The devs probably already know this but the way I handle this is to have each ai measure its origin offset from the last frame (use that to figure the actual velocity by dividing the offset by frametime), send that offset to the client if it has a different value than the previous value, along with the current actual origin. Then on the client set that actual velocity as the velocity, and find the offset from the client's current ai origin to the actual origin that was sent. Use that offset to add a secondary velocity to the movement each frame to catch up to where the ai should be, at any designated speed. This should keep your ai very close to where they actually are on the server.
Edit: Also if the actual velocity on the server becomes '0 0 0', you then tell the client the ai is stopping, and where it is stopping, then lerp to that point on the client.