AI Agents

We build agents for a defined task.

An agent can move a piece of work through several steps, using the permissions it’s been given. We define the task, the conditions for continuing and the point where a person takes over.

The procedure Relay ran with, beside its recorded run on one supplier invoice

How the workshop handles messages

Relay ran with exactly this procedure. The clauses its notes cite are marked.

Hands the decision to a person

  1. Clause 1

    Match every sender to a customer or supplier record before acting on their message.

  2. Clause 2

    Check supplier invoices line by line against the purchase order, the goods receipt and the contract prices. Put an invoice with any discrepancy on hold.

  3. Clause 3

    Never pay, dispute or accept an invoice alone. Prepare the dispute and ask the owner to approve it before anything goes to a supplier.

    The workshop owner approves

  4. Clause 4

    Parts orders up to 50 EUR for a customer repair can be placed without approval.

  5. Clause 5

    Quote customers from the workshop price list. Hold a slot for up to 24 hours while they decide.

  6. Clause 6

    Brakes, a crash or an injury go to the mechanic on duty. Safety warnings are fine; do-it-yourself repair instructions are not.

    The mechanic on duty takes over

    In the recorded photo run, the crash inspection goes to the mechanic on duty.

  7. Clause 7

    Write to customers and suppliers in the language and tone they used. Address suppliers formally.

  8. Clause 8

    Before every step, note briefly what you are checking and why.

RelayBike workshopBuilt by Version2
Parts supplierRačun VP-2026-0917Email · 09:02

Its notes from the run14 recorded steps

  1. 01

    Match the sender racuni@•••••.hr to a supplier record before acting on invoice VP-2026-0917, per clause 1.

    suppliers.findClause 1

  2. 03

    Pull PO-5521 lines to compare ordered quantities and prices against invoice VP-2026-0917, per clause 2.

    purchase.orders_getClause 2

  3. 06

    Chain overcharge on line 1 of VP-2026-0917: 10 units at 38.90 invoiced against contract price 31.50 in Prilog 1.

    calculator.evaluate

  4. 12

    Invoice VP-2026-0917 has four discrepancies totalling 218.80 EUR net, so it goes on hold under clause 2.

    invoices.set_statusClause 2

  5. 13

    Prepare the formal Croatian dispute letter for VP-2026-0917 listing the four items totalling 218.80 EUR net, for the owner to approve under clause 3.

    mail.draftClause 3

  6. 14

    Clause 3 requires the owner to approve the dispute before it goes to the parts supplier; draft D-3301 covers 218.80 EUR net.

    approvals.requestClause 3

  7. Waiting for your approval

    The workshop owner approves

What one is made of

  1. 1

    The trigger

    Where work arrives: one number, one inbox, one folder.

  2. 2

    What it may touch

    Each system is a tool with its own limits. It can only use what it’s been given.

  3. 3

    The steps

    The procedure, written in your words, where your team can read it and change it.

  4. 4

    The stop

    Every case it must not decide, written down before it runs.

  5. 5

    The approval

    A named person decides, and someone else when they’re away.

  6. 6

    The log

    What it did and why, in sentences you can search.

An agent is a piece of software. It has no judgement about your business beyond the rules you write with us, and the responsibility for what it does stays with people.

The work

Each of these jobs has a point where the agent stops.

These are six jobs people ask us for. Each one repeats, has a written rule for the ordinary case and has a moment software shouldn’t decide.

A customer writes an order the way they speak, on WhatsApp or by email. It’s matched to your items and prices, checked against stock and the delivery cutoff, and prepared.

  1. A message
  2. Matched to your items
  3. Draft order
    1. You post itStops and waits for a person

Log

  1. A message
  2. Matched to your items
  3. Draft order
  4. You post it

If the job is mostly answering questions from your own documents, that’s AI Integration.

One of ours, working

Relay reads a bike workshop’s inbox.

We built Relay to work through a small business’s inbox under a written procedure. Each step comes from a recording of an actual model run, kept in the order it happened.

It stops at the approval and waits. Sending the letter it prepared, or taking the job over, is the decision a person makes.

Relay

Bike workshop

Built by Version2

Parts supplier

Email, racuni@•••••.hr

CV-3109

SubjectRačun VP-2026-0917

Poštovani,

u privitku Vam šaljemo račun VP-2026-0917 za Vašu narudžbu PO-5521. Molimo uplatu do 14.10.2026.

