AI as a feature is exhausting. AI as a tool is something else. We build dashboards as a two-founder team and the help of AI in places where it actually pays off. Below is the short list of where we let AI lift the work, and the four places where we refuse to.

Where AI earns its keep

1. Schema mapping

Connecting an unfamiliar HubSpot portal to your warehouse used to take half a day. The same job, with an LLM helping to map custom fields against your vocabulary, takes about forty minutes. AI is doing what humans hate: reading other people's column names and proposing reasonable matches. A human signs off.

2. The change tile

Almost every dashboard we ship has a tile that says, in two sentences: here is what moved this week, in plain English. An LLM writes those sentences on a schedule. Without AI, that tile is a meeting. With AI, it is a note that lands in your inbox before standup.

3. Anomaly first-pass

Classical anomaly detection flags too much. An LLM that has seen the last month of values can answer is this worth waking someone up at 2am with more nuance than a z-score. We still send the underlying numbers, but the AI ranks them.

4. The build itself

We use AI to write the boring parts of a dashboard build: scaffolded SQL, named widgets, type-safe API clients. This is what makes a two-founder team competitive with a fifteen person consultancy. It also lets the two of us spend more time on the parts that need taste.

The four places we refuse to use AI

Each of these has a story behind it. None of them are moralistic. They are operational.

1. Numbers on the page

Every number visible on a dashboard is calculated by deterministic code. No exceptions. An LLM can describe a number, label a number, contextualise a number. It does not compute the number. If a finance team cannot trace the value back to a query, the dashboard has failed.

2. Permission decisions

Who sees what, who can write to what, who is allowed in the admin panel. Permission logic is hand-written and reviewed line by line. An LLM that can be talked into a permission change is a security incident waiting to be filed.

3. The first widget the user sees

The top-left tile is the one users learn the dashboard from. It does not get a generative summary. It gets the same data, laid out the same way, every single day. Consistency is the product.

4. The deadline

AI is allowed to comment on whether a deadline is at risk. It is not allowed to move the deadline, soften the deadline, or rewrite the deadline in friendlier terms. A deadline that shifts to spare someone's feelings is no longer a deadline.

The rule that keeps us honest

We use AI everywhere it does grunt work, and nowhere it makes decisions a human has to defend. If the work is read this and propose, AI helps. If the work is read this and decide, a person owns it.

That line moves as the tools get better. We expect to revisit it every six months. We do not expect it to disappear.

A short field test

Before any AI-generated text lands in front of a user, we run it through three questions.

  • Would a careful human, given the same numbers, write roughly the same sentence?
  • If the sentence is wrong, is the cost recoverable inside a day?
  • Does the dashboard still work if we delete the AI part?

Three yeses and the AI ships. Anything else and we cut it. So far, that test has cost us about as many ideas as it has saved us from. We are at peace with the trade.