Skip to content
WorkNivo
Free beta

AI Use & Responsibility Policy

Version 0.3.0DraftLast updated 29 October 2026Not yet in effect

WorkNivo uses AI to read purchase orders and prepare draft orders for your team to check. This page explains, in plain terms, what the AI does, what it sends to our AI Provider, what it can get wrong, which safeguards exist and which do not, and what your team must do before relying on the result.

The rules in Permitted and prohibited uses are part of the Terms of Service. The rest of this page explains how the Services work and does not limit the rights you have under the Terms or the law.

Key things to know

  • WorkNivo can make mistakes. Always review AI-prepared information before approving business decisions.
  • An AI Draft is a suggestion, not a verified fact. The AI reads the document; WorkNivo's own code then matches it against your records, checks prices and calculates totals; a person in your team decides.
  • The checks only catch what does not fit your records. A misread product code that happens to match another product in your catalogue, or a line the AI skipped, can pass every check.
  • Each uploaded PDF is sent in full to our AI Provider, Anthropic, and is processed outside the EU. Your customer, product and price lists are not sent.
  • Approval records that a person in your team accepted the order. It does not mean the order is free of errors, and it is never done by the AI.
  • Do not use WorkNivo to make decisions about people or for medical or other regulated decisions. Take extra care with documents it is not designed for, such as invoices or contracts.

This summary is not a substitute for the full text below. If they differ, the full text applies.

WorkNivo uses AI in two places only: to read Uploaded Documents in WorkNivo Intake, and to answer questions in the website assistant on worknivo.app. Everything else, including customer and product matching, price checks, totals, approvals, exports, team activity and analytics, is done by WorkNivo's own code following fixed rules, or by people in your team.

In WorkNivo Intake the AI's job is narrow. It:

  • reads supported documents (PDF purchase orders, up to 20 MB and 30 pages);
  • extracts structured information from them: who the buyer is, the order details and each order line;
  • provides the information that WorkNivo's code then uses to suggest customer and product matches and to identify possible discrepancies;
  • helps prepare draft order records, so your team does less retyping.
Who does what when a purchase order is processed
StepDone byWhat happens
1. UploadA person in your teamYou upload a PDF. Before anything is sent to the AI, WorkNivo checks that it is a PDF, within the size and page limits, and unchanged since upload (by comparing a digital fingerprint). Uploading exactly the same file twice in a Workspace is blocked.
2. Reading (extraction)AI model (Anthropic)The AI reads the whole document and returns, in a fixed format: the buyer's name and customer reference, purchase order number, order date, requested delivery date, currency, delivery notes, and for each line the product code, description, quantity, unit, unit price and the line's text copied from the document. It also says whether the document looks like a purchase order and whether it is readable.
3. Checks and calculationsWorkNivo's code (no AI)WorkNivo checks the AI's response against the fixed format, matches the buyer and products against your own records, compares prices with your agreed customer prices or list prices, calculates line totals and the subtotal, and flags issues.
4. ReviewPeople in your teamReviewers see the original document next to the draft, correct fields, choose the right customer or product where needed, and resolve flagged issues.
5. Approval or rejectionA person with the required roleA person approves or rejects the order. WorkNivo blocks approval while required information is missing or an error-level issue is unresolved.
6. ExportA person in your teamApproved orders can be downloaded as CSV or XLSX files, generated in your browser. WorkNivo does not send them to any other system.

Our AI Provider is Anthropic. The full details of this data flow, including the legal basis and transfer safeguards, are in the Privacy Policy (AI processing) and on the Subprocessors page. In short:

