# Use AI as a classifier

> Generative AI is most reliable in Microsoft 365 and Dynamics 365 when it picks a category, fills a metadata column or routes work, and least reliable when it rewords text that people then act on.

Source: https://ali.id.au/blog/use-ai-as-a-classifier/ (last edited 2026-10-16)

AI is reasonably accurate at picking a category or an urgency level: in my experience, right about 19 times in 20. When it rewords original text, it can introduce subtle data loss or errors. The occasional junk is foundational to how the maths works, so I don't think that problem goes away with the next model.

## What it's good at

* **Triage.** Read incoming tickets and emails, assign them to the right person, and raise critical ones to high priority. It augments the person doing triage; it doesn't substitute for them.
* **Metadata.** Fill in the columns that [replace folder paths](/blog/replacing-your-file-server-with-sharepoint/#dont-copy-the-mess) and drive search, filtering and navigation. Tagging is what turns a library into something you can browse with refiners.
* **Review flags.** Audit policy content and flag what's due for review.
* **Extraction.** Pull a date, an amount or a party name out of a document into a column.

## Where to find it in Microsoft 365 and Dynamics 365

* **SharePoint autofill columns**, the one people forget. In a library, add a column and choose Autofill > [Set up](https://learn.microsoft.com/microsoft-365/documentprocessing/autofill-setup). You write a prompt per column, and the answer for each file lands in that column. They work with choice and managed metadata columns, so the answer comes from your list ([overview](https://learn.microsoft.com/microsoft-365/documentprocessing/autofill-overview)). With Microsoft 365 Copilot licences, autofill is included in [Copilot in SharePoint](https://learn.microsoft.com/sharepoint/copilot-in-sharepoint-get-started), formerly Knowledge Agent. Real-time autofill across a whole site costs Copilot Credits. Without Copilot, it's pay-as-you-go [document processing](https://learn.microsoft.com/microsoft-365/documentprocessing/syntex-licensing), the old Syntex, and so is [taxonomy tagging](https://learn.microsoft.com/microsoft-365/documentprocessing/taxonomy-tagging-overview) against your term store.
* **AI Builder in Power Automate.** [Category classification](https://learn.microsoft.com/ai-builder/text-classification-overview) comes as a [standard model](https://learn.microsoft.com/ai-builder/prebuilt-category-classification) that sorts customer feedback into six fixed categories, or a custom one trained on your own labelled data. Either returns a [confidence score](https://learn.microsoft.com/ai-builder/text-classification-model-in-flow) with each label. A [prompt](https://learn.microsoft.com/ai-builder/use-a-custom-prompt-in-flow) does the same job when you can't label enough examples. Check the licensing: new customers [buy Copilot Credits](https://learn.microsoft.com/ai-builder/administer-licensing), and new or renewed licences stop including AI Builder credits from 1 November 2026.
* **Dynamics 365 Customer Service.** Unified routing can attach skills to incoming work with the [intelligent skill finder](https://learn.microsoft.com/dynamics365/customer-service/administer/set-up-isf-model), an AI Builder classification model, and Copilot [labels incoming emails](https://learn.microsoft.com/dynamics365/customer-service/implement/faq-responsible-ai-copilot) as positive, negative or neutral. For drafting, Microsoft says to ["always review the response"](https://learn.microsoft.com/dynamics365/contact-center/use/use-copilot-email) before a Copilot-drafted email goes to a customer.

## Rules that keep it honest

* **Give it a fixed list.** A choice column, a term set or a set of queues. A free-text answer is back to rewording.
* **Use the confidence score.** Below your threshold, a person decides.
* **Keep the original.** The label sits next to the source text. It never replaces it.
* **Design the categories yourself.** Machine learning can't yet work out what's important to your business. AI-made structure is better than the mess most organisations have, but it's not as good as structure people designed. It's why I think information architecture will be one of the last jobs to go.
* **Treat the content as untrusted.** Text inside a document or email can carry instructions, and there's [no fully safe way](https://role-confusion.github.io/) to screen for that. Don't let a label trigger anything you can't undo.
* **Don't send unreviewed AI output to people**, internally or externally. Drafts are fine; a person checks them and sends them.