Lijep pozdrav,
Sanja
Računovodstvo

    Client work

    Maestral works through a wholesaler’s evening orders.

    Tramontana Gastro supplies hotels and restaurants on the Dalmatian coast. Maestral, the company agent we built for them, turns orders sent on WhatsApp into draft orders overnight. Anything that commits the company to money, a price or a new promise goes to a named person in Slack.

    Ordered on WhatsApp at 22:47, drafted as order 18834 at 22:48 and on the van at 5:40. In between, at 23:14, a credit note went to the director for approval in Slack.

    WhatsAppAn evening order, written the way the customer speaks
    Order deskEvery draft line traced to the customer’s words
    SlackA credit note waiting for the director’s approval
    KaštelaOn the van, route Makarska 2

    Screens from Tramontana’s workspace, in Croatian

    It may
    Draft orders in the ERP, reply to customers inside WhatsApp’s service window, and answer the team in Slack with its sources.
    It stops before
    Moving money, changing a price, a credit limit or a customer’s terms, and posting any document in the ERP.
    Who approves
    Credit notes go to customer service, finance or the director depending on the amount. Out of hours, finance approvals go to the director.

    Rules it follows: Fiskalizacija 2.0, EU AI Act transparency, GDPR, processing in the EU

    Limits

    What an agent won’t fix.

    • A process without clear rules

      If two people on your team would answer the same case differently, settling that comes first, and it’s your team’s work to do.

    • A decision nobody owns

      Approving a payment or sending a binding reply needs a named person and an approval boundary. An agent doesn’t create an owner where there isn’t one.

    • A job with no judgement in it

      Sometimes a rule and a form are enough. Ordinary automation is cheaper to build, cheaper to run and easier to check, and we’ll say so when it fits.

    • A rule nobody updates

      Procedures change, so an agent’s rules and checks need reviewing after launch, which is what care is for. A stale rule goes wrong without anyone noticing.

    Responsibility

    Your process owner decides which actions may run and which need review. We define the workflow and its checks, and a named person handles the cases it can’t complete.

    Questions

    The things worth asking first.

    What’s an AI agent, in one sentence?

    Software that carries one repeating job through its steps inside permissions you set, and stops at the cases you told it not to decide.

    Will it do things without asking?

    Only the things you’ve listed. Everything else waits for a named person, and the list itself is a document you can read and change.

    What happens when it gets something wrong?

    The work goes to the person who owns it, with what the agent read and the likeliest answers, and the log says why it chose what it chose. Before launch it runs in parallel with your team on your own past cases, so the wrong answers surface while nothing is at stake.

    Can it work with the systems we already have?

    It goes through the same door a person uses, so your system’s own checks still apply. Inboxes, WhatsApp, Slack, calendars, an ERP, a web shop, bank exports and e-invoices are all ordinary. If a system has no door at all, you’ll know in the first week.

    How long does it take?

    3 to 5 weeks for one job end to end, 6 to 9 weeks for three to five that share what they know. The first of those weeks is spent on your real cases.

    What does it cost to run after the build?

    Care is €190 a month, and the models are billed at what they cost. Usage depends on how much work the agent does, and it’s read back to you every month with the rest of the review.

    Who is responsible for what it does?

    You are, the same as for anything sent in your company’s name. Your process owner decides which actions may run alone. We define the workflow, its checks and its exceptions, and we answer for those being what we agreed.

    Do our customers have to be told?

    Yes. Article 50 of the EU AI Act has applied since 2 August 2026: a person must be told they are dealing with an AI system at the first interaction. Every reply an agent sends carries that line, and a person stays one tap away.

    Do we own it?

    Yes. The code, the rules and the records are yours and they sit in your accounts. The handover includes the procedure your team wrote, so the part that decides how the agent behaves is a document you keep.

    Scope and price

    The first engagement ends in a written verdict.

    We map one task and its exceptions, then test a bounded version against agreed examples. The written verdict records whether it’s ready for a pilot and what remains unresolved.

    AI Agent

    One workflow run end to end, human approval on exceptions, guardrails, and a dashboard that shows what the agent did.

    from €3,300per project, 3 to 5 weeks

    Several AI Agents

    Three to five workflows with shared memory, escalation to your team, and evaluations on your own cases.

    from €7,400per project, 6 to 9 weeks

    Agent Care

    Monitoring, guardrail updates and a monthly review of what the agent did. Model usage is billed at cost.

    €190a month, from handover

    The build and the care are separate lines on the quote.

    ContactWhatsApp