TL;DR: An AI that connects support, sales, product and engineering could see what no single team sees. The hard part isn't the model. It's permissions, sources and privacy, and a debugging problem most teams don't expect.
Support knows what customers complain about. Sales knows which objection stalls the deal. Product knows what got pushed to the next release. Nobody sees all three at once, and the answer to a decision often sits right where they overlap.
Working with LLMs, RAG and vector search in products, I started thinking about an AI that knows the company, not just the product. One that connects support, sales, product, engineering, meetings and documentation, and hands context to whoever has to decide, with the source of every claim right next to it.
In most companies, that information reaches decision makers through summaries, reports and slide decks. Often nobody got it wrong. The information just didn't survive the trip.
Picture asking "how do we sell more in this segment?" and getting an answer that combines objections from sales calls, customer requests, support tickets and what the latest release already fixes. The most important insight might live in the relationship between departments.
The technology exists. Governance is the hard part
- Permissions follow the source. If you can't open a document in the original system, the AI shouldn't bring its content to you either.
- No source, no answer. Every important claim points to the passage it came from.
- Privacy law is part of the design. In Brazil that means the LGPD: purpose, necessity, a legal basis and clear rules before capturing meeting data or other sensitive sources.
- Aggregate analysis, no surveillance. "Which objections come up most?" helps the business. "Which salesperson is the worst?" is a different system.
Who's asking? How I'd sketch the retrieval side
- Store every chunk with its source ID and the access list from the original system.
- Filter by the identity of the person asking at query time, before ranking. Filtering after the model answers is too late.
- Permissions change. An index that syncs access lists once a night will show yesterday's permissions all day, so sensitive sources need a check at read time.
- Log what was retrieved: document IDs, scores and which filters applied. Not the content.
- Let the teams that own the data own the evaluation set.
A less obvious problem
A system like this can retrieve information that even the people who built it aren't allowed to read. How do you improve a system that can access what you yourself can't open?
That's where the logs above matter. Builders debug through what was retrieved, by IDs, metadata and relevance scores, while the data owners write the test questions and judge the answers.
The permission mess was already there
In a 2024 Gartner survey of 132 IT leaders, 40% said data oversharing led them to delay their Microsoft 365 Copilot rollout by three months or more. The permission problems were already there. The AI just made them easier to find.
The opposite risk
A study commissioned by AWS from Strand Partners, released this September, found that 50% of Brazilian companies already use AI, but only 15% of those have reached an advanced stage. Being so cautious that the company never learns to use it is expensive too.
Starting with the least sensitive source, understanding what the system actually needs to do and expanding gradually seems like the safer way to experiment.
So here's the question I keep coming back to: would an AI that reads the whole company help us see what no single team can see, or would it just amplify the permission problems the company already has?
Sources: Gartner survey reported by Computerworld; AWS/Strand Partners study reported by About Amazon Brasil.
Top comments (0)