In an e-commerce business, AI pays off almost entirely in one place — and it is not the chatbot
Where it is actually worth putting: catalogue, internal search, returns and forecasting. With a map of what each costs and returns, and the three things not to do first.
If you run an online shop, someone has pitched you a chatbot in the last two years. It is the thing you can see, show in a meeting and install in a week — and it is also the one that, in the vast majority of cases, moves the numbers least. It answers questions a well-written order status page would already have settled, and to do it well it needs data almost nobody has prepared for it.
The real value sits further back, in three places nobody demos because they have no face: product records, internal search, and handling what comes back. This article takes them apart one by one, and at the end puts everything on a plane with two axes — what it costs, what it returns — because the real question is not “is AI useful”, it is “which one do I start with on Monday”.
1 · The catalogue, where it pays off most
An e-commerce business with a few thousand SKUs almost always has the same problem: the product records are whatever the supplier sends. Titles written for an internal price list, descriptions missing or copied identically from other sellers, attributes absent exactly where filtering would need them. It is not laziness: rewriting four thousand records by hand is months of work nobody can afford.
It is precisely the kind of task where a language model is useful: high volume, cheap errors, and a person checking before publication. Look at the same row before and after — the switch at the top right.
catalogue · 3 of 4,180 SKUs
18V cordless driver with two batteries and carry case
VT-4180
Compact driver for assembly and maintenance, with torque adjustable across 20 settings. Includes two 2 Ah lithium batteries and a fast charger in a hard carry case.
18V
code 4180
category: power tools (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
max torque 45 Nm (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
batteries included: 2 (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
use: assembly (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
Disposable nitrile gloves, size L — box of 100
VT-2255
Powder-free blue nitrile gloves, resistant to oils and detergents. Suitable for workshops, laboratories and professional cleaning. Box of 100, size L.
size L
code 2255
material: nitrile (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
powder-free (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
pieces: 100 (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
sector: workshop (Marked fields are proposed by the system and pass human review before publication. The distinction stays visible in the record forever.)
Why this, before everything else
Because those records are the raw material for everything that follows. Internal search cannot filter on attributes that do not exist, recommendations cannot connect products nobody classified, and Google cannot show a record that says nothing. Every euro spent downstream on a poor catalogue returns less than the same euro spent here.
2 · Internal search, which almost nobody looks at
In an online shop, people who use the search bar buy far more often than those who browse categories: they are after something specific, so they have already decided. It is also the part of the site that gets the least attention, because it does not show up in presentations.
Classic search compares words. Someone typing “cordless drill” does not find “18V driver” even though that is what they need, because neither phrase contains the other's words. Semantic search compares meanings, and it finds that match.
Your industry's synonyms are in no dictionary. “Allen key” and “hex socket”, “angle grinder” and “disc cutter”: people in the trade know them. A system that learns from the searches actually made on your site collects them by itself, and within months knows more than any hand-written list.
Zero-result searches are the most valuable information and the most ignored. Someone searching for something you do not have is telling you what to stock, or that your catalogue calls it something else. That list should be read weekly, and in almost no company does anyone open it.
Be careful not to overdo it: an engine that “understands” too much returns everything and stops being trustworthy. The right threshold is the one where a search with no match says there is nothing, instead of showing something vaguely similar.
3 · What comes back, and what will run out
Two unglamorous jobs with a direct return, because they do not touch sales: they touch the costs that eat them.
Returns, classified as they arrive. The model reads the reason the customer wrote and puts it in a category: wrong size, not as described, damaged, changed mind. After three months the list tells you which product records are lying — and fixing those is worth more than any campaign, because every return costs two shipments plus warehousing.
Demand forecasting. You do not need a sophisticated model: history, seasonality and supplier lead times are enough to say what will run out and when to order it. It has the most direct return of all, because a stock-out is measured in lost sales and idle stock in tied-up cash.
Routing incoming requests. Emails get classified and routed, and for the most frequent ones a draft is prepared for a person to approve. This is the right place for AI in support: it does not answer the customer, it prepares the answer for whoever does.
The console where all this becomes visible — live
An order queue that moves, exceptions being resolved, and an agent answering over the data at the bottom. It is the demo from the agents article, and it is where the three jobs above reach the people actually doing the work.
