AI assistance

Reviewing AI suggestions

AI in Kenseme always leaves you the last word, but the way you review its work depends on the feature. There are four patterns. Knowing which one you’re in tells you what’s already been saved and how to take something back.

1. Inline suggestions: accept or reject

The suggestion appears next to the current value, and nothing changes until you choose.

  • Where: Pipeline column mappings after Run AI Rename, where each suggestion reads “Suggest: name (confidence)”.
  • What to do: Click Accept to use the suggestion or Reject to keep your value. Typing your own value works too.

2. Review lists: tick what you want

The AI produces a batch, and you choose which items to keep before anything is saved.

  • Where: the Ontology Generator’s review, the Glossary Extractor’s candidate terms, the Suggest Names dialog for a table and its columns, and the binding plan for lakehouse and database bindings (Review plan, then Apply N bindings).
  • What to do: Read each item, clear the ones you don’t want, and apply the rest. In the database binding plan, Reject unchecked suggestions also dismisses the items you cleared so they don’t come back.

3. Proposal cards: create or apply

A chat answer comes with a card for each object the AI wrote. Nothing changes until you click the card’s button.

  • Where: the schema’s Schema Agent tab.
  • What to do: Read the SQL, then click Create for a new object or Apply for a change to an existing one. The card then shows Created or Applied with a link to open the object.

4. Edits you asked for: review and undo

When you ask an assistant to make a change, it makes it, and you review afterwards.

  • Where: the Ontology Assistant. Before a destructive change it asks you to confirm with Yes or No. Every change appears on the drawer’s Activity tab with Undo and Redo. See Chat with the Ontology Assistant.
  • Also here: description generation writes descriptions directly. Before it overwrites one, such as a schema’s description with Auto Generate Description, the confirmation tells you that the current text will be replaced and can’t be restored.

Note: The Data Agent is different again: it answers questions and never changes anything. It lists the assumptions it made next to each result, and you can change any of them with one click. See Ask the Data Agent.

Confidence scores

Some suggestions show a confidence, such as Pipeline AI renames and binding suggestions. Treat it as the AI’s own estimate, not a guarantee:

  • High — the names and context matched well. Still read it.
  • Middling — a reasonable guess. Check it against what you know.
  • Low — a hint at best.

When not to accept

  • It renames something your team has already agreed on.
  • It breaks your naming conventions.
  • The reasoning is vague or generic, which usually means the AI didn’t have enough context.

Get better suggestions

  • Describe your objects. Tables, columns, and ontology entities with descriptions give the AI far more to work with than cryptic names alone.
  • Set naming conventions for the organization and each schema. See Organization-wide settings.
  • Keep a glossary. A curated glossary gives the AI your organization’s vocabulary, and the Ontology Generator can use it as a source.
  • Give instructions. Most AI features accept guidance, such as the Ontology Generator’s instructions or the Star Schema designer’s AI Guidance (optional).