Your AI Agent Doesn't Need More Data. It Needs Context.

Your AI agent is not stupid. It is guessing, confidently, because nobody ever told it what your numbers actually mean.

It is Sunday night in Toronto, and Priya is doing the report herself, again

Priya runs a two-person wedding photography studio. Bookings live in a Google Sheet. Deposits land in Square. Client messages are split across three inboxes.

Every Sunday she pulls the same numbers by hand: how many inquiries this month, how many turned into a booking, which referral source is actually paying off.

She finally tried an AI agent to save herself the hour. She asked it one simple question: "Which month had the most cancellations this year?" It answered instantly, specifically, and confidently. It was also wrong, off by three months, and it missed a January refund that should have counted.

The agent was not broken. It had access to the spreadsheet. It just had no idea what "a cancellation" meant in Priya's business, which tab was current, or that she had corrected a data-entry mistake back in March.

Here is how reporting works right now, for almost every small business

• One person, usually the owner, is the only one who knows which spreadsheet is "the real one."
• Definitions live in someone's head: what counts as a lead, a cancellation, a returning client.
• Every report means stopping other work to pull the numbers by hand, again.
• Connecting an AI tool to that same data does not fix any of this. It just makes the guessing faster.

You are not missing an AI tool. You are missing a decision: what your numbers mean, and where the true version of each one actually lives.

What is AI agent context, and why does it matter?

Context is everything an agent needs to know before it can be trusted with the data in front of it: your definitions, your rules, and which source wins when two disagree. Without it, an agent connected to your spreadsheet is just a faster way to guess.

Nate Herk, who teaches AI automation on YouTube to a fast-growing small-business audience, splits this into two buckets worth stealing for your own setup:

• Expertise context, the things that are true in every conversation: your definitions, your policies, who you are. This gets loaded every single time, like a rulebook the agent never forgets.
• Situational context, the things that are true right now: this week's bookings, one client's message thread, yesterday's sales. This gets fetched fresh, only when it is actually needed.

Mix the two up and you get exactly what happened to Priya. Old situational data, March's uncorrected numbers, gets treated as permanent expertise, and the agent repeats it with total confidence.

Why this is suddenly worth fixing this month

LangChain, the company behind one of the most widely used AI agent frameworks, published a detailed account of rebuilding its own internal reporting around a data agent. Before, nearly every data question went through one person on a small team. After giving the agent clear metric definitions, a semantic model of the business, and explicit rules about which source to trust, adoption changed fast: nearly all provisioned employees used it inside 30 days, generating roughly 2,200 agent conversations, an average of 23 questions per person per month.

That is not a story about a bigger AI model. It is a story about a small team doing less manual reporting and more system-building, because the agent finally had the context to be trusted with the question.

Isn't this just a chatbot connected to my spreadsheet?

No, and this is the part worth sitting with. A chatbot with spreadsheet access will answer any question you ask, correctly or not, because nothing stops it from guessing at a definition it was never given.

A data agent with real context does three things a plain chatbot does not: it knows which source is the truth when two disagree, it knows your specific definitions instead of generic ones, and it gets checked and refreshed on a schedule, so it cannot quietly drift for months without anyone noticing.

How do you build this with no data team?

You do not need LangChain's engineering budget. You need five decisions, most of which take an afternoon, not a sprint.

1. Write down your five most-asked numbers, in plain English. What counts as a booking. What counts as a cancellation versus a reschedule. What counts as a returning client. Priya's fix took eleven minutes: a cancellation is a client who paid a deposit and did not shoot, full stop, refund status is tracked separately.
2. Name the one true source for each number. If bookings live in a Google Sheet and payments live in Square, decide out loud which one wins when they disagree, and tell your agent that rule directly, in writing.
3. Split what the agent should always know from what it should look up. Your definitions and pricing rules are expertise context, load them every time. This week's bookings are situational context, fetch them fresh each time someone asks.
4. Connect the agent to the live source, not a copy. A pasted-in export from three weeks ago is exactly how two versions of the truth end up disagreeing, with the agent picking one at random.
5. Put a ten-minute audit on your calendar every week. Ask the agent what it thinks your numbers mean, and check it against reality. Nate Herk runs this as a scheduled check on his own systems specifically because stale context is the single most common reason an agent that used to work well "suddenly" starts getting things wrong.

What you have after this

A week in, your five definitions are written down, and your agent stops making up its own version of "cancellation."

A month in, the person who used to spend Sunday night pulling numbers is spending it doing something else, because anyone on the team can ask the question directly.

Six months in, you have a system that gets more useful the longer you run it, instead of one that quietly rots the way an unaudited spreadsheet does.

Same data. Same business. A machine that finally knows what your numbers mean.