.
Article
A Responsible Guide to Using LLMs at Work
This guide, written by Ben Martin, Wordsmith’s Head of Privacy, serves as a starting point for leveraging Generative AI with integrity and security across your organization. Built upon the core elements of Wordsmith's own approach to use of AI for in-house legal and business use, this document sets out some of the key building blocks to build your own data ethics or responsible AI usage policy. It is a practical summary to help you balance innovation with compliance when determining your own approach and includes information and examples about how we are building responsible AI tools here at Wordsmith.
Generative AI is a fantastic tool that can streamline our workflows and spark creativity. But we need to ensure that we’re steering it and using it in the right way.
To use these tools effectively, we need to balance innovation with integrity. This guide empowers you to use AI safely, ensuring we protect our data, our colleagues, and our reputation.
1. Data Privacy & Security
If it’s not a private AI system, keep it out of the prompt. Public AI models often "learn" from what you type, meaning your trade secrets could end up in someone else’s answer. Never input PII (personal info), financial specs, or IP unless you are using a secure, company-approved instance.
Scenario: You need to summarize confidential strategy notes for a merger.
Best Practice: Don't paste the raw text into a public chatbot. Anonymize the data first (e.g., change "Company X" to "Target A") or use approved internal tools only.
Wordsmith does not train its models on any data you add into the system, and this is stated in Wordsmith’s Terms of Service (see clause 3.2.1). Your AI providers should state whether they train their models on your data in their Terms or Privacy Policy. |
2. Bias Awareness
AI mirrors the internet - including its stereotypes. Models can unintentionally output biased content regarding gender, race, or culture. It’s up to you to spot this and correct it.
Scenario: You use AI to help draft a job description.
Best Practice: Watch out for hidden stereotypes. AI can default to gendered language, for instance, describing men as "strategic" and women as "helpful" for the exact same achievements. Critically review the text to ensure it focuses on objective results, not unconscious tropes.
At Wordsmith, we prevent the common issue of AI systems learning from potentially inaccurate or biased customer inputs because we do not use input data to train our AI models or improve our systems. To uphold this commitment, we employ strict data segregation, provide auditable logs, and use a model governance framework. This framework automatically selects the highest-performing, compliant large language models from leading providers (such as Anthropic, OpenAI, Google DeepMind), ensuring we always use the most accurate tools available while maintaining the highest data protection standards. You should ensure that tools you use have measures in place to minimize the risk of bias. |
3. Verification & Fact-Checking
Most LLMs are designed to sound convincing, not necessarily to be truthful. They can "hallucinate" facts and figures confidently. You are responsible for the accuracy of your work.
Scenario: You ask AI for a competitor's revenue growth last year. It gives you a precise number.
Best Practice: Treat that number as a guess. Go to the actual financial report or a trusted news source to confirm it before putting it in your deck.
Wordsmith takes a multi-layered approach to preventing hallucinations and ensuring accuracy. We maintain comprehensive internal benchmarks covering all key system components - main chat, workspaces, document review, and redlining functionality. These benchmarks are created and maintained by our legal team using our own proprietary data, ensuring they reflect real-world legal scenarios without compromising customer confidentiality. When looking for tools - ask about how developers ensure that the output is valid and correct. |
4. Human oversight
AI is a drafter, not a decision-maker. Never let algorithms make final calls on things that impact people's lives, like hiring, promotions, or access to services.
Scenario: You use AI to scan through a pile of resumes to save time.
Best Practice: Use an approved LLM to format data, but never let the AI decide who gets "rejected." A human must review candidates to ensure qualified people aren't filtered out by a biased algorithm.
Wordsmith allows you to check your co-counsel’s work. Administrators can review and approve all Wordsmith responses within dedicated channels prior to the responses being shared, which ensures human validation for every interaction. You’re also able to highlight sensitive matters with alerts which notify specific users. Check your tools have features which allow this, or build in processes to ensure that you have human oversight of outputs. |
5. Transparency & Integrity
Be open about it. Using AI isn't "cheating" if it helps you be more productive, but passing off machine-generated text as your own original thought isn't right.
Scenario: You use an LLM to write the first draft of a departmental update.
Best Practice: Add a simple note when you send it: "Drafted with the assistance of AI, edited and reviewed by [Your Name]."
6. Scope of Competence
AI should not be used as the sole source of advice on high-stakes matters, check your answers yourself and use licensed professionals where appropriate.
Scenario: A client asks a tricky question about a contract clause.
Best Practice: Don't ask a non-specialist chatbot to interpret the law. Consult the Legal team. The AI can help you write the email, but the advice must come from (or be verified by) a human expert.
Whilst Wordsmith is a specialist tool for in-house legal teams. It does not provide legal advice, but instead helps to automates routine legal tasks, such as contract review and document drafting, which can then empower qualified lawyers to work more efficiently and gain better visibility into legal workflows. Wordsmith is designed to support and augment legal professionals' expertise, rather than replace it. |
7. Consent
Ask before you upload. Just because you have a colleague's data doesn't mean you have permission to feed it to a third-party AI.
Scenario: You want to write a fun bio for a coworker using their CV details.
Best Practice: Ask them first. Processing their personal data through an external tool is a new use of their info that requires their consent.
8. Respect Workplace Policy
Stick to the approved toolkit. Don't use "Shadow IT" (browser extensions or unvetted apps) that haven't been checked by Security.
Scenario: You find a cool new tool that makes better slides than our standard software.
Best Practice: Submit a request to IT or procurement before using it. We need to make sure it doesn't have security holes before we trust it with our work.
Summary: Treat AI like a junior intern. It’s helpful and fast, but you need to check its work, protect sensitive info, and take final responsibility for the output.
Quick Guide: AI Dos and Don'ts
✅ DO
Fact-check everything: AI hallucinates. Verify numbers and facts with a trusted source.
Be transparent: It’s okay to use AI as a drafter, just let people know (e.g., "Drafted with AI assistance").
Review for bias: Check if the output relies on stereotypes, especially regarding gender or culture.
Stay in the loop: You are the decision-maker. Never let AI make final calls on hiring or personnel issues.
Ask first: Get consent before uploading a colleague’s info or resume into a prompt.
❌ DON'T
Feed the bots secrets: Never input personal data, confidential client data, or trade secrets into public models.
Use it for un-checked legal advice: AI is not a lawyer (or a doctor). Don't use it for compliance or safety-critical answers, without verification.
Go rogue: Stick to IT-approved tools. Avoid unvetted browser extensions.
Copy-paste blindly: Always refine the tone and check the logic before hitting send.