What is sent
For each document: the entire PDF, together with fixed instructions written by WorkNivo. Everything in the PDF is sent, including any names, email addresses, phone numbers, delivery addresses, prices and commercial terms it contains.
What is not sent
Your customer, product and price lists, the names or email addresses of your users, and your Workspace name. Matching against your records happens inside WorkNivo after the AI has replied.
Where it is processed
Outside the EU. Anthropic processes requests globally or in the United States and stores data in the United States; it does not currently offer EU-only processing. Every document WorkNivo processes therefore leaves the EU. How transfers are protected.
How long Anthropic keeps it
Under Anthropic's published terms, API inputs and outputs are kept for up to 30 days, and longer if they are flagged under Anthropic's Usage Policy (inputs and outputs up to 2 years, trust-and-safety scores up to 7 years) or if the law requires it. WorkNivo does not have a zero-data-retention arrangement with Anthropic [To be confirmed: confirmation that no zero-data-retention arrangement is in place].
Training
WorkNivo does not train AI models and does not use Customer Content to train models. Anthropic's Commercial Terms say Anthropic may not train models on customer content from its services; this applies to WorkNivo's account only if that account is under those terms [To be confirmed: confirmation that WorkNivo's Anthropic account is under Anthropic's Commercial Terms].
Which model
Documents and website-assistant questions are handled by an Anthropic Claude model [To be confirmed: model used in production]. If the chosen model refuses a request or is overloaded, Anthropic may answer with a different model; WorkNivo records which model answered for each document [To be confirmed: which models Anthropic's fallback can use].
What WorkNivo keeps from the reply
The extracted fields (saved as the draft order, its lines and its issues), each line's text copied from the document (up to 200 characters), the name of the model that answered and the amount of processing used. WorkNivo does not keep the AI's raw reply and does not write document content to its logs.
Anthropic's role
Anthropic acts as WorkNivo's Subprocessor for this processing, and as a controller only for limited purposes under its own Usage Policy, such as safety monitoring [To be confirmed: Anthropic data processing addendum in force for WorkNivo's account].

AI models read documents in a way that is usually right but never guaranteed. WorkNivo does not guarantee:

  • perfect extraction, or the correct reading of every document;
  • a correct customer or product match every time;
  • accurate results from poor-quality, unusual or incomplete documents;
  • that every error or discrepancy will be found;
  • financial accuracy of quantities, prices or totals without your verification;
  • that exported files or records meet any accounting, tax, customs or other regulatory requirement;
  • that flagged issues, suggested matches or explanations are always correct.

The table shows the kinds of mistake that can happen on a purchase order, with an example, and whether WorkNivo's checks would flag it. The examples are illustrations, not a complete list.

Kinds of AI mistake on a purchase order, with examples, and whether WorkNivo's checks flag them
What can go wrongExampleWill WorkNivo's checks flag it?
Misreading a character or numberOn a faded scan, a quantity of 8 is read as 3.No. 3 is a valid quantity. Only a person comparing the draft with the original will see it.
Confusing similar product codesAB-1007 is read as AB-1001.Not reliably. If AB-1001 exists in your catalogue, it is matched without an issue. A price-mismatch error appears only if the document price differs from AB-1001's price.
Matching the wrong customerThe AI picks up the delivery-address company, or your own company name, as the buyer. Or the buyer is "ABC Trading GmbH" and your records hold a different company called "ABC Trading Ltd".Only partly. If no customer matches, or several match equally, an error is flagged. If the wrong name matches one of your customers, it is accepted. Names are compared ignoring capitals, accents, punctuation, a leading "The" and legal-form endings such as Ltd, GmbH or Inc, so two different companies with the same trading name are treated as the same customer.
Misinterpreting quantities and units"10 boxes of 12" is read as 120 units, or "1.000" (one thousand, written with a dot) is read as 1.Partly. A unit that differs from the catalogue unit raises a warning, and a fractional quantity for a product sold per piece raises an error. A plausible but wrong number is not flagged.
Missing informationThe requested delivery date is in a handwritten note or a footnote and is not picked up.Partly. A missing purchase order number or delivery date raises a warning; a missing quantity or price is flagged. Delivery notes and other text are not checked.
Producing an incomplete orderThe order table continues on page 2 and the last three lines are skipped.No. WorkNivo calculates the total from the lines that were found and does not read or compare the document's own total, so a missing line simply makes the total smaller.
Inaccurate descriptionsA long description is shortened, or two lines are merged into one.Only if the result no longer matches a product in your catalogue.
Misreading the layoutA freight or discount row is read as an order line; the seller's address is read as the buyer's.Partly. A row that matches no product is flagged as not found. A wrongly assigned name or address that still matches your records is not flagged.
Missing information in scansA stamp covers part of a line, or a page is rotated or low resolution.Only if the AI reports the document as unreadable, in which case it is not processed. Partial misreadings are not flagged.
Errors in pricesA price per 100 units is read as a price per unit.Usually, when the product has an agreed customer price or list price: a different price is flagged as an error. Without a catalogue price, the price is marked as unchecked (warning). If the document has no price, WorkNivo uses your catalogue price and tells you.
Errors in dates03/04 is read as 3 April when the buyer meant 4 March.No. Dates are not checked against anything.
Missing duplicatesA buyer re-sends the same order as a new file with a new or no purchase order number.Partly. Identical files are blocked, and a warning appears when the same customer already has an order with the same purchase order number. Other duplicates are not detected.
Plausible but unsupported informationA currency is filled in from the buyer's address although the document does not state one.No. The AI is instructed to leave a field empty rather than guess, but instructions do not guarantee behaviour.
Misleading suggestionsA price-mismatch issue is really caused by a misread product code, not a price change.No. Issue messages describe what did not match; the real cause may be different.