Here are the seven most-discussed options placed on a plane. Horizontally what they cost to stand up, vertically how much they move the numbers. Tap a point for the details — and look especially at the top-left quadrant: low effort, high return.
effort
return
Indicative cost
Time
Record enrichment
22
84
€6,000 – 15,000
3 – 6 weeks
Semantic search
34
76
€8,000 – 20,000
4 – 8 weeks
Reorder forecasting
40
80
€10,000 – 25,000
6 – 10 weeks
Returns classification
18
55
€3,000 – 8,000
2 – 4 weeks
Request routing
30
58
€7,000 – 16,000
4 – 6 weeks
Recommendations
62
46
€12,000 – 30,000
8 – 12 weeks
Website chatbot
74
30
€15,000 – 40,000
8 – 14 weeks
Record enrichment — Titles, descriptions and attributes generated over the existing catalogue, with human review before publication. It improves conversion, internal search and organic ranking together, because all three read the same records. It is almost always the starting point.
Semantic search — The search bar understands meaning instead of words. It returns a lot if the catalogue is broad and the records are already decent — and almost nothing if the records are the supplier's, which is why it comes after enrichment.
Reorder forecasting — History, seasonality and supplier lead times to say what to order and when. The most direct return of all: fewer stock-outs, less capital tied up. It does require two years of tidy history, and few companies have that.
Returns classification — The return reason read and categorised. Small, quick, and it produces the list of product records that lie. It does not move sales tomorrow, but it removes a recurring cost and improves the catalogue with real data.
Request routing — Incoming emails classified, enriched with customer history and routed, with a draft ready for the most frequent ones. It returns in proportion to volume: below a hundred requests a month it is not worth the maintenance.
Recommendations — “Customers who bought this also bought…”. It works well on large catalogues with heavy traffic, and hardly at all below a certain data threshold: without enough purchases to learn from, it suggests random pairings that erode trust. Consider it later, not first.
Website chatbot — The thing everyone pitches and the one that returns least, for two reasons: it answers questions a good order status page would already have settled, and to answer well it needs a clean catalogue and clean order data — that is, all the work lower down this map. Done last it works; done first it becomes the example people cite to say “AI does not work”.
start here
The figures are orders of magnitude for a shop with a few thousand SKUs and need checking against your case: an eight-hundred-item catalogue costs less than a forty-thousand one, and one with tidy history costs far less than one that must be reconstructed. The relative position of the seven points, though, changes little from company to company — and that is what you need in order to decide the order.
5 · The three things not to do first
These are not technical mistakes: they are sequencing mistakes. All three become reasonable later, and all three do damage at the start.
The chatbot as a first project. It is the bottom-right quadrant of the map: high effort, low return, and built on data that is not ready. When it answers badly — and at the start it does — the company concludes AI is useless, and closes the door on the things that would have worked.
Publishing generated text without review. On a few hundred records you can spot-check; on all of them you cannot. A description attributing a feature a product does not have is not a style error: it is a promise made to a customer, and product information carries legal weight. Human review is not excessive caution, it is the condition that makes the work usable.
Buying a tool before deciding the use case. Choosing the tool is the least important decision and the easiest to change; the use case and the data are the hard ones. Whoever starts from the subscription ends up looking for a problem to justify the spend, which is the fastest route to a project nobody uses.
Half an hour on your catalogue, not on AI in general
In the audit we open your catalogue and look at three concrete things: how many records have a real description, how many useful attributes exist for filtering, and what people search for that finds nothing. From there, half an hour tells you whether the first point on the map applies to you and what it would cost.
If the catalogue is already in order, we move to the second point and look at internal search. And if the volume does not justify any of the seven, I will say so — it happens below a certain size, and in that case the most useful thing you can do has nothing to do with AI.
On product pages. It is where the same spend produces the largest return, because every incomplete page is a page that does not rank and a filter that does not work: the supplier's code becomes a readable title, a description, and the attributes people actually search by. It costs between €6,000 and €15,000 and ships in three to six weeks. The chatbot, which is the first thing everyone asks for, is last on the list.
How much does an AI chatbot for e-commerce cost, and is it worth it?
Between €15,000 and €40,000, with eight to fourteen weeks of work: the most expensive of the seven uses we mapped, and the one that returns least. The reason is not the technology. A useful chatbot has to read the catalogue, orders and returns policy — so it rests on data that in most shops is not in order yet, and fixing that data is the project that actually matters, one that usually pays for itself on its own.
How do you know whether an online shop's data is ready for AI?
Look at the catalogue. If descriptions are missing, attributes are empty and the same product appears under three different names, no language model fixes that downstream: it will answer with the same confusion it finds. The practical test takes a minute — search your own site for one of your products using a customer's words, not your internal code. If you cannot find it, the problem is the catalogue, not the search engine.
What does semantic search change compared with ordinary site search?
Ordinary search compares strings: someone typing «cordless drill» never finds the «18V battery driver», which is exactly what they wanted. Semantic search compares meanings, and recovers the sales that today end on an empty results page. It costs between €8,000 and €20,000. It is worth looking at, because it is the function almost nobody measures and the one used by people who have already decided to buy.