AI features inside your product
A model inside the interface: search by meaning, generation, tags and moderation, recommendations, documents into data — with measured quality and predictable costs.
04 05 service 4 of 5 in Automation & AI
What’s included
- Semantic search that understands intent, not keywords
- Generated descriptions, summaries and replies
- Classification, tagging and moderation
- Personalised recommendations
- Model routing to keep costs predictable
- Guardrails, evaluation and logging so quality is measured, not assumed
- Private or on-prem models when data can't leave
Who it’s for
- SaaS products competing against rivals who already shipped AI
- Marketplaces and catalogues where search is the main funnel
- Products sitting on data they've never turned into a feature
In short
AI features put a model right inside your product’s interface: search that understands meaning, generated descriptions and summaries, automatic tags and moderation, recommendations, parsing of documents into data. We start from a task and a metric, check the quality on a set of your examples, and only then build the feature into the product. Costs stay predictable thanks to routing between models and caching; quality is protected by guardrails, format checks and a fallback if the model is unavailable. When data must not leave your perimeter, the model runs on your server. The price is counted in hours and fixed before the start.
Which AI features we build in
Not a chat window bolted on for the sake of it, but features that save users time and turn your data into value.
-
01
Search by meaning
Understands what a person means, not just the words: synonyms, descriptions in their own words, other languages.
-
02
Text generation
Product descriptions, summaries of long documents, draft replies — with a person approving where it matters.
-
03
Classification and moderation
Tags, categories, sentiment, spam and rule-breaking content — sorted automatically.
-
04
Recommendations
Similar products, related articles, “people also take” — based on meaning and behaviour.
-
05
Documents into data
Fields are extracted from PDFs, scans and emails and land in the product as structured data.
-
06
Help inside the product
Answers about how the product works, from your documentation, right where the user got stuck.
Keyword, semantic or hybrid search
In catalogues and marketplaces search is the main path to a purchase. Each kind of search has strong and weak sides.
| Criterion | By words | By meaning | Hybrid |
|---|---|---|---|
| Synonyms | only those written in by hand | understands by itself | understands by itself |
| A request in one’s own words | often finds nothing | finds what is meant | finds what is meant |
| Exact code or SKU | finds precisely | may miss | finds precisely |
| Other languages | a separate setup for each | works across languages | works across languages |
| Cost | the lowest | vectors are computed once per item and for each query | as by meaning |
| Best for | small catalogues and exact codes | knowledge bases and articles | catalogues and stores |
How we build an AI feature into a product
The model is the easy part. What matters is what is around it: the metric, the checks and the behaviour when something goes wrong.
-
A task and a metric
What exactly should get better and how we will know: fewer empty searches, faster processing, fewer manual edits.
-
A set of test examples
Real requests and documents with correct answers — every version of the feature is checked against them.
-
Comparing models
Several models are run on the same set: we choose by quality, speed and cost per request.
-
Building into the interface
The answer appears as it is written, there are clear loading states, and the user can always correct the result.
-
Guardrails and a fallback
Input and output are checked; if the model is unavailable, search switches to the ordinary one and generation waits in a queue.
-
Measurements in production
A log of requests, answers and costs; the metric is compared with the numbers “before”.
A cloud model or your own
The model is chosen for the task, the budget and the requirements for the data — sometimes both kinds work side by side.
| Criterion | Cloud API (Claude, GPT) | Local model on your server |
|---|---|---|
| Quality | the strongest models for complex tasks | good for narrow tasks: tags, extraction, search |
| Data | go to the provider by API; the terms forbid training on them | stay on your server |
| Cost | pay for actual use | a server with a graphics card — a fixed cost |
| Start | fast | needs hardware and setup |
| When to choose | most tasks, especially at the start | data must not leave the perimeter, or the volume is large and steady |
Predictable costs
An AI feature should not turn into a bill that grows faster than the product.
-
01
Routing between models
Simple requests go to a cheap, fast model; complex ones go to a strong model.
-
02
A cache for repeats
The same request is not paid for twice.
-
03
Vectors are computed once
For search, an item is processed when it is added or changed — not on every request.
-
04
Limits
Per user and per month — so one active account or a bot cannot eat the budget.
-
05
The cost of each request
Visible in the log: you know what a feature costs per user and per month.
-
06
Keys in your name
The model accounts belong to you — you pay the provider directly.
Common mistakes with AI features
-
AI for a press release
A feature without a task and a metric is used for a week and forgotten.
-
No test examples
Without them every change is a lottery: something got better, something quietly broke.
-
Generation straight to the storefront
Unchecked descriptions bring wrong specifications and claims you will have to answer for.
-
One expensive model for everything
Most requests are simple — paying the top price for each one is a waste.
-
No plan B
Model providers have outages too. Without a fallback the whole product goes down with them.
-
Extra personal data in requests
The model is sent only what it needs for the task — names and contacts are removed where possible.
Questions about AI features
How much does an AI feature cost?
Development is counted in hours and fixed before the start. Model usage is paid to the provider separately, by actual use; the estimate is free.
How long does it take?
As a guide, one AI feature takes one to four weeks, including the test set and the checks. The exact dates are fixed before the start.
Can you add AI to our existing product?
We do not edit someone else’s code, so the feature is built as a separate service with an API and documentation — your team connects it. In products we build, it is built in directly.
Which models do you use?
Claude, GPT and local models — chosen on a set of your examples by quality, speed and cost.
How do you measure quality?
On a set of real examples with correct answers, before launch and after every change, and in production — by the metric agreed at the start.
Can you generate descriptions for thousands of products?
Yes. Generation goes in the background from your specifications, and a person checks a sample and everything the model is unsure about before it reaches the storefront.
Is our data safe?
The work goes through the API, where the provider’s terms forbid training on your data; the model gets only what the task needs. When data must not leave the perimeter, a local model is used.
What if the model is unavailable?
The feature does not break the product: search switches to the ordinary one, generation waits in a queue, and the user sees a clear message.
Does semantic search work in other languages?
Yes: multilingual models understand a query in one language and find material in another.
Who owns the feature and the keys?
You. Code, model accounts and keys are registered to you and handed over at launch.
Online form
Discuss
an AI feature
Tell us about your product, the data it has and what users should be able to do — we will suggest which AI features pay off and estimate the work. The consultation is free.