By Michael Domanic, Section Head of AI

You should start working as if agents are listening.

Most organizations have started realizing they need to hook up connectors and give AI access to everything they do. An LLM becomes 100x more useful if it can answer in the context of your company’s customers, roadmaps, calls, and decisions. 

The problem is that even when you do that, a lot of information isn’t in those connected sources – or isn't structured in a way that AI can actually use. A vague file name, a decision made verbally and never written down, a CRM record nobody updated after the last call: Your agents can’t get value from any of that, no matter how many connectors you add.

This is where behavior comes in. There are four adjustments in how a workforce operates that will meaningfully fortify the context your agents have to work with, on top of everything the connectors are already giving them.

1. Write (and speak) like an agent is reading it, because one will be

Most of what gets written inside a company is written for a person. Nobody writes a Slack message imagining it gets pulled into a context window six months from now to brief an agent on what happened with an account.

That’s a problem, because an agent doesn't have the shared context that a human reader brings. A human reads, “They pushed back on it” and understands who “they” is. An agent can only work with what’s explicit. 

The habit you need to build: Adding context to messages and verbal decisions. Name your files as if someone unfamiliar with the project needs to find them. Say the client's name out loud on the call instead of "they." This behavior allows agents to retrieve and reason over what you wrote. 

2. Capture decisions, not just discussions

A transcript of a meeting is not usually a clear record of what the meeting decided. It's a record of people talking, at times restating the same point four different ways. An agent trying to extract a decision from that has to guess, and a guessing agent is a liability the moment you give it any real autonomy.

The fix is one sentence at the end of every important conversation: we decided X because Y, owner is Z, next review is whatever date makes sense. 

3. Default to public

Direct messages (DMs) in Slack and Teams are dark data, and dark data might as well not exist as far as your agents are concerned. Sometimes that’s a good thing - we don’t want to surface sensitive conversations about employee performance, comp, etc. to the LLM. 

But for information you do want the LLM to see, you need to use channels. The habit of defaulting to public channels is one of the cheapest, highest-leverage changes a company can make. 

4. Keep the systems of record current

Stale data actively pollutes every agent that reads it. An agent pulling from a CRM that is weeks out of date will give you a wrong answer, stated as fact, with no indication that anything is off.

Maintaining your systems of record has always been an important behavior, but it’s even more critical now, because every agent built on top of that system inherits whatever staleness is sitting in it. 

How to turn access into intelligence

Every company is going to have access to roughly the same models and connectors within a year. The difference is whether agents can use that connected data to derive a greater understanding of your company.

Writing for the agent, capturing decisions, defaulting to public channels, and keeping data clean are not abstract good habits. They directly impact how useful your agents are.

The good news is, none of them require new tools. They do require a mindset shift in how your company operates now, and they require holding people accountable in the same way you'd hold them to any other operating standard. Very few companies have done that so far, and that gap is where the supercompany lives.

See you next week,

Michael

Your fellow Head of AI

‍