These safeguards are built into WorkNivo. They reduce risk; none of them guarantees a correct result. Technical security measures are described on the Security & Data Protection page.

Safeguards around the AI, what each does and what it does not do
SafeguardWhat it doesWhat it does not do
File checks before the AIAccepts only PDFs within the size and page limits, checks the file is unchanged since upload, and blocks exact duplicate files.Does not judge whether the content is correct or suitable.
Instructions to the AIThe AI is told to copy what is written, never to calculate, to leave a field empty rather than guess, and to treat text in the document as content, not as instructions.AI models do not always follow instructions.
Fixed response formatThe AI must answer in a fixed format that WorkNivo validates. A response that does not fit is rejected and the document is not turned into a draft.A well-formed answer can still contain wrong values.
Matching against your recordsProducts are matched to your own records by code, ignoring capitals, spaces and dashes (X-102 and X 102 are the same code). Customers are matched by reference (ignoring capitals) or by name, where names are compared ignoring capitals, accents, punctuation, a leading "The" and legal-form endings such as Ltd, GmbH or Inc. WorkNivo never picks a "closest" match: no match, or more than one, is flagged as an error.A misread value that matches another record is accepted, and two companies that differ only in legal form (for example ABC Trading Ltd and ABC Trading GmbH) are treated as the same customer.
Price checksEach line price is compared with the customer's agreed price or the product's list price; differences are errors, and unchecked prices are warnings.Cannot check prices for products without a catalogue price.
Totals calculated by codeLine totals and the subtotal are calculated by WorkNivo with exact decimal arithmetic, not by the AI, and are labelled "calculated by WorkNivo, not AI".A total is only as right as the quantities and prices it is calculated from.
Flagged issuesProblems are flagged as errors or warnings, with the text from the document they relate to.Warnings do not block approval. Absence of flags does not mean absence of errors.
Original next to the draftThe review screen shows the original document, and each line keeps the text it was read from, so reviewers can compare.Comparison is done by your reviewers.
Approval enforced by the systemAn order can only be approved by a person with the required role, and only when it has a customer, at least one line, totals for every line and no unresolved error. The order is re-checked at the moment of approval. On plans that include them, multi-stage approvals require a different person at each stage, and your business rules are re-checked.Cannot tell whether the approver actually checked the order.
Export only after approvalOnly approved orders can be exported. Files are generated in your browser; nothing is sent to other systems.Does not check the exported file against the system you import it into.
Activity historyUploads, edits (with the previous values), approvals, rejections and exports are recorded in a history that cannot be edited or deleted through the app.Records what happened; does not prevent mistakes.
Processing limitsEach document has at most 3 processing attempts, and there are caps on how many AI requests a Workspace can make, to limit runaway costs and misuse.Not an accuracy measure.

Notices in the product tell users about the AI at the point of use: the website assistant is labelled as AI, and the upload, review, approval and export screens explain that drafts are AI-prepared and must be checked.

WorkNivo is designed on the basis that a person reviews every order before it is used. The AI and the checks are there to save time, not to replace that review.

Extraction (AI)
The AI's reading of the document. It can be wrong in any field.
Automated checks (WorkNivo's code)
Fixed rules that compare the extraction with your records and calculate totals. They find some mistakes, never all of them, and they cannot see what is on the original document.
Approval (people)
A decision by a person in your team that the order, as reviewed and corrected, can be used in your workflow. This is the only stage that compares the draft with what the document actually says.

As the Customer, you are responsible for:

  • deciding who in your team may review and approve orders, and giving approval rights only to people who can check orders properly;
  • making sure reviewers understand what the AI can get wrong (this page is written to help) and have the time to check each order;
  • comparing each AI Draft with the original document before approving it, especially the business-critical details:
    • customer and delivery address;
    • product codes and descriptions;
    • quantities and units;
    • prices, discounts, currency and totals;
    • dates and delivery instructions;
    • whether the order is a duplicate or a revision of an earlier one;
  • reading warnings, not only errors, before approving;
  • keeping your customer, product and price lists accurate and up to date;
  • checking exported files before importing them into other systems or sending anything to your customers or suppliers;
  • getting financial, legal, accounting, tax or other professional advice where you need it. WorkNivo is not a substitute for it.

