Skip to content

Connect an AI assistant to the information you choose

Start with a question your team asks regularly, choose the emails, documents or data needed to answer it and connect a compatible MCP or REST client. Preview the shared information and keep important answers open to review.

Supplier update

Delivery is planned for 18 October.

Supplier
Cedar Works
Internal price
£12,400

Selected AI context

Supplier
Cedar Works
Delivery date
18 October
Internal price
Excluded
Original files
Not included

Let the assistant read the delivery date while prices and original files stay private.

The AI context includes the delivery date; internal price and original files are excluded.

Readiness checklist

Information for the question
Choose the email accounts, files and pages needed, Open examples to check the available text
What the assistant can read
Choose fields and masking for the shared view, Choose supporting references and original downloads separately, Give the connection the appropriate token and capabilities
Connect your client
Enable the MCP or REST connection the client supports, Choose purpose requirements and retrieval limits
Freshness and review
Check whether the assistant uses current information or a saved version, Review important values and answers against their sources

Connection checks

Starter use cases

Start with the questions people keep asking

Write down a recurring question such as which supplier delivery dates changed or which invoices are due next. Identify the records and fields needed to answer it.

Choose whether the assistant needs a few fields, readable document text, supporting references or original files. Share only what helps with the question.

  • Questions about status, such as what is open, overdue, or waiting on a reply
  • Questions about evidence, such as where a promise, date, or obligation came from
  • Questions about change, such as what updated in a thread, file, or monitored website
  • Questions about handoff, such as what a new owner needs to know after offboarding or reassignment

Bring in the emails and files needed for the question

Connect the accounts and folders you’re authorized to use. Include the messages and documents needed for the question; import an archive for earlier history and use active connections for later updates.

Open examples to see which text and fields are available before sharing them with the assistant.

  • Live inboxes and shared mailboxes where new work arrives
  • Hosted intake addresses that receive requests, notices, or third-party updates
  • Mailbox exports used for account handovers, investigations, or offboarding
  • Selected drive folders, blob storage roots, or SFTP directories with supporting documents
  • iCal feeds and supported calendar events derived from ICS attachments
  • Websites and portals that publish notices, policy updates, or changing requirements

Choose the information the assistant can read

Create a shared view of the chosen emails, documents, extracted data or current profiles. In Publication settings, choose Included fields, apply any masking and enable the MCP or REST connection your client supports.

Choose separately whether the assistant can open supporting references or download original files. Set purpose recording and retrieval limits to suit its task.

When source evidence is needed

Include permitted references to the email or document behind a value so a reviewer can check it. Decide separately whether the connection can download the original file.

When selected fields are enough

For invoice due dates, share the invoice references, amounts and dates. Keep unrelated details private and turn off original downloads when the assistant only needs those fields.

Preview the connection using the assistant’s credentials

Use the intended token to search and read the shared view. Check its visible fields, supporting references and original-file access, with your sharing policies applied.

Choose an AI-context token for one shared view or a catalog token for the eligible views your assistant needs to discover. An existing API-only token needs the appropriate context capabilities before it can use these tools.

  • The shared views and fields needed for the question
  • Supporting references and original downloads the assistant may receive
  • A compatible MCP or REST connection and credentials
  • The AI provider’s data-handling terms and your team’s review requirements

Choose how the assistant receives updated information

Use the latest captured information for a current working view, a schedule for regular updates or a fixed snapshot for a particular review. Capture and refresh details show how recent each version is.

Choose a refresh schedule when new information should reach the assistant regularly. Set up alerts separately when a change should also notify a person.

Review answers and important fields

Check fields used for deadlines, reporting, or customer communication against source evidence. Review and correct extracted working values when needed.

Use the sources and reviewed values to check the assistant’s answer. Review your external provider’s retention and training terms alongside the information you choose to share.

Test one connection and one question

Connect a compatible REST client or an MCP client that can send bearer credentials. Try a known question and check the returned fields, supporting references and freshness information.

To ask questions inside Polytrace, an administrator connects your organization’s AI provider and model. Where the feature is enabled, Ask this publication uses your signed-in access and the same permitted searches and calculations.

Common questions

Can I use an external AI client or ask inside Polytrace?

Use compatible external clients through the enabled MCP or REST connection with scoped credentials. For questions inside Polytrace, the optional assistant needs deployment enablement, an eligible shared view and your organization’s connected AI provider.

Does exporting context require a Polytrace AI provider connection?

Sharing read-only information through MCP or REST does not call a model in Polytrace. The in-product assistant and model extraction use your organization’s connected provider.

When is human review worth keeping in the process?

Check important source values and answers before using them for a deadline, customer response, compliance decision, or external report.

Can I connect it directly to ChatGPT?

Use a compatible client that can send bearer credentials. Direct ChatGPT installation through OAuth needs a separate authorization integration in your deployment; connecting Gmail or enabling company SSO does not supply it.

Try the question your team wants an assistant to answer

Choose its information, preview what the assistant can read and check the answer using the intended connection.