An AI agent can take over the sorting of new cases, write drafts of replies to common questions and gather the picture for the person who replies. A person reads and sends, and decisions about money, exceptions and sensitive cases stay with you. Start with a single kind of question and decide in advance how you will know it works.
What the agent can take over
- Sorting new cases. By urgency and subject, so the person replying starts with what matters most.
- Drafting replies to common questions. The draft follows the way you reply today, and a person reads it before it is sent.
- Gathering the picture. Background from email and systems in one place, so the person replying does not have to search.
- Compiling a weekly status. A short overview for the person in charge, with what needs a decision marked.
Where a person stays
- Money and exceptions. Compensation, credits and departures from terms are decided by a person on your side.
- Sensitive cases. When a customer is upset or the situation is hard, a person is needed who listens and chooses the words.
- Contracts and legal questions. Questions that call for judgement about a contract go to the person responsible.
- Everything that is sent. The agent writes drafts. Nothing is sent until someone on your side approves it.
Start with a single kind of question
Choose the question that comes most often and has a clear answer, for example a status question or a question about delivery or terms. Write down how you reply today and let the agent draft in that tone.
Let a person read every draft during the first weeks. Then look at how many drafts were sent as they were and how many were changed. It shows what works and what needs adjusting before you add a new kind of question.
An example of a first flow
You receive many questions about delivery status. The agent reads the case, gathers the picture from your systems and writes a draft with the date and the next step. The person replying reads the draft, changes what needs changing and sends it.
A flow like that is small enough to evaluate in a few weeks, and it is possible to stop after it.
What you measure
- Reply time. How long it takes from the case arriving to the customer getting an answer.
- Time per case. How much time each case takes from the person replying.
- Drafts sent unchanged. It shows how well the drafts match the way you reply.
- Colleagues' view. A short question about whether the agent makes the day easier.
Stop points
Decide in advance what should get better, for example how long it takes to reply and how much time each case takes from you. Decide also when you stop and look at the result.
With us there is a stop point after every step. You choose whether we go on, change course or stop, and nothing is left hanging if you stop.
So that colleagues use it
That is the first design question. Short messages in the right language, in tools they already have and without a new login. The agent should be where colleagues already work, and they should see that it saves time in the first week.
Customer data
Information about customers is personal data. You are the controller and we are the processor, and a data processing agreement is signed before we start. Read more on the page about trust and data protection.
Common questions
Can an AI agent reply to customers on its own?
We build it so that the agent writes drafts and a person reads and sends. Nothing is sent until you approve it.
Which cases suit a first try?
Those that come often, have a clear answer and where a draft saves time, for example status questions and common questions about delivery or terms.
Do we need to change systems?
No. We build on top of the systems you already have.
How long does it take to get started?
Count on weeks. The length is set in the mapping.
What does it cost?
A fixed price per step, with a cap. The price of the next step is set in the mapping before the step starts.
What happens to customer data?
We sign a data processing agreement before we start, and you decide which information the agent may see. See the page about trust.
How do we know it works?
You decide in advance what should get better, for example reply time and time per case, and look at it at the stop point after the first step.