AI query assistant: what can it handle?
When many similar inquiries arrive, the first deadlock often occurs before sales even reads the message. An AI assistant can prepare data and draft responses, while the team decides what works for the client.
Determine what the assistant is preparing
Start with work that you can clearly describe. Let's say: when a request for an offer arrives, select the service, deadline and contact information, then forward the message to the person who manages that type of project. Data not entered by the client remains empty. Don't ask an assistant to deduce the scope of work or budget from one vague sentence.
When the messages are sorted properly, add a draft response. He can confirm what the client asked for and suggest two additional questions. The team gets a useful starting point and a source query to go with it.
Give him content to rely on
Prepare a short list of services, forwarding rules and answers to frequently asked questions. For each piece of information, specify who maintains it. If the offer changes, the change must also reach the assistant.
For prices, deadlines and availability, use confirmed data from the appropriate system. If there are none, let the outline mark what needs to be checked. A convincing answer can be wrong and is therefore checked by a person who can confirm the offer.
Separate the proposal from the submission
In the first version, let the assistant prepare the drawings in the internal space. The employee reads, corrects and sends them. If the message contains a complaint, an unusual request, or an attachment that the assistant cannot reliably process, the system should assign it to the team without trying to resolve the case itself.
Ne tražite od asistenta da izvršava instrukcije iz primljenih poruka. Poruka je sadržaj za obradu; dozvoljene radnje određujete vi. Ograničite pristup na podatke i radnje potrebne za taj posao.
Also test the messages that disturb the routine
Prepare a sample without actual contact information. Include a simple query, a message requesting two services, an incomplete request, and a message with conflicting deadlines. For each, write down the outcome you expect in advance. If the assistant does not know who to forward the query to, a good outcome is to clearly mark it for manual review.
Measure the time to a usable draft
Compare how well the team works without and with the assistant, including checking and corrections. Keep track of misassignments and how many drafts actually helped. If the team often goes back to the original and rewrites the answer from scratch, narrow the task down to the part that has proven to be reliable.
Further reading: NIST: Generative Artificial Intelligence Risk Profile