| Minimum Software Version | 8.16.0 |
| Solution(s) | Cases International ✓ Cases US ✓ Institutions ⨉ Counsel X |
AI Workbench is a forward-looking enhancement to document analysis workflows, supporting faster setup and deeper document insight. It combines the intelligence of the Single Document Analysis tool with an interactive workspace, enabling clients to quickly structure, validate, and iterate on the key information within legal content.
If you have a question about the AI Workbench that doesn't appear here, please contact Opus 2 Platform Support.
FAQs on Functionality
1. Can I switch on the AI Workbench feature myself?
No, the AI Workbench the feature must be enabled by Opus 2 Platform Support.
2. What usage reporting is available?
Select Project menu > AI Usage. This shows total account usage by month, year to date, and for the last twelve months. It also shows how many documents and words have been processed.
System level reporting is also available for System Administrators at System Admin > AI Usage.
Note: We do not currently capture usage data per user.
3. How is the 600 million word maximum limit for fair usage of AI applied?
Once you do any analysis on a document its word count is added to your usage count. Single document queries add to the word count (full document text, question, and if enabled, also the history).
For multiple-document queries, only the fifty fragments that are being considered for answering the query are counted . Go to How does querying work in AI Workbench lower down this document for more details.
If the usage limit is reached, no further results will be generated.
At approximately 550 words per A4 page, 600 million words equates to 1.09 million pages of text.
4. What is the maximum number of documents I can add to a canvas section?
Each canvas section is limited to holding 50 documents, and this includes the top-level section of a canvas.
A canvas section is what we use in AI Workbench for analysis, so the maximum number of documents you can drop in to be analysed or query on at one time is 50 documents.
There is no limit to the number of sections in a canvas, so technically a canvas can have as many documents as you would like, however this is for structuring documents only. Having more than one canvas section does not change the number of documents you can work with in AI Workbench.
Taking the image below as an example:
- Even though the parent canvas is called My Canvas, there are three canvas sections here: My Canvas, Section A, and Section B.
- Each section can contain 50 documents, so the whole canvas itself has 150 documents.
- In AI Workbench, the user can select only one of the sections mentioned above to use for querying.
- Selecting the top level section My Canvas for querying would NOT select the documents in all 3 sections, only the 50 documents in that selected My Canvas section are used.
5. Can AI output include data not present in the document?
It shouldn't, but guardrails vary depending on whether the task is an information extraction or a summarisation.
6. Can AI output include data from the internet?
None of the AI features in the Opus 2 Platform involve the Large Language Model (LLM) making additional web requests for content from the internet. However, LLMs are trained on data from the internet. Our prompts include specific instructions to base responses ONLY on information found in the documents provided and not to use any prior knowledge from the training data.
7. Why do my AI document analysis results not have page links?
Unless the None reference option is selected, then all text based and query responses should have page link references.
8. Is there a limit on the number of queries I can run?
No, each query contributes to the client's fair use policy but there is no set limit on the number of queries. Query results can be saved in a project using the Add to project as docx option.
9. How does multiple document querying work in AI Workbench?
When several documents are dropped into a canvas, they are chopped into segments of approximately two hundred words each. When you enter a query, the AI Workbench identifies the fifty most closely-related fragments and generates a response to your query from the content of those fifty fragments. Hence the need for specific commands or prompts; vague questions won't give you useful results.
10. Can users re-run document analysis?
Not currently, but we expect to include this in a future release.
11. Can the summaries be ingested into Relativity via the Opus-Relativity integration?
No, the Relativity integration works in one direction only - from Relativity into the Opus 2 platform.
12. Can summaries be exported in bulk?
No, not at the moment. You can add them as documents to the project and then export from there, but you'd have to add the summaries to the project one by one.
13. Can I request a document summary of say, 150 words, for a criminal investigation?
Yes, you could command the query feature to provide a 150-word summary of this document.
14. Can I query for potential inconsistencies across multiple documents? For example, if I have a canvas that includes several deposition transcripts from the same expert across many different cases, can I find out if that expert has been inconsistent?
