You ask AI to help prepare for a meeting. It gives you a perfectly reasonable agenda for a meeting that bears only a passing resemblance to the one you’re actually having.
So you explain the customer, the project, what happened last month and why the apparently obvious next step won’t work. Then you find the relevant email, copy in a conversation from Slack, Teams or whichever chat tool your organisation uses, and explain which parts of the project plan are still current.
The answer gets better. But assembling the briefing has become a job in itself, and you’ll probably need much of the same background next week.
I think this is a useful point to ask how much of that explanation you should be doing manually. If the information already exists in the tools you use, your AI assistant may be able to find and read it for you.
What context is AI missing?#
A short question often relies on quite a lot of background.
When you ask a colleague to help with a customer forecast, they might already know that a launch has slipped, the customer is changing their ordering pattern and the sales team is still waiting for confirmation of a promotion. You don’t include all of that in the request because you’ve been talking about it for weeks.
Your AI assistant needs a way to get that background too.
The sales data might show a change in orders. The explanation could be in a customer email. An internal Slack discussion might explain why the team thinks the change is temporary. A task in Jira or Monday might show who’s meant to confirm it and whether they’re still waiting for an answer.
Those sources each contribute something useful. Together, they give a much better starting point for deciding what to investigate than the numbers alone.
Saved preferences and project instructions can help with the stable background: what your role is, which customers you work with and where the main information lives. The changing details of the work need somewhere to come from as well.
Reading is useful work#
I recently suggested someone connect their AI assistant to Slack. Their first thought was that it could post updates and communicate on their behalf.
What I had in mind was that it could read the conversations they were already having.
Instead of explaining a request again, they could point to the conversation where it had been discussed and ask the assistant to prepare a plan. They could then check whether it had understood the request before doing the work.
I think this is an easy benefit to overlook. We associate connecting tools with automating actions: sending an email, updating a ticket or posting a message. But finding the right information and working out what it means can take a substantial part of the day.
Claude, for example, can search connected Slack conversations and email. Jira and Monday also provide ways for compatible AI tools to retrieve project information. The connections available to you depend on your platform and your organisation’s setup.
There’s useful work here before you ask the assistant to send or change anything in those systems.
How do I give AI context without copying and pasting everything?#
I’d start with a task you already understand, and point the assistant towards the sources you’d normally use to prepare for it.
For example, imagine you’re preparing for a customer forecast review. You might ask:
Prepare a briefing for next week’s forecast review for this customer. Read the customer email thread about the launch, the recent discussion in the account’s Slack channel and the launch board in Monday. Summarise what has changed, what remains unresolved and which assumptions we should check against the current forecast. Link to the sources and include their dates. Flag any conflicting information or sources you couldn’t access.
Suppose the customer email says the launch has moved by a month, while the project board still shows the original date. The useful output would flag that disagreement and identify the forecast assumption that needs checking.
You can review the evidence and bring a specific question to the meeting. You haven’t had to collect every message yourself, and the assistant hasn’t had to guess why the launch matters.
Giving it a customer, a task and some likely sources also makes the request easier to check. You know roughly what it should find. “Go and understand our whole business” is a rather harder place to start.
Do we need to document everything first?#
You probably have more written down than you think. It may just be spread across several tools and lack anything resembling a sensible filing system.
Customer priorities appear in emails. Reasons for a change turn up in Slack. Unresolved dependencies sit in project updates. Some of the explanation you’d type into a prompt would be a summary of those records anyway.
I’d use that material as a starting point. You can ask the assistant to assemble a first account of what’s happening, then correct what it missed and turn useful findings into documentation people can maintain.
That doesn’t remove the value of an agreed project brief, a current plan or clear definitions of your measures. They make future searches more reliable. But you don’t need to finish reorganising the company’s knowledge before trying a useful task.
AI also can’t recover a decision nobody recorded. If an important agreement only happened in someone’s head, someone still needs to write it down.
How do we know which information to trust?#
A message tells you what someone said at a particular time. It might be a proposal, a concern or a plan that was abandoned the following morning. Even a recent message may be someone’s recollection of an older decision.
Ask the assistant to distinguish confirmed decisions from suggestions and open questions. Check the underlying sources when the distinction matters. If the customer email and project board disagree, establish which reflects the agreed plan. The summary should help you find that question, rather than quietly settle it for you.
Keep important agreements somewhere people can find and maintain them. Finding five versions of a plan quickly doesn’t establish which one everyone has committed to. That’s part of the human work in getting a decision made.
I’d also check the permissions before connecting a source. Use your organisation’s approved setup and individual sign-in where supported. Claude’s connector documentation distinguishes access through an individual’s permissions from custom connections using shared credentials. Know whose access you’re granting.
Choose read access where the connection supports it and that’s all the task requires. A prompt saying “don’t change anything” isn’t a substitute for restricting the connection’s permissions. Existing access deserves a look too: information being easy to find can expose problems that were previously hidden in an obscure folder.
Start with one recurring task#
In that conversation, I could have explained the reading benefit earlier. Sending messages is an obvious use for a connection, but helping someone arrive better informed can be valuable on its own.
Take one meeting or recurring question where you spend time gathering background. Connect the relevant approved sources, point the assistant towards them and ask it to prepare a briefing with links to the evidence.
Check what it found, what it missed and whether it saved you useful time. Then keep the instructions that helped, including where the current plan lives and who can resolve conflicting information.
The next time you ask for help, you should have less explaining to do.
