Let's talk about shared context layers, often referred to as "company brains"

To answer hard questions about your company like:
"what's true about X, since when, according to who, and where do accounts disagree?"
You need context that is a derived, citable fact layer over all your communications and documents
It's the system of record for what the company or organization "knows" which is different from the systems of record containing the outputs from work that is done
Jack in accounting knows why a vendor invoice was paid that didn't match the PO, whereas the accounting system just contains the data elements - the PO and the payment
So if you want to know "why did we pay a different amount than what was on the PO" you need to go ask Jack
Unless you have a shared context layer that is connected to the Slack and email threads showing Jack confirmed with the business manager that the order was changed and approved the payment
For a shared context layer to be useful, it has to be grounded in evidence, able to surface contradictions, and supersede old truths with new information
It has to answer:
- Who is this person? Jack from accounting, not Jack from Engineering
- What is the current state? PO doesn't match invoice and can't be paid until resolved
- Where do accounts disagree? original PO was approved by Jane in email, but change order was requested by Bob in Slack
- What changed? Bob needed to increase the number of seats to include an additional group
Retrieval-based systems will not be reliable at these questions, they're good at finding text, but not so good at solving claims like why was this invoice paid when it didn't match the PO?
The architecture for a properly functioning shared context layer needs to account for:
1. Evidence - the raw data, sourced
2. Facts - the entities and claims within the company
3. Judgement - the reasoning across evidence and facts
4. Access - what evidence and facts can AI reason over
This is of course a simplified breakdown as the building of a reliable shared context layer can be quite complex, but if you get this right, it serves as the foundation for workflow automation
Jack in Accounting doesn't need to go track down Jane and Bob to determine if the invoice should be paid - an agent can do that and Jack can focus on more strategic activities like planning and analysis
The point is, if you want to automate workflows and deploy agents beyond pilots, you need to first focus on a shared context layer so agents have reliable information to act