There is a Slack channel I monitor closely. It is where Remote's accounting team, more than 30 of us, asks questions about our accounting systems and mostly about SAP.
Someone needs an exchange rate. Someone is getting an error. Someone remembers a process exists but cannot find where to run it. Those questions tend to be repetitive over time, and for four years the answer to most of them was often me.
Now the first answer is a bot: a Notion AI agent that sits in the channel, searches past questions and threads and our accounting knowledge database, and replies with links and a suggested next step in around 30 seconds.
I'm Christos Petsis, Associate Accounting Systems Manager at Remote. Part of my role is helping the team use our accounting systems correctly, so when a question came into that channel, I would take the time to go through it, investigate, and respond manually.
That worked, but it depended on my calendar.
If I was working on something else or sitting in a meeting, I would not see the message until later, which meant a slow response. If I was out of office, a colleague would pick it up and the reply could be slower still. Most of these questions were not new either. They had been answered before, in an older thread or in our accounting knowledge database. The answer existed. The problem was that nobody could reach it at the moment they were blocked.
How the bot answers a question
When someone posts, the bot replies immediately to say it is searching past solutions, then comes back around 30 seconds later with an answer, links to the prior questions it drew from, and a proposed next step. A question I answer manually takes me around half an hour on average, so the bot saves close to 29 minutes each time it responds. In practice the saving is larger, because several hours can pass between someone asking and me seeing the message.
In the video I ask the bot for the AUD to GBP exchange rate in SAP as of a given date. It finds related questions from the past, links back to them, and returns a valid answer with a next step. This is the kind of question the team asks regularly, which is the point of building this around past threads.
There is an escalation path as well. If the answer is not satisfactory, the person reacts with a red cross and the bot tags me. Accounting systems are no place for confident guesses, and I would much rather find out quickly that the bot was unhelpful than find out later that someone acted on incorrect information.
What went wrong while I was building it
I wanted to build this for a while, but Notion agents could not work in private Slack channels at first. That limitation is fixed now, though I still had a difficult time connecting the agent to a private channel. It was solved in the end. Then the bot started responding in channels it was only supposed to read. It had access to several of them so it could learn from past threads, and I adjusted the agent rules so that it writes in one channel only.
Thread handling needed the same treatment. The main question is often not in the first post of a thread, it appears in the second or third message, and the bot was not picking those up. I changed the rules so it reads the full thread before deciding what is being asked.
Some things cannot really be fixed, and I am fine with that. A messy question produces a messy answer, and a question that is too vague gives the bot nothing to match against, so it returns nothing at all. A genuinely new question has no past answer to retrieve either. The bot handles the recurring cases, which means the unusual ones reach me with less noise around them.
The next thing I want to build is a dashboard, so I can monitor the agent's success and escalation rate and see where it can improve.
Remote's AI for Actual Work course is about pointing AI at a real workplace problem, and mine was already sitting in a Slack channel: the same SAP questions, slow first responses, and answers scattered between Slack and Notion. What the course made clear is how much of this is setup. Start from a real pain point, give the agent the right context, and keep a human path for anything uncertain.
The SAP Knowledge Bot is simple, but I think it is effective enough. It does not transform accounting. It gives a few dozen people a faster first answer and sends the harder cases to me, which is what I built it to do.
Remote’s AI for Actual Work course is free. Sign up at aiforactualwork.com and learn how to turn a workplace problem into something useful.
