AI Terms
Last updated: 9 September 2026
These terms cover the AI features in Summarly. They supplement our Terms of Service and Privacy Policy: those apply to the service as a whole, while this document says specifically what the AI does, what data reaches a model, and what you should never rely on it for.
Where Summarly uses AI
AI is not a separate product bolted onto Summarly — it is how several core features work. These are all of them.
| Feature | What is sent to the model | What comes back |
|---|---|---|
| Transaction categorization | A batch of transactions — description, the raw bank text, amount, currency, type — together with your category names and the vendors already known in your workspace. | A proposed category, a tidied merchant name, and a vendor match. |
| Receipt & invoice reading | The image or PDF itself, however it arrived — uploaded, captured from your phone, or sent to your receipt inbox. | Total, VAT, date, vendor, document type, a one-line summary and tags — used to propose a match to a transaction. |
| Document reading | The contents of PDFs and images you store in Documents. Office files are not read. | A title, a one-line summary, and organizing tags. |
| Assistant | Your question, the page you asked it from, any files you attach to that question, and the results of the read-only queries it runs against your own workspace. | An answer, and sometimes a chart of figures it read. |
| Firm assistant | For accounting firms only: the question, plus read-only receipt-completeness and transaction figures for exactly the client companies that firm member is already entitled to see. | An answer across that member’s own book of clients. |
Nothing here reaches beyond your own workspace. No AI feature sends one customer’s data into another customer’s request, and none of them sends your data anywhere except the provider named below.
Who processes it
Today our only AI sub-processor is Anthropic (Claude). Requests are made to its API under our commercial agreement with them. The current list is always the one in the sub-processor table of our Privacy Policy, which also sets out how transfers outside the EEA are protected — we will update it there, with notice, before any new provider handles your data.
Features use different models for different jobs: a fast one for categorization and the first pass over a receipt, a more capable one for the assistant and for a closer second read when the first pass is unsure. We change models as better ones arrive, and we do not treat any particular model as part of this agreement — every commitment below applies whichever model, and whichever provider, is doing the work.
Whichever provider it is, your data reaches it under a commercial agreement rather than a consumer one: it is processed to answer the request in front of it, and it is not used to train models.
What we do not do with your data
- We do not train AI models on your data— not our own, and not our provider’s. Your transactions, receipts, documents and assistant conversations are used to answer your request and nothing else.
- There is no data-sharing or model-improvement programme to opt into, because we do not run one. If that ever changes it will be an explicit, opt-in choice with notice — never a default.
- We do not sell your data, and we do not use it for advertising or profiling.
- Each workspace’s data is isolated at the database level, and the assistant’s queries run under that same isolation — it can only read what your workspace can read.
The assistant is read-only
It cannot create, edit, categorize, send, or delete anything in your books. It answers from tools that read your workspace, and every figure it states has to come from one of those reads rather than from its own arithmetic. Questions about Cyprus tax are answered only from the reference material we give it, and it cites the source; questions about how Summarly works are answered only from Summarly’s own help content.
You are still the one who checks
AI output is a suggestion, not a decision. That is built into the product: a suggested category is written to a separate field and only becomes your category when you confirm it in Review, and a receipt read from a photo is proposed as a match rather than applied.
Categories, totals, VAT figures, matches, summaries and assistant answers can be wrong — and they can be wrong in ways that read as confident and plausible. Review them before you file, pay, or rely on them. Summarly is not an accountant, auditor, or tax adviser, and the assistant does not give tax, legal, or financial advice; nothing it says replaces a qualified professional.
Limits & availability
- Assistant messages are capped per calendar month — 15 on the free plan, 300 on Premium — and the count resets at the start of each month.
- Free-tier AI can be paused when the free population as a whole reaches our monthly cost ceiling. Premium access is unaffected.
- A question can carry at most six attachments, and very large files are shortened before the model sees them — the assistant will say so rather than pretend it read the whole thing.
- AI features are provided “as is”. We may change, limit, or withdraw them, and the disclaimers and liability limits in our Terms of Service apply to them in full.
Your controls
- You can delete any assistant thread from the assistant panel, which removes it from your history.
- Files you attach to a question are context for that one question — the file contents are never kept as part of your chat history.
- You can export everything, or permanently delete your workspace, from Settings → Privacy & data. Deleting the workspace deletes its assistant history with it.
- Categorization and receipt reading are how the product works, so they are not individually switchable. If you would rather a document were never read by a model, do not upload it.
Changes
We may update these terms from time to time. Material changes — in particular any change to the commitments above — will be communicated through the app or by email before they take effect.
Questions? Email founders@summarly.com.