An Approval means that an authorized person in your team accepted the AI Draft, as reviewed and corrected, for your own workflow. WorkNivo records who approved it and when.

Approval does not mean that the order is free of errors, that WorkNivo has verified it, or that WorkNivo has accepted or confirmed the order on your behalf. "No issues flagged" does not mean "no errors".

Your review duty does not take away WorkNivo's own obligations. WorkNivo remains responsible for providing the Services as described in the Terms of Service, for describing the AI's role and limits honestly, for the safeguards described on this page working as described, and for protecting your data as set out in the Data Processing Agreement. Nothing on this page limits liability that cannot be limited by law, or excuses WorkNivo's own breach, negligence or unlawful conduct. How liability is shared is set out in the Terms (limitation of liability).

WorkNivo staff do not review or approve your orders for you.

This section is part of the Terms of Service. Breaking it is a breach of the Terms.

WorkNivo Intake is designed to read purchase orders that a business receives from its business buyers, and to prepare draft order records that the business's own staff review, correct and approve before using them in its order-handling workflow. You may use it for that purpose and for related back-office work within your plan's features.

You must not use WorkNivo, its AI features or its Outputs:

  1. to act on AI Drafts without human review, for example by approving orders in bulk without checking them, or by building a process that treats drafts or exports as verified;
  2. to make or support decisions about individual people that have legal or similarly significant effects, such as creditworthiness or credit scoring, insurance, access to essential services, education, law enforcement, migration or justice;
  3. to have the AI, or any automated process based on its Outputs alone, evaluate, score, rank or monitor employees or job applicants, or make decisions about hiring, performance, discipline or allocating work based on their personal behaviour or traits. This does not cover WorkNivo's team activity and report features, which count recorded actions such as orders approved, do not use AI, and are information for people to interpret; decisions about an employee must not be based on them alone [To be confirmed: confirm this boundary against AI Act Annex III(4) and employment law];
  4. for medical diagnosis or treatment, prescriptions, patient records or other clinical decisions;
  5. to process special categories of personal data (such as health data), or data about criminal convictions, in documents uploaded for processing [To be confirmed: final scope of the sensitive-data rule and any regulated workflows that are allowed];
  6. to upload documents you do not have the right to process, or that contain personal or confidential information you are not allowed to share with WorkNivo and its AI Provider;
  7. to manipulate the AI, for example by hiding instructions in documents, trying to extract WorkNivo's instructions to the AI, getting around its safeguards or limits, or using it to produce content unrelated to your orders;
  8. to access or infer another Workspace's data, or to test or attack the AI features without WorkNivo's written permission;
  9. in breach of the law, of the acceptable use rules in the Terms, or of the AI Provider's usage rules that WorkNivo makes available to you [To be confirmed: whether and how Anthropic's Usage Policy is passed through to Customers].

If you use WorkNivo for a purpose listed above, you, not WorkNivo, are responsible for the legal obligations that come with that use, including under the EU AI Act (see EU AI Act). WorkNivo may restrict or suspend access as described in the Terms.

The following are not prohibited, but WorkNivo is not designed or tested for them, and results are likely to be less reliable. Take extra care if you try them:

  • documents other than purchase orders, such as invoices, contracts, delivery notes or price lists;
  • handwritten orders, photographs of documents and poor-quality scans;
  • orders where a mistake in product identification could cause safety, legal or regulatory harm, such as dangerous goods, controlled or age-restricted products, or medical devices, without an additional check by a qualified person;
  • using exported files as accounting, tax or customs records without checking them.

The chat assistant on worknivo.app answers questions about WorkNivo using Anthropic's AI. It is labelled as an AI assistant and tells you its answers may be imperfect.

  • What is sent: your last messages in the conversation (up to 10 messages of up to 1,000 characters each), fixed instructions written by WorkNivo and WorkNivo's plan prices.
  • What WorkNivo keeps: nothing. The conversation exists only in your browser tab and is not saved or logged by WorkNivo. Anthropic keeps it as described above.
  • Please do not share confidential or personal information in the chat.
  • Answers are not commitments. If an answer differs from WorkNivo's legal documents or your agreement, the documents and the agreement apply. Ask support@worknivo.app if you are unsure.

