How the platform uses AI

This page explains, in plain terms, where the platform uses AI, what leaves the platform when it does, what never does, and what stays your responsibility. It describes the platform as built; where a setting depends on how Nia has configured its environment we say so.

The short version

• AI on this platform proposes; people decide. Extracted indicators, reconstructed figures and suggested actions are all reviewed by a person before they count. Advisor answers are shown to you straight away, with their citations, as advice for you to check — they never change your data.
• Spreadsheets, CSV, Policy Hub documents and Advisor attachments are always read on the platform. PDF, image and scanned files, Word files with no readable text, and any PDF or Word data submission in which the platform's own reader finds no reportable figures are sent to the external parsing service (section 2). Only if the platform's own reader finds nothing in a sheet may a bounded portion of that sheet's text be sent to an AI model (section 1).
• PDFs, images and scanned files are sent to one external parsing service, with caching switched off.
• Text sent to AI models goes only to named hosts that do not store or train on it; for every request, including embeddings, only to hosts with a zero-data-retention policy.
• Nothing you upload is used to train any model — not by Nia, and not by the services named below, under their published policies.
• Usage is capped. When your plan's monthly allowance is reached, AI work stops rather than continuing at cost.
• No decision about you is made solely by a machine.

1. What happens to a document you upload

Excel and CSV files are read by the platform's own, rule-based extractor. The file is not sent to any AI service to be read. If a sheet yields nothing, the platform may send a bounded portion of that sheet's text (at most 12,000 characters per sheet, a limited number of sheets, within a time budget) to an AI model to find the indicators it asks for.

Word files are read on the platform first. A Word file with no readable text (for example, a document made entirely of images) is treated like a scan, and a Word or PDF data submission in which the platform's own reader finds no reportable figures is also sent to the parsing service so the figures can be found in the narrative.

PDFs, images and scans cannot be read reliably by rules alone. The whole file is sent to an external document-parsing service, LlamaParse by LlamaIndex, which returns text and tables. The platform sends every job with the "do not cache" instruction, and LlamaIndex's published policy is that files are used only to return your result, never for model training. LlamaIndex offers processing in Europe as well as the United States; the platform uses the European region, so files are processed in the European Union.

For investee and fund submissions the order is strict: the platform's rule-based reader goes first, and a file it can read is never sent anywhere. Only a scan, a photo, a PDF whose columns cannot be told apart, or a prose-only Word document goes to the parsing service.

After parsing, the platform maps the returned tables to indicators using rules, not AI. If that produces nothing, the document's text may be sent to an AI model to extract indicators from prose.

The platform keeps a copy of the parsed text and tables, keyed to your organisation, so that re-uploading an identical file does not need a second parse.

2. Figures reconstructed by AI are flagged, and must be acknowledged

When a submission's figures were reconstructed from a scan or prose rather than read from cells, the platform marks the whole submission as AI-reconstructed. The mark is written into the data-quality record, shown on the review screen, and the submission cannot be approved until a named person records a written reason for accepting it. The same rule applies to fund-level data. The flag is recorded per file, not per individual figure.

3. The Advisor and Policy Review

The Advisor answers only from material it has retrieved: Nia's knowledge base and, if you have uploaded them, your organisation's own documents. It has no web access and is instructed not to use outside knowledge. What is sent to the model for a question is: your message, the passages retrieved for it, earlier turns of the conversation, any document you attached to that conversation (in full), and — when you ask about your portfolio — a read-only summary of your own submission status, indicator trends and data-quality counts, computed by the platform. Retrieval is confined to your organisation's data.

Documents you add to the Policy Hub or attach to a conversation are read on the platform and split into passages; each passage is sent to an embedding model so it can be found again. Every embedding is created on a zero-retention endpoint at a named host that does not store or train on the input. Policy Review sends the policy's text to an AI model to compare it against Nia's best-practice guidance; the result is a report for you to read, not a change to your documents.

Conversation history is kept so you can return to it. The platform's retention setting for conversations is 12 months. The automated purge that enforces it runs daily.

