AI agents
An agent takes the first round on what comes in: reads it, assesses it, fetches what is missing and prepares a reply. Then a human looks before anything becomes binding.
Short answer
An AI agent is a program that can read, assess and act in your systems: sort incoming email, pull records from the ERP, put together a set of material and propose a reply. The difference from a chatbot is that the agent does things rather than only talking, and that it works in several steps without anyone directing each one.
A chatbot answers what you write. A classic automation does exactly what it was programmed to do, every time, and stops when reality deviates from the rule. An agent sits between the two: it is given a goal and a set of tools, and decides for itself which steps are needed to get there.
A concrete example. An order email comes in with an attachment. The agent reads the email, interprets the attachment, looks the customer up in the ERP, notices that the delivery address is missing, formulates a question back to the customer and puts the case on hold. Nobody wrote a rule for that particular combination. The agent could read, fetch and act, and it chose the steps itself.
That is also why agents are harder to build well than they look in a demo. A program that picks its own steps can pick the wrong ones, and the whole craft lies in deciding what it may do on its own and where a human comes in.
The honest version, which is the only one that holds after three months in production.
01
Reading unstructured text, working out what a case is about, pulling the right records from several systems and putting them together into one set of material. It is work that takes a human five to fifteen minutes per case and an agent a few seconds, with more even quality at volume.
02
An agent built properly knows the difference between a case it can handle and one it cannot. Uncertain cases go to a human with the full material attached, instead of being guessed at. You set the threshold for what counts as uncertain, and it can be adjusted over time.
03
Language models can phrase something that sounds right but is not. The countermeasure is that the agent may only answer from approved sources in your systems, that the answer points back to where the information came from, and that someone reviews it before it goes out. The risk does not disappear, it becomes manageable and visible.
04
If the customer records live in three systems with three different spellings, the agent will make the same mistake a new employee would have made, only faster. Sometimes the conclusion of a mapping is that the data needs cleaning first, and then we say so.
05
Pricing, hiring decisions, customer relationships and everything that requires someone to take responsibility stay with people. The agent removes the groundwork so that whoever decides gets a complete set of material instead of an empty folder.
Five patterns that recur in almost every industry.
01
Reads what comes in, works out what the case is about, fills in records from your systems and passes it on to the right person with a summary and the original material still attached.
02
Gathers the enquiry, previous deals, product data and prices into a draft quote. Uncertain lines are flagged for manual checking. The salesperson reviews, adjusts and sends.
03
Answers recurring questions about where a case, a delivery or an order stands, taken from approved system sources. It removes the stream of small questions that otherwise interrupts someone all day.
04
Pulls the figures from your systems at set times, compiles them in the same format every time and puts the draft where the report belongs. Nobody spends a Friday afternoon pasting in columns any more.
05
Answers employees' questions about your own routines, contracts and documents, with a reference to the source. The new colleague does not have to ask the person at the next desk, and that person does not have to answer the same thing for the eleventh time.
The same technology, a different working day. Six examples of where we usually start.
01
Enquiries arrive by phone, email and web form and often get typed in several times before the right technician takes over. The agent receives the case, categorises it, proposes a time and prepares a draft quote that someone reviews before it goes out.
02
Fault reports come from several channels and almost always miss a piece of information. The agent asks the follow-up questions, prioritises according to your rules, creates work orders for the right contractor and keeps the tenant updated until the case is closed.
03
Every month, time goes to chasing the same missing documents and sorting incoming material. The agent sees what is actually missing per client, sends targeted reminders and files documents to the right client before an employee takes over.
04
Bookings, feedback and documentation eat the time that should have gone to the relationships. The agent matches interview slots, gathers candidate material ahead of the assessment and keeps the CRM up to date. Selection and hiring decisions stay with people.
05
Orders arrive by email, PDF and phone call, and customers ask about status over and over. The agent turns incoming orders into structured records, spots what is missing, answers status questions from the system and escalates deviations to traffic control.
06
Customers order using their own names for the articles. The agent matches the customer's language against your product register, flags uncertain matches, fetches stock status and delivery time and prepares a draft quote for the salesperson to review.
Four steps from idea to something that runs every day.
01
We look at four things for every candidate: volume (how often does it happen?), complexity (how many exceptions are there?), data quality (can the records be trusted?) and business value (what does it cost you today?). The candidate that wins is rarely the one that sounded most exciting in the meeting.
02
The agent is connected to your existing systems through their APIs and gets only the permissions it needs. We build against your real data from the start, because it is the exceptions in real data that decide whether the solution holds.
03
The agent prepares and proposes, a responsible person approves. Every run is logged so that you can follow what was read, what was proposed and who made the decision. The checkpoint can be moved later, but it always starts in place.
04
In production we track how often the agent gets it right, how often it escalates and what it costs in model calls. Then we adjust. An agent nobody follows up gets worse over time, because the reality around it changes.
Riktpriser
Indicative prices for the build. Bigger than that? Then we start with a conversation about whether two people are enough, and we answer honestly.
Indicative prices excl. VAT. Your project is always quoted at a fixed price before we start.
A responsible person reviews before anything becomes binding. That covers quotes, contracts, customer replies, payments and everything else that would be unpleasant to take back. The agent does the groundwork and presents the material, the human makes the decision.
We set the line together with you, and it does not have to stay in the same place forever. A common pattern is that the agent first gets to propose everything, that after a few weeks of logs you see which case types it gets right every time, and that those workflows are then let through automatically while the rest keep passing a human.
It is also what makes the solution survive. Automations that get switched off almost always do so because they were once allowed to act alone in a situation where they should not have.
Common questions
Yes, just as a human can. That is why we build with source requirements, logging and a human review before anything becomes binding. We also measure how often the agent gets it right, so that the errors are known figures and not surprises.
Most of them. The most common are Fortnox and Visma, Microsoft 365 and Google Workspace, HubSpot, Salesforce and Pipedrive, and Slack and Teams. If your system has an API, or even an export, we can connect it.
A simple agent is in production in 1-2 weeks, a pilot with several integrations in 2-4 weeks. What decides is rarely the build but how quickly we get access to the systems and who at your company can answer questions about the exceptions.
We keep the solution up to date as part of maintenance and test against the same set of real cases every time something is swapped out. A new model being better in general does not mean it is better at your particular task, so we check before we switch.
No. You need to know your business and be able to point at what takes time. We show what can be automated, what it requires and what it costs. If more people in the organisation are to work with the tools, AI training is a separate track.