CODEAXIS OÜ — TALLINN — USE CASES
Four workflows where Estonian teams save hours.
Concrete outcomes of AI adoption across support, sales, documentation and learning. Each section names the input, the AI step and the human check — so you can spot your own starting point.
- 01 Inbox
- 02 Extract
- 03 Draft
- 04 Review
- 05 Signed
How to read this page
Each case is a short chain: input, AI step, human check.
Estonian teams do not buy AI as an idea — they buy it as a working route through the day: an inbox that fills while you sleep, a quotation that drafts before you reach the coffee, a catalogue that writes itself from a photo, a newcomer who stops interrupting the senior engineer. Below are the four routes we see most often, in the order a business tends to adopt them. Read them as patterns, not as a pitch: each one is honest about what the AI does, where a person still has to sign off, and what the pilot should measure before anyone calls the project a success.
They run in Estonian and English side by side, because most firms here serve both without a translation layer. If more than one of the chains below sounds like your week, that is normal — the four usually connect.
Case 01 · support
Customer support with quoted sources.
Live stage — draft from sources, staff approve the sensitive topics before send.
This is the quietest of the four — and often the one that pays first. A help desk's real cost is not the average question; it is the twentieth variant of the same question, asked by a customer in a different language. The assistant reads your own help articles and past resolved tickets, drafts an answer with the source link attached, and either sends it if the topic is safe or leaves it for a colleague to approve. Answers stay consistent whether the customer wrote in Estonian or English, which matters more than it sounds: inconsistent policy explanations are what generate the next ticket.
Input
Help articles and past resolved tickets, plus the current stock, delivery and returns policy.
AI step
Draft or direct answer with a source link the customer can read, offered in Estonian and English from the same knowledge base.
Human check
Staff approve warranties, refunds and anything with a legal edge before it leaves the queue.
Case 02 · sales enquiry handling
From procurement email to quotation.
Live stage — match, draft, wait for sales confirmation.
-
01 · Arrives
A procurement email or RFQ lands in the shared inbox with attachments — a spec sheet, a spreadsheet, sometimes a photographed form.
-
02 · Matched
Items are matched against the price list and your pricing rules; ambiguous lines are flagged instead of guessed at.
-
03 · Drafted
A quotation is drafted in your template, with delivery terms and validity dates filled from your own shelves of rules.
-
04 · Confirmed
A person confirms margin, price and delivery, then signs. The AI never signs — it only clears the paperwork so the signature is fast.
This is the route that shows up most in Estonian B2B procurement, and it is no longer a theoretical pattern. An automated agent that watches the inbox, matches line items to a price list and drafts the quotation is now core tooling in several shops — what used to cost a day of chasing attachments now takes the time it takes a salesperson to read and confirm. The value is not speed for its own sake; it is fewer missed enquiries, because the ones that used to sit unread until Friday now come back with a number before lunch.
Meanwhile, at 22:40 — a salesperson confirms a quotation the pipeline drafted after hours.
Case 03 · documentation and listings
Listings and documents from photos and tables.
The second-hand trade is the sharpest example, and it is also the one Estonians tend to recognise first — because several local studios already build it. A camera, a jacket, a box of spares, a stack of thirty photographed items that need to be listed by end of week: the assistant reads each photo alongside a short note or a spreadsheet row and drafts a structured, search-friendly description. The result is not a sales gimmick — it is the same template logic a reseller or service firm uses for a parts catalogue or a product page, scaled to the volume of a Sunday's photo dump.
The three inputs a draft begins from
-
01
Photographs of the item itself, taken the way an honest listing ought to be taken — the marks and wear visible, not hidden.
-
02
A spreadsheet row or a short typed note — dimensions, era, condition, what the seller actually knows about the piece.
-
03
The house template: your category rules, your title length, the attributes your own listings always carry.
Human check
A quick edit before publishing. The draft earns its keep when the editor's first reaction is 'close enough to fix', never 'start again'.
Case 04 · learning support
Learning and onboarding help.