4. Where AI text goes — the model hosts

All AI model calls (other than document parsing) are routed through OpenRouter. Every request carries three restrictions written into the platform:

1. A fixed list of allowed hosts — Anthropic, Google Vertex AI, Amazon Bedrock, Microsoft Azure and OpenAI. A request that cannot be served by one of them fails; it is never sent elsewhere.
2. "Deny data collection" — no host that stores prompts or trains on them may be used.
3. Zero data retention — every request, including embeddings, may only be served by an endpoint with a published zero-retention policy.

Today the Claude models are served on Amazon Bedrock and Google Vertex AI zero-retention endpoints, and embeddings on Microsoft Azure, the one zero-retention endpoint listed for the embedding model. Anthropic's own service and OpenAI's own service are permitted hosts but receive no requests, because neither lists a zero-retention endpoint for the models the platform uses.

OpenRouter's own policy is that it does not retain prompts unless the account opts in to logging.

Models in use today are Anthropic's Claude models (Claude Haiku 4.5 for extraction, quality checks, the Advisor and guidance drafting; Claude Sonnet 4 for report narratives) and an OpenAI embedding model (text-embedding-3-small). Nia can switch a task to a Google Gemini or OpenAI GPT model only from a short vetted list; this page is updated if that happens.

5. What never happens

• Your files, figures and conversations are not used to train models — the platform sends nothing for training, and the named services state that they do not train on customer data.
• No file that the platform can read itself is sent to an external service.
• No AI output is written into your submissions, reports, indicator dictionary or action tracker without a person approving it. (Advisor answers are stored in your conversation history so you can return to them; they do not change any of that data.)
• No decision with legal or similarly significant effect on anyone is made solely by automated processing. AI on this platform assists with reading documents, drafting and answering questions; approvals, submissions and reports remain human decisions.

6. Usage limits

Each metered AI operation reserves one "run" from your plan's monthly allowance before it starts. At the cap, the operation is refused and nothing is sent to a model. See Usage Caps for what counts as a run.

7. Your responsibility

AI output on this platform is assistance. The fund remains responsible for the data it submits, approves and reports. Check reconstructed figures against the source document, read Advisor answers with their citations, and record your reason when you accept an AI-reconstructed submission — that reason is part of the audit record.

8. Who processes your data

The services named on this page — LlamaIndex (document parsing), OpenRouter and the five model hosts above (AI models and embeddings) — act as processors for your organisation's content. The full list of sub-processors, with regions and safeguards, is published on our sub-processor list and summarised in the Privacy Policy. Nia has accepted each provider's standard data-processing terms as at 4 September 2026. Provision to be finalised in a later version

Who we are and how to reach us

Nia Impact (trading as Nia Impact Invest) · company number 221851
Lot 02, Floor 1, CentrePoint, Trianon, Mauritius

Legal and privacy: sarah@niaimpactinvest.com · Support: hello@niaimpactinvest.com

