The cost depends on the task the agent takes on and the data and interfaces you already have. Separate one-off integration and testing from running costs. Include model and channel fees, operations, human review and work on exceptions. Count the benefits you can observe against the current workflow.
A first calculation
Suppose a merchant saves six minutes of service time on each of 100 tasks a month. That is ten hours. At €30 an hour, the time is worth €300 in the calculation. If model use, operations and extra reviews cost €180, that leaves €120 per month before setup. Under these assumptions, an integration costing €1,200 would pay back in ten months.
These figures are illustrative. Saved time has economic value if the team can put it to other work or reduce its workload. Count the extra work from clarification and mistaken orders too.
When do sales enter the calculation?
If the pilot produces additional orders, use the extra contribution margin after fees, shipping and expected returns. Check how many of those orders would have happened through another channel anyway. Decide in advance what improvement would justify expanding the pilot and what error rate would make you pause it.