Robert Ambrogi Blog
A few weeks ago, I wrote here about the imminent launch by Josef of the beta version of Josef Q, its new product that uses advanced AI to transform policies and regulations — such as those pertaining to privacy, data security, HR and procurement — into digital Q&A tools, so that employees or clients can easily get answers to policy and compliance questions.
Today, that launch arrives, as Josef makes the beta version available for anyone to try for free. The product will remain free throughout its as-yet undetermined beta period, after which Josef plans to begin selling the product.
An Australia-based company whose primary product is a no-code automation platform for legal professionals, Josef has partnered with OpenAI to develop the tool, which is currently using GPT-4, the latest version of Open AI’s large language model (LLM). However, Josef is “remaining agnostic” about which LLM it uses and may experiment with other companies’ models, Sam Flynn, cofounder and chief operating officer, told me.
The company sees Josef Q as particularly useful to businesses, where legal, compliance, information security, and HR teams can use it to provide employees with instant answers about corporate policies and procedures.
It can also be used by law firms, both to answer internal employee questions and as a client-facing tool to provide answers to policy and regulatory questions based on the law firm’s knowledge and expertise.
Flynn said the tool could even be used by legal aid organizations to provide consumers with answers to questions in areas such as landlord-tenant law.
How It Works
In a brief demonstration yesterday, Flynn showed me that creating a digital Q&A tool using Josef Q takes just minutes and requires only three steps.
First, the user uploads the policy document, such as a privacy policy or parental leave policy, to be analyzed by the LLM. In Flynn’s demonstration, he used a parental leave policy.
Second, the user trains the tool by asking three questions. The questions should reflect those that others are likely to ask of the tool. For the parental leave example, one question might be, “Am I eligible for parental leave?” Josef Q will suggest an answer, which the user can edit if need be.
Once the three questions have been asked and answered, the tool is trained and ready to launch.
Josef Q uses “human-in-the-loop” training and moderation. Even after the initial three-question training, the moderator can go in and see all the questions that have been submitted and the answers Josef Q has given, and refine the model by correcting or editing any of the answers.
Also, the person who submits the question can help further train the tool by giving a thumbs up or down to indicate whether the answer was helpful.
There is no danger of Josef Q “hallucinating” and making up answers, Flynn told me, because the answers it provides will always be limited to the policy document. Thus, if someone were to ask the parental leave Q&A the name of the current U.S. president, it would respond, “I’m sorry, I don’t have an answer.”
A future version of the tool will also include confidence scores for each answer.
During the beta period, Josef is capping the size of a document that can be uploaded to 8,000 characters. Once the product is ready for general release, users will be able to upload any size document or even multiple documents related to a policy.
See my prior post for more information about Josef Q.