What WorkNivo records to limit misuse of the assistant is explained in the Privacy Policy. Whether chat answers also need a machine-readable AI label is under review [To be confirmed: approach to machine-readable marking of website-assistant answers].

The EU Artificial Intelligence Act is Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. This section sets out how WorkNivo understands its position. It is WorkNivo's own assessment, not legal advice to you.

Roles under the EU AI Act
PartyRoleWhat it means here
WorkNivoProvider of an AI system (and downstream provider of a system built on a general-purpose AI model)WorkNivo builds WorkNivo Intake and the website assistant on top of Anthropic's model and offers them under its own name. It is responsible for the system's design, its transparency and this information.
The CustomerDeployerYour business uses WorkNivo Intake under its own authority in its professional activity. You decide who uses it, for what, and how outputs are reviewed.
AnthropicProvider of the general-purpose AI modelAnthropic develops the underlying model. The duties for general-purpose AI models (such as technical documentation and information for downstream providers) are Anthropic's. WorkNivo does not train or modify the model.

WorkNivo has assessed WorkNivo Intake against its intended purpose: reading business purchase orders and matching them against a business's own records for human review. That purpose is not one of the high-risk uses listed in Annex III of the AI Act, and WorkNivo is not a safety component of a regulated product. WorkNivo's assessment is therefore that WorkNivo Intake is not a high-risk AI system, and that it does not involve any practice prohibited by Article 5 [To be confirmed: confirmation of the written non-high-risk assessment].

This assessment depends on the intended purpose. Using WorkNivo for one of the prohibited uses, such as decisions about employees or creditworthiness, could fall into a high-risk category and could make the business that does so a provider with its own obligations. WorkNivo does not use AI to evaluate, score or monitor people; its team activity and analytics features count recorded actions and do not use AI.

  • The website assistant tells people they are talking to an AI before they start (Article 50(1)).
  • Drafts prepared by the AI are labelled as AI-prepared in the product, and values calculated by WorkNivo are shown separately from values read by the AI.
  • Machine-readable marking of AI-generated content (Article 50(2)): WorkNivo's approach for the website assistant follows the European Commission's guidance once settled [To be confirmed: Article 50(2) approach for Intake drafts and website-assistant answers].
  • This page is WorkNivo's description of the AI's purpose, capabilities and limits, and is kept up to date when they change.

The AI Act asks providers and deployers to take measures to support the AI literacy of the people who operate AI systems on their behalf. WorkNivo does this for its own staff. As a deployer, your business does this for the people in your team who use WorkNivo; this page and the notices in the product are designed to help you, but they do not replace your own measures.

If the AI produces something wrong, unexpected or harmful, please tell us at support@worknivo.app. Reports help us improve the instructions and checks. It is especially useful to hear about:

  • the same kind of mistake happening repeatedly;
  • a draft containing information that is not in the document;
  • a check that should have flagged a problem but did not;
  • a website-assistant answer that is wrong or misleading.

Tell us the Workspace, the order or file name and what was wrong. Please do not email the document itself unless we ask for it.

Security issues: see Report a vulnerability. Illegal content: see Report illegal content. All contact points are listed in the Legal Notice.

Support and general questions
support@worknivo.app
Security reports
[To be confirmed: dedicated security contact address] General contact meanwhile: support@worknivo.app

We update this policy when the AI features, the AI Provider, the data sent to it or the law change. The version history at the end of the page lists every change.

Changing the AI Provider, sending new kinds of data to it, or allowing Customer Content to be used for training would be a material change. We would give notice before it takes effect, as described in the Terms [To be confirmed: notice period for change notifications (in-app notice exists; email channel once available)].

Questions about this document

Email support@worknivo.app. Every contact point is listed in the Legal Notice.

Related documents

Version history

  1. 0.3.0 · 29 October 2026Legal review: key point aligned with the body (unsupported documents are not prohibited); customer-name matching described as it is built; the employee-evaluation prohibition narrowed to AI-based or automated evaluation; in-product AI notices described as built.
  2. 0.2.0 · 28 October 2026Full draft for legal review: AI role, data sent to the AI Provider, limitations with examples, verified safeguards, human review, permitted and prohibited uses, website assistant, EU AI Act position.
  3. 0.1.0 · 28 October 2026Structure created; text not yet drafted.

All documents are listed in the Legal Center.