Responsible AI

1.Availability

1.1Dynamo AI is switched on for a workspace on request, once the AI service that serves the workspace has been set up. Until then, AI features are not available in the workspace, even if they are part of the Plan.

1.2We do not publish accuracy or speed figures for Dynamo AI. Quality and response time depend on the model and the hardware it runs on.

1.3The rest of this page describes the design as built. It applies to a workspace once Dynamo AI is switched on for it.

2.What Dynamo AI does

• answers questions from your company’s records, files and attachments that you are allowed to read, and shows the sources;
• summarizes a document, answers questions about a document, and compares two versions;
• finds a certificate, a report or a passage in company knowledge;
• drafts text, and proposes actions such as creating a task, as drafts that you confirm.

2.1The ERP does not depend on AI. With AI not included in the plan, switched off, out of allowance or unreachable, AI features give a clear message and records, workflows, search and files work as usual.

3.Your permissions apply

3.1What a user cannot open, Dynamo AI cannot use for that user. This is enforced below the interface. Records are used only if the user has read permission on them, and files only if the user can read them in Drive or on the record they are attached to.

3.2Permissions are never copied into the index. They are checked on every request, so removing a role takes effect at once. Sources the user may not read are dropped before anything is sent to the model, and no title, count, snippet or citation of them appears in the answer.

3.3Tools run as the signed-in user, with the user’s own permissions, and cannot switch identity. Password, code, attachment and HTML fields are not indexed.

4.Tenant isolation

4.1Each customer’s workspace has its own database. The index, embeddings, conversations, drafts and usage counters of Dynamo AI live in that database and nowhere else. The shared inference service keeps no data between requests. One workspace can never retrieve another workspace’s data, and this is covered by tests with two workspaces.

5.Sources and transparency

5.1Answers cite the sources they are based on so that people can check them. Citations are kept only for sources that were actually given to the model, and an invented citation is removed.

5.2An answer may still be incomplete or wrong. Check important facts against the source. The user is responsible for what they do with an answer.

6.Actions are drafts that you confirm

6.1Dynamo AI cannot change your data by itself. A proposed action is a draft that runs only when its owner clicks Confirm. It then runs through the normal document permissions, validations and workflow. Permissions are checked again at confirmation, a draft expires after a day, no other user can confirm it, and a confirmed draft cannot run twice.

6.2There is no database query language and no shell command available to the model. The tools it can use are a closed list, chosen from the user’s request and not by the content of documents or by the model. Dynamo AI never bypasses accounting controls, approvals, segregation of duties or HR permissions.

7.Defences against prompt injection

7.1Text from PDFs, Office files, CRM notes, e-mails, imported data and Drive files is treated as untrusted data. It is fenced in the prompt, and characters that could close the fence or fake a source are neutralised. The system policy tells the model to answer from the sources and never to follow instructions found inside them.

7.2The model has no way to act, because tools are chosen before retrieval and actions need confirmation. A heuristic flags attempts, the answer says that a source contained instructions that Dynamo ignored, and the message is marked for audit. This was tested against a document that tried to reveal other data, run SQL and shell commands, create a task and forge a citation, and none of that succeeded. No defence is perfect, and we keep testing.

8.Models and where they run

8.1Dynamo AI runs on open-weight models served centrally, never installed on employee computers. They run on infrastructure that Dynamo operates. If a third party ever hosts the inference service, it is listed on the Subprocessors page before it is switched on.

8.2We check the licence of every model before production use. We do not use a paid per-query service from a model vendor to answer your questions.

8.3Model roles are configurable, so a model can be replaced without changes to your data or your permissions.

9.Your data and training

9.1Customer data does not train models. We do not use your records, files, questions or answers to train or fine-tune any model, ours or anyone else’s.

9.2Conversations are stored in your workspace, are visible only to their owner, and are purged after the retention period that your administrator sets in the AI settings. Usage is metered for allowance and billing: the number of requests and tokens, indexed pages and similar counters, not the content.

9.3The administrator controls which roles and apps may use Dynamo AI, which sources are indexed, whether actions are enabled, and the conversation retention.

10.Limits and responsible use

10.1Dynamo AI is an assistant, not an adviser. Its answers are not accounting, tax, legal, medical or employment advice. Do not rely on it as the only basis for a decision that has legal or significant effects on a person. Always apply human judgement.

10.2The Acceptable Use Policy applies to Dynamo AI. Do not use it to produce unlawful content, to get around permissions, or to try to reach data you may not see.

10.3Features have a monthly allowance counted in requests. When it is used up, users are told and the administrator is pointed to the subscription page.

11.Feedback and questions

11.1To report wrong, harmful or unexpected output, or a security issue, write to security@dynamoos.com for security matters or sales@dynamoos.com for anything else. For privacy questions about AI, write to privacy@dynamoos.com.

11.2We update this page when Dynamo AI changes, and will not describe a capability here before it is available.