The fourth case looks like training software and behaves like a patient colleague. It sits on top of your internal guides and process notes and produces a study plan, an explanation, or a challenge — not a ready-made answer. The distinction matters: a learner who is handed an answer learns to ask the machine again; a learner who is asked to find the rule and apply it learns the work.
Estonia has an unusual advantage here. The national AI Leap programme and the Institute of the Estonian Language invested heavily in exactly this pedagogy — helping learners plan and think independently, in Estonian, with teachers in the loop. Companies borrow the same pattern for onboarding: the assistant explains where to find the rule, then asks the newcomer to apply it, and a mentor reviews whether the newcomer actually understood.
Input
Internal guides, process notes, the folder of files a senior would send a newcomer, plus last month's meeting notes.
AI step
Study plan, explanation of where to find the rule, and a prompt asking the newcomer to apply it — never a finished answer to copy.
Human check
A mentor reviews. Onboarding is where a small mistake is cheapest to correct, so the mentor's review is where trust in the system is built.
Underneath all four
Knowledge upkeep behind all four.
Every one of the four chains above stays honest only because someone keeps the facts underneath them in order. This is the least glamorous part of AI automation and the part that decides whether a pilot is still useful nine months in. Scattered files and a stale price row are what make a support assistant quote a discontinued return window or a sales draft offer last autumn's price.
The AI step
Structure the facts, flag contradictions between sources, and suggest review dates so nobody has to remember when the policy last changed.
The human check
The owner confirms the change. Attributable facts and scheduled review are what make an assistant quote something a visitor can trace back to a source the team trusts.
What a pilot should measure
Four numbers before you call it done.
Pick the metric the workflow actually moves, keep the baseline for two weeks before you switch anything on, and review weekly. If the acceptance rate stays low, adjust the sources or the rules — do not adjust the metric.
| Signal | Why it reads | Watch for |
|---|---|---|
| Drafts accepted without edit | Direct read on whether the assistant's sources and rules match the work. | A dip after a policy change — often a stale source, not a model problem. |
| Time from enquiry to quoted | The clearest sign that procurement chasing has actually gone away. | Speed for its own sake — a fast wrong quote is worse than a slow right one. |
| Tickets resolved with a quoted source | How often the answer can be traced back to the article a customer can read. | Tickets closed without a link — those stop building the knowledge base. |
| Listings published per week | Volume the editor can actually keep up with, not raw output. | A rising count with falling quality — pull it back to what editing can hold. |
Keep the baseline for two weeks before launch and review weekly. Stop or adjust if acceptance stays low. That is what keeps automatiseerimine honest, and it is the only way anyone in the room will trust the numbers at the end of the quarter.
The work happens here — Tallinn, and it happens after hours.
AI landscape in Estonia →Questions we get
What usually decides the answer.
Three questions come up in almost every first conversation. If yours is different, write to us — a one-line description of the workflow is enough to get a specific answer.
Which use cases are strongest: government, education, or enterprise automation?
In daily business buying, enterprise automation for support and sales leads — those two workflows pay the pilot back inside a normal quarter. Education pilots like the national programme and public-sector controls like Reason Reserve shape expectations about safety and human review rather than driving direct sales. Read them as the standards your own pilot will be judged against, not as the first thing to buy.
How do startups handle compliance?
The pattern we see is unglamorous and it works: limit the sources an assistant is allowed to read, log every output along with the source it came from, and keep a human approval step in front of anything that touches money or legal text. That way, if a regulator or a customer asks why the system said what it said, the answer is in a log rather than in a memory.
Can a small team run this?
Yes, on two conditions: one person owns the knowledge files and keeps them current, and somebody reviews the flag list every week. A team of three can run one of these workflows comfortably. What fails is not size — it is a project with no owner, where the sources drift and the assistant slowly starts answering from last year's policies.
If more than one chain above sounds like your week, that is a normal starting point.
Bring a single workflow — the inbox queue, the RFQ pile, the catalogue backlog, or the onboarding folder — and we will name the input, the AI step and the human check for it in a first conversation. If the fit is right, we draft the pilot plan together, including what the pilot will measure and what the baseline should be before anyone switches the workflow on.