Version history

1. v1.4Across the legal set — Version 1.4, effective 11 September 2026: the company you contract with changes. The contracting company is now Nia Impact, company number 221851, registered at Lot 02, Floor 1, CentrePoint, Trianon, Mauritius, still trading as Nia Impact Invest; the Terms, the Privacy Policy and the sub-processor list carry its details in place of the previous South African company's; that company's tax-registration number is removed from clause 1.1 and from the signature block, and the sentence in the fees clause that stated fees exclusive of that tax is removed, with nothing put in its place. The clauses that rest on South African law — governing law, jurisdiction and notice timing, the intellectual-property indemnity, the fee currency and interest rate, the electronic-acceptance and consumer-protection provisions, and the data-protection framing of the Privacy Policy — are unchanged in wording and now carry a visible note that they are under review by our legal counsel following the move; counsel's pass on that text is a separate step. Because the company you contract with is not the one named in version 1.3, the acceptance-registry identifiers change to terms-2026-09-11-draft, privacy-2026-09-11-draft and ai-notice-2026-09-11-draft, and everyone is asked to accept the re-issued documents once on their next sign-in. The product name is unchanged from 1.3; all seven documents are re-issued as 1.4 so the set shares one version.
2. v1.3Across the legal set — Version 1.3, effective 9 September 2026: the product is renamed. The application these documents govern is now called Nia Within, and every document is re-issued with the new name; the wording is otherwise unchanged, and the company you contract with at this version is still the South African company named in the documents. Because the Terms, the Privacy Policy and the AI-processing notice are re-issued under a new name, their acceptance-registry identifiers change to terms-2026-09-09-draft, privacy-2026-09-09-draft and ai-notice-2026-09-09-draft, and everyone is asked to accept the renamed documents once on their next sign-in. The Terms, the Privacy Policy, the Cookie Notice and the sub-processor list carry the new name; the AI-processing page and the German summaries carry no text change; all are re-issued as 1.3 so the set shares one version.
3. v1.2Across the legal set — Version 1.2, effective 7 September 2026: a factual correction to what version 1.1 said about embeddings. Embeddings — the numerical representations of your text used for search — are now sent only to an endpoint that keeps no copy of the input and does not train on it; today that is Microsoft Azure, the only such endpoint offered for the embedding model we use. Version 1.1 said that no zero-retention endpoint existed for embeddings and that they were sent to Microsoft Azure or OpenAI. Both statements are corrected: Privacy Policy section 6(a) and the section 7 table, the AI-processing page and the sub-processor list now say that embeddings go only to the zero-data-retention endpoint at Microsoft Azure, and that OpenAI remains a permitted host under our routing rules but receives no requests today. This tightens how your text is handled; nothing is taken away and nothing new is asked of you, so no one is asked to accept again. The Cookie Notice, the Terms and the German summaries carry no text change and are re-issued as 1.2 so the set shares one version. Privacy Policy and Terms of service — Version 1.2, effective 7 September 2026: two statements in section 10 of the Privacy Policy are replaced because the Platform now does what they said it could not yet do. Deleting your account no longer deletes the audit-trail entries recorded in your name: each entry keeps what was done and when, and your name is replaced by a label that does not identify you, such as "deleted admin #a1b2c3d4" — never your name and never your e-mail address. Administrator accounts can now be deleted the same way as any other account; a setting an administrator changed for an organisation keeps its value, with the setter's name removed. The sentences saying the trail was deleted with the account, and that some administrator accounts could not yet be deleted, are removed because they are no longer true. Nothing is taken away, nothing new is asked of you, and you are not asked to accept again. Clause B8 of the Terms of service (closing your account) is corrected to match: it no longer says that audit entries recorded in your name are deleted when you delete your account yourself, or that you should ask your administrator to remove you instead — whichever way your account is closed, the entries are kept under the same non-identifying label. The other five documents carry no text change and are re-issued as 1.2 so the set shares one version.
4. v1.1Across the legal set — Version 1.1, effective 6 September 2026: a wording and link tidy-up of version 1.0. Nothing is taken away and nothing new is asked of you; the only change to what the Terms require is one new commitment by Nia, in clause 21.1, which section 21.3 makes effective on publication, so no one is asked to accept again. Privacy Policy section 7: the GitHub row now says the database backup job is not running on a schedule at present, as sections 9 and 10 already said; the opening sentence of section 7 is reworded by one word. Privacy Policy sections 9 and 10: the database-backup sentences now say the off-site copies were taken from 1 to 3 September 2026 and the job is not running on a schedule; the promise itself is unchanged. Sub-processor list: Resend's legal entity is written as Plus Five Five, Inc. Terms 21.1: Nia still provides a copy of any earlier version on request, and now commits to 10 business days where the request is made in writing — the same turnaround as an export under 7.8. All documents: the legal hub and the versions page are linked from the site footer. The Cookie Notice, the AI-processing page and the German summaries carry no text change and are re-issued as 1.1 so the set shares one version.
5. v1.0Version 1.0, effective 4 September 2026: the first published version, adopted by Nia Impact Invest. It supersedes the draft of 2 September 2026. Later revisions will be listed here with a summary of what changed.

All versions