Unfortunately AI Workbench cannot do that. Our multi-document query functionality uses a two stage RAG process in which we first query the document set for the fragments of documents that are most closely related to the question and then the LLM is asked to generate a response based on those fragments.
As such, it works best with detailed questions, which, for example could mention particular people, places or events. It is not able to answer more abstract questions that would need a multi-stage process. This is under consideration for a future version. What users can do, is run a query such as Are there any inconsistencies against the statement [content] or information which argues it's not true? - this input can then be used as a starting point for a comparison.
FAQs for data security and responsible use of AI
15. Does the AI Workbench and the underlying service retain client prompts or data?
No content is stored by the LLM (Large Language Model) provider (AWS Bedrock).
The results of the standard analyses are stored in the database and retained in the client's own environment in the same way as all client work data is currently stored.
The questions posed by the user in the interactive query feature, and the responses generated by the LLM are not stored unless the user explicitly chooses to save them as a docx file within the system.
16. Are Opus 2's AI features trained in any way?
Opus does no training whatsoever. The Large Language Model is pretrained when we receive it.
17. What is an AI guardrail, and what guardrails does AI Workbench use?
AI guardrails are protocols and tools designed to ensure AI systems operate within ethical, legal, and technical boundaries, promoting safety and fairness. They act as barriers, preventing misuse, monitoring AI innovations, and safeguarding data privacy and public safety. The purpose of AI guardrails can include:
- Ethical and Legal Compliance
- Risk Mitigation
- Trust and Transparency
- Bias Detection and Reduction
In AI Workbench, querying has the strictest guardrails to ensure that it operates as per the four bullet points above for client-entered prompts. All other prompts are engineered to include required guardrails as part of their pre-defined configuration.
18. Are any guardrails visible to users?
No guardrails are visible to users. They are configured as part of our AI service backend.
19. What referencing mechanism is used for the Query functionality?
Opus 2 uses a three-step referencing approach:
A. Scoped Contextualization
If a query is run against a single document, the whole document is used to generate a response.
When a user submits a query against multiple documents, our system identifies the most relevant documents or excerpts from the client's uploaded data such as transcripts, pleadings, or witness statements. These selected excerpts are then injected into a structured prompt, which is passed to Claude, our commercial LLM (Large Language Model). This approach ensures that Claude only operates on the specific data relevant to the matter, and has no access to external content or general training data at runtime.
B. Implicit and Explicit Referencing
Because Claude is limited to the scoped content injected into the prompt, all answers are implicitly grounded in the client's case data. However, explicit citations, for example, “this came from page X of document Y”, are only present if that metadata is included in the original text and preserved during prompt construction. Otherwise, references are implicit but limited strictly to the data provided by the client.
C. Post-Response Semantic Matching
Once Claude generates a response, each paragraph of the output is semantically compared to the original set of excerpts using a method for measuring similarity between text excerpts (cosine similarity). This post-response semantic matching ranks the most relevant excerpts to each paragraph. The system then displays these high-confidence matches beneath each paragraph, grouped by source document. This allows users to quickly verify and trace the origin of the model’s statements providing transparency and confidence without relying on traditional RAG (Retrieval-Augmented Generation) infrastructure. Note: Post-Response semantic matching is used to identify related excerpts after a response is generated. Source type references, by contrast, ask the LLM to prove its output during generation. Only then are those references displayed in-line.
20. Is Opus 2’s AI technology leveraging some sort of RAG approach where you’ve also pre-trained your AI on a set of legal documents, or are you only using the commercial LLM?
Opus 2 leverages Anthropic’s Claude Haiku 3.0 as our commercial LLM (Large Language Model), and our current implementation does not involve fine-tuning the model on proprietary legal data. However, we have developed a set of domain-specific queries and structured prompt templates optimised for case preparation and strategy workflows.
These prompts are designed to guide Claude’s responses within the context of legal reasoning, helping ensure accuracy, relevance, and alignment with litigation best practices. While we do use a full Retrieval-Augmented Generation (RAG) architecture for answering multi-document queries, we don't use RAG for any pre-training.
Related Articles