Last week a customer showed us a report their agent had produced. They had asked for the average sales cycle, and the number was off by weeks. The agent had access to every opportunity in Salesforce, and every field on those opportunities was described. It had counted the placeholders.
Their Salesforce creates proposed renewal opportunities months ahead of time. Many sit at $0 in an early stage until the renewal is real. Others are deals that were pushed back to Stage 0 but still carry a dollar amount. A rep knows not to count them. The agent didn’t, because nobody had decided what it should receive.
That is the state of most AI context layer work today. Agents can reach the data, and they increasingly understand it. They still give confident answers that are wrong, and the cause is a third level of maturity that most teams haven’t reached yet.
Level 1: The agent can reach the data
Over the past year, access went from a project to a setting. Salesforce made its hosted MCP server generally available, and MCP (the standard way an agent connects to another system) became ordinary infrastructure alongside connectors and integration platforms. Some reach one system and some reach thousands.
This is real progress, and it is no longer where the gap is. An agent with access to everything still has to decide what to look at. Give it an opportunity with 212 fields and it will choose some of them. One might be the field your forecast depends on. Another might be a field nobody has filled in since 2021. Access gets the agent to the record. It does not tell the agent which parts of the record matter.
Level 2: The agent understands what elements of the data mean
This is where most of the industry’s work is happening now. Data catalogs, metadata and warehouse semantic layers describe your fields, define your metrics and serve those definitions to agents.
An agent at Level 2 is better. It knows “Amount” is ACV and not TCV. It knows Stage 3 is not commit. Your definitions are finally somewhere an agent can read them.
But the agent still has to fetch the record itself, decide which related records matter, and assemble the picture from many separate requests. Each request is a chance to guess wrong, and definitions don’t tell it which records to leave out. The customer above was at Level 2. Their placeholders were well described. They were still counted.
Level 3: The agent receives the correct and complete data
Level 3 is about what actually lands in front of the agent when it asks about one account or one deal, as opposed to what it could reach or look up.
Someone who knows the business decides in advance what the agent should see for that kind of record: which fields, which related records, which records to leave out, and which rules apply. Renewal placeholders at $0 are not pipeline. Stage 0 deals that still carry an amount are not pipeline either. The agent then receives all of it in one response, as the signed-in user, with that user’s permissions.
That is where a wrong answer is made or prevented. More access won’t fix the placeholder problem. Better definitions won’t either. Only deciding what the agent receives will.
This is the level the market hasn’t built yet for live Salesforce data, and it is the level we built ContextWorks for.
Which level are you at?
Look at the agents your team uses today and ask one question: when my agent asks about one account, what exactly does it receive?
- “Whatever it queries.” You are at Level 1.
- “Field descriptions and metric definitions.” You are at Level 2.
- “This record, these fields, these related records and these rules, in one response.” You are at Level 3.
Each level needs the one below
You need Level 1 or nothing works. Level 2 improves everything built on top of it. Level 3 is where the agent does its reasoning, so it is where the answer goes right or wrong.
Access was last year’s work and meaning is this year’s. What the agent receives is what will decide whether your company trusts agents with real work on real customers.
Most of the companies we talk to are at Level 1, and many are working on Level 2. Very few are at Level 3, which is why their agents sound confident and still get it wrong. If that sounds like your org, ask the question above about your own agents, then see what an agent receives from ContextWorks.
