A hands-on guide for organizations starting with AI - covering prompt engineering, building your first FAQ assistant and AI agent with AnythingLLM, plus a 14-day pilot plan to measure real ROI.
Most AI guides tell you what AI can do. This one tells you how to actually do it - step by step, from opening your first chat window to building a working AI assistant that answers questions from your company documents, with a 14-day plan to prove it is worth the effort.
1. Where should your organization start?
The biggest mistake is trying to "use AI everywhere" at once. Start with one pain point that is measurable and repetitive:
- HR policy questions - "how many annual leave days" asked 50 times a month
- Document search - finding the right contract clause in 40-page agreements
- Report summarization - condensing weekly department reports for management
- Email drafting - standard responses to common customer inquiries
Pick one. Measure how much time it takes today. That is your baseline.
2. What tools to use and where to start
For your first experiments, you need nothing more than a web browser:
- ChatGPT (chat.openai.com) - the most capable general assistant
- Claude (claude.ai) - excellent for document analysis and writing
- Both are free to start - upgrade to paid when you hit usage limits
Important: for company use, do not paste confidential information into personal accounts. Set up team/enterprise accounts, or use local tools (covered in section 6).
3. What is prompt engineering?
Prompt engineering is the skill of communicating with AI to get the output you need. It is not magic words - it is clear communication:
- Bad prompt: "write an email"
- Better prompt: "write a professional email to a Thai client apologizing for a delayed delivery, referencing order #12345, offering a 5% discount on the next order, in polite business Thai"
The more context you give, the better the output. AI is not a mind reader - it is a very fast assistant who has never worked at your company.
4. Practical prompt techniques
- Role assignment: "Act as a senior accountant reviewing this reconciliation..."
- Output format specification: "Return as a bullet-point list with headers" or "Return as JSON with fields: name, email, department"
- Examples (few-shot): "Here are 3 examples of good product descriptions. Write a similar one for: [product name]"
- Chain-of-thought: "Think step by step before answering. Show your reasoning."
- Refinement loop: If the first answer is not right, say "make it shorter", "more formal", "add a specific example" rather than starting over
5. What is an AI agent and how is it different from chat?
A chatbot answers one question with one response. An AI agent can take multiple steps: search your documents, call an API, draft an email, check a database and compose an answer from all of those sources - all autonomously.
Think of it as the difference between asking a colleague a question and delegating a task to them.
6. Hands-on: build a FAQ assistant and your first agent with AnythingLLM
AnythingLLM is a free, open-source tool that lets you build a document-based AI assistant without coding:
- Download and install AnythingLLM Desktop (free from anythingllm.com)
- Create a workspace - name it something like "HR Policies"
- Upload your documents - PDF, Word, text files of your employee handbook, policies, FAQ documents
- Start asking questions - the AI answers grounded in your documents with citations
- Test with real questions - "how many sick days per year?", "what is the dress code policy?"
Everything runs on your machine - no data goes to the cloud (depending on the model you choose). This is a safe starting point for confidential documents.
7. From experiment to organizational adoption
Once the pilot works, the path to adoption is:
- Document what works - which prompts, which documents, which use cases produced real time savings
- Train others - share the working prompts and workflows with the team
- Standardize tools - choose one platform (AnythingLLM, ChatGPT Team, Claude Team) for consistency
- Set data boundaries - create a simple policy: what data can go into AI, what cannot
- Measure continuously - track hours saved, quality improved, adoption rate
8. The 14-day pilot plan and how to know if it is worth it
| Days | Action | Success metric |
|---|---|---|
| 1-2 | Choose one use case and measure current time spent | Baseline: hours per week |
| 3-5 | Set up AnythingLLM, upload documents, test 20 real questions | 80%+ answers correct |
| 6-10 | Daily use by 2-3 team members; refine prompts and add documents | Daily active usage |
| 11-14 | Measure time spent on the same task with AI assistance | 30%+ time reduction |
If you see 30%+ time reduction with acceptable quality, the ROI case is clear. If not, the pilot cost you two weeks and zero software licenses - far cheaper than an enterprise AI project that fails after six months.
Ready to go beyond the pilot? Devsign’s AI training programs take teams from first prompt to production deployment.
