If the AI counts the number of people going in and coming out, the only information it would be able to extract is the average time a client takes in a meeting. For more precision, how about recognizing the clients and categorizing them. For example, if you could categorize them by age and then calculate the average time spent by an age class in the meeting, you could (slightly more) precisely determine how much time an appointment with a future person of the same age class might take.
Also, if there is a microphone that recognizes the voices in the meeting, it could be able to divide the time of the meeting into that spoken by the client vs the officer. If the conversation is highly confidential, a simple wavelength sensor could be installed that cannot identify the words but can identify who is speaking. The time of the meeting is noted. When the same client books another appointment, the time of the meeting could be more precisely predicted.
The meetings could be further categorized by the reason of the meeting. Some reasons might take more time than others. For example, "meeting a new client to listen to their terms" might take more time than a "casual in-office lunch with a friend".
If an important client, whom we do not want to lose, is consistently late for the meetings, that client's meetings could be scheduled 15 minutes or half an hour early (depending upon how late the client usually is), overlapping with the previous client, thus, saving time.
Also, all these inputs when used together could lead to another level of precision. Using these inputs, finer categories could be made, for example, "a 40-year-old male coming for a casual meeting" or "a 30-year-old woman coming with a business proposal". Based on the "appointment time" history of these categories and the personal history of each client, the future appointment time could be precisely predicted.