AI suggestions
An assistant is a model the firm has approved to look at one kind of record and propose improvements: a corrected counterparty name, a tidier address, a likelier classification. Insight › AI suggestions is where those proposals arrive for a person to decide.
The principle: assistants propose, the queue decides
An assistant produces suggestions into a queue that already exists. Nothing applies one on its own.
- There is no auto-apply setting. No configuration switches the queue off, and no confidence score turns a suggestion into a write. A corrected counterparty name sits next to the data a conflicts engine reads, and Datum issues no clearance decision (ADR-0011). A person looks at every suggestion.
- Accepting records your decision. It does not change the record. The screen says so: "Accepting records your decision. It does not change the record — that is a separate, versioned write." The value reaches the record through the normal write path, with its own version, rationale and audit entry, exactly as if you had typed it.
- Walls are respected. An assistant only ever gathers records through the same reads a person has, as a system with no special power. A record screened from a person cannot be sent to a model and cannot have a suggestion made about it.
- Personal data is refused by default. An attribute classed Personal or Special is neither sent as context nor targeted, unless the firm has deliberately switched that guardrail on.
For the decision behind this design, see ADR-0033, assistants propose, the queue decides.
The screen
In the demonstration firm no assistant has been configured, so nothing is proposed.
Use Suggestion type at the top to narrow to one kind of suggestion, or leave it on All.
When the screen is empty
The empty state is deliberate and distinguishes two cases that look the same:
- "Nothing is proposed. An assistant watches an entity and proposes into this queue; configure one under Administration › AI › Assistants, and enable it when you are ready for record content to reach the model." Nothing is configured to look. This is not the same as the assistant finding nothing.
- If your session cannot decide suggestions, the message says that accepting and rejecting needs the stewardship class, and that the proposals are readable either way.
An assistant is created disabled. Enabling it is a deliberate step that needs an administrator to re-authenticate, because from that moment the configured attributes reach a processor. Disabling also asks, so that an assistant a firm believes is watching its data cannot be quietly stopped.
When there are suggestions
Each suggestion shows:
| Field | Meaning |
|---|---|
| Suggestion / Proposal | What is being suggested, and the new value. |
| Current | What the record says now, so you can compare. |
| Proposed by | The assistant, and the Model behind it. |
| Why | The model's stated rationale. |
| The model was shown | Exactly which attributes were sent as context. |
| Confidence | How sure the model says it is. Treat it as a hint, not a verdict. |
Choose Accept or Reject. Along with the list, the screen reports how often your stewards agree: "Stewards accept N% of suggestions". A low rate means the assistant needs configuring again, not that stewards should accept more.
A high confidence on a wrong suggestion is still wrong. Read the Current value and the Why before you accept, as you would a colleague's proposed correction.
What is counted and what is thrown away
A run of an assistant records how many records it considered, how many it suggested, how many fell below the confidence floor and how many were invalid (the model returned something the schema refused). These counts, shown on the administration side, let you tell a well-calibrated assistant on a quiet day from one whose every proposal is being discarded.
Who can do what
| You can… | Steward | Administrator |
|---|---|---|
| Read the suggestions | Yes | Only where a role reads records |
| Accept or Reject | Yes | No (needs the stewardship class) |
| Configure, enable or disable an assistant | No | Yes, with re-authentication |
Related
- Dashboards — the acceptance rate and steward decisions are metrics too.
- Match thresholds — duplicate detection is the match model's job, not an assistant's.