> This is the Markdown of a Backed Help Centre page, https://admin.backed.giving/help/integrations/ai-assistance.md. Every page is listed at https://admin.backed.giving/help/llms.txt. > This page uses Backed's default terms, such as Funding Round and Application Template. An organisation may rename them, so read each as your organisation's own term if you've renamed it. AI assistance is where your organisation decides whether Backed may use an AI model on its behalf, and for what. Every capability is off until an Administrator switches it on. Set it up under `Settings > Profile and terms`, on the `AI assistance` tab. ## Before you start You need the `Can access Organisation settings` permission to see the tab, its `Usage` and its `Privacy statement`, and the permission to manage integrations to switch a capability on or off. The `Decision log` is read with the same permission as the organisation's audit log, `Can access Organisation settings`, and your Data Access Rules decide which of its entries you see. ## The capabilities Each capability is switched on and off on its own: - `Assessor aids` - `Administrator drafting` - `Applicant help` - `MCP access` - `Form drafting` - `Acquittal comparison` Nothing calls a model for a capability that is off. ## Switch a capability on 1. Go to `Settings > Profile and terms` and open the `AI assistance` tab. 2. Turn on the capability. A terms dialog opens. It shows: - the AI provider Backed's operator has set up for your site, and where it processes requests; - that the provider does not train its models on your organisation's inputs or outputs; - the capability's model card: its version, when it was published and last evaluated, the model and provider, where requests are processed, what it is for, what it is given, what it never receives, what it never does, where a person approves its output, and its limits. 3. Read the terms and accept them on behalf of your organisation. Backed records who accepted and when. Turning a capability off takes effect straight away for new requests. A request already running is allowed to finish. Every change is recorded in the audit log as `ai.capability_changed`. ## What the model never sees Every request passes through a data boundary before it leaves Backed. Unless your organisation has allowed a category for that capability, the boundary removes: - uploaded file content; - bank account details; - tax and business identifiers: ABN, ACN, NZBN, IRD and TFN; - contact details: email addresses, phone numbers and postal addresses. Form answers are sent only for `Assessor aids`, `Administrator drafting` and `Applicant help`. `Form drafting` receives the text of the guideline document your organisation chooses for it. That is the only file content any capability receives. Files applicants upload are never sent. ## Usage and the audit log Every model call is recorded in the audit log as `ai.call`. The entry names the capability, the record it was about, who asked, the model, how many tokens went in and came out, and which data categories the boundary applied. It never holds the prompt or the answer. The `Usage` tab counts calls by month and by capability. ## The decision log The `Decision log` tab lists every AI output, newest first, 50 at a time. Select `Load more` for the next 50. Each entry shows: - when the output was made, and which capability made it; - the record it was made for, with that record's version where it has one; - the model and the exact model version, and the prompt template and its version; - what happened to it: `Accepted`, `Edited`, `Noted` or `Discarded` by a person, `Awaiting a person`, or `Not used` once 30 days pass with nobody acting on it; - who acted on it, and when; - which data categories the boundary withheld; - a short form of the input and output hashes. Hover over one to see the whole hash. The log keeps references, versions and hashes. It never holds the prompt or the output, so it can show that an output was used without keeping what it said. To narrow the list, choose a capability, type a record as its type and id (for example `Assessment 123`), choose the person who acted, or set a date range, then select `Apply filters`. `Clear filters` shows everything again. `Export CSV` downloads the entries that match the filters you applied. No one can edit or delete an entry. Who can read the log, and which entries they see, follows the same permissions and Data Access Rules as the audit log. When an accepted or edited AI output changes a record, the record's history and the audit log show an `AI assisted` line. Select `View the decision log entry` to open the `Decision log` tab. From the audit log entry's details, it opens on that one entry; select `Show every entry` to see the whole log again. From a record's history or the audit log list, it opens filtered to that record; select `Clear filters` to see the whole log. ## The privacy statement The `Privacy statement` tab holds a statement about automated decisions for your organisation's privacy policy. Backed writes it from the capabilities you have switched on, and writes it again whenever one changes. It shows when it was last written and the capabilities it covers, and says: - the kinds of personal information the capabilities use; - the kinds of decision each capability helps with, and where a person approves its output; - that no decision is made by a computer program alone. With no capability on, the tab says so, and the statement says that your organisation does not use AI assistance in its grant-making. While Legal Counsel has not yet ruled on whether the Privacy Act's automated-decision provisions apply to grant assessment, the tab shows a `Legal review pending` note. Ask your own adviser to read the statement before you add it to your privacy policy. Select `Copy statement` to copy the whole statement as plain text, ready to paste. ## Your public AI model cards page Your organisation's applicant portal has a public `AI model cards` page at `/ai/model-cards`. Anyone can read it without signing in. It shows one card for each capability, with the same facts as the terms dialog, and marks each card `In use` or `Not in use` for your organisation. Link to it from your privacy policy beside the statement, so applicants can see what each capability does and where a person decides. ## MCP access `MCP access` is the same switch as `Allow AI assistants to connect (MCP)` on the `API & MCP` tab. Changing either one changes both. Connected assistants only read your data; Backed calls no model for them. See [Connect an AI assistant with MCP](mcp.md). ## Administrator drafting Drafting outcome report narratives runs under `Administrator drafting`, and so do `Draft with AI` on a decision's personal message and letter and on a conversation reply, and `Draft summary` on a data report (see [Drafting with AI](../administrator/decisions/ai-drafting.md)). The `Tone notes` you set here guide each draft's voice. While it is off, the drafting controls are hidden and a drafting request is refused. ## When a switch is unavailable - **No provider is set up for your site.** The terms dialog has no provider to show, so a capability can't be switched on. Backed's operator chooses the provider for each site; ask your Backed contact. - **The capability has no published model card.** A capability can't be switched on until its model card is published. It becomes available once it is. ## See also - [Connect an AI assistant with MCP](mcp.md) - [Integrations home](integrations-home.md) - [Drafting with AI](../administrator/decisions/ai-drafting.md)