Smart Paste in Dynamics 365 Sales parses unstructured text a user has copied — an email, an RFP, a document extract — and maps what it finds onto the matching Dataverse fields for review. It proposes values; the user confirms them before anything is saved.
The problem it addresses
CRM data quality does not fail because people disagree about its importance. It fails because filling twelve fields from an email is tedious, and a rep with a call in four minutes fills three of them.
Everything downstream inherits that: forecasting, segmentation, routing, and every report anyone builds on top. Removing the tedium is a data quality intervention dressed as a convenience feature.
How it works
Form Fill Assist handles the pasted-text case on a form. For documents — scanned purchase orders, invoices, structured PDFs — Azure AI Document Intelligence does the extraction, and the mapping into Dataverse is configured against the model's output fields.
What extracts reliably, and what does not
| Content | Reliability | Handling |
|---|---|---|
| Names, email addresses, phone numbers, job titles | High — structurally distinctive | Accept with a glance. |
| Company names and addresses | High | Check against existing accounts to avoid duplicates. |
| Dates and deadlines | Moderate — relative dates are ambiguous | Verify anything that drives a task or reminder. |
| Commercial terms stated in prose | Low | Read it. Do not accept a value you have not confirmed. |
| Implied or inferred information | Not extraction — inference | Out of scope. Treat any such value as unsourced. |
The distinction that matters is between reading a value out of the text and producing one that fits. The first is extraction. The second is a guess with good formatting.
Where the data goes
Smart Paste operates within the Microsoft Azure compliance boundary. Content stays inside your tenant and is not used to train public foundation models. For most procurement reviews that is the whole question, and it is worth having the answer ready with the documentation reference rather than reconstructing it later.
What still needs a local decision is which record types it is enabled on. A feature that reads whatever a user pastes is one a user can paste the wrong thing into, and some record types warrant more care than others.
Rolling it out
- Enable it on lead and contact forms first — highest volume, lowest risk.
- Watch what gets corrected in the first fortnight; that tells you where the mapping is wrong.
- Extend to opportunity forms only once the mapping is settled.
- Keep human confirmation on. There is no version of this where skipping review is the right call.