong refund is money. Price the failure before buying the capability.
Can I write down the right answer for fifty cases?	If you cannot, you cannot measure it, and you will be buying on impression. This question disqualifies more projects than any other.

One pattern worth noticing across all nine: the hard part is rarely the model. It is having your policies written down, your metrics defined, your catalogue coherent and someone named to own the output. That work is reusable — do it for the first agent and the second is dramatically cheaper.

Next in the series: asking your own data a question — the third job above, and what took it from one right answer in ten to nine. All articles.

Nine jobs a retailer can hand to an AI agent today