AI Assistants in the Company: From Prompt Engineering to Useful Skills

Now many companies have reached the stage of implementing AI assistants, but not everyone has an understanding of where exactly to start and how to build a useful corporate service based on the new fashionable technology.

The first thing you will have to figure out is prompt engineering , a basic skill that not only programmers but also every employee who will encounter AI in one way or another will have to learn to use.

The experience of creating clear instructions for artificial intelligence directly affects whether the assistant becomes a powerful tool or remains an expensive toy.

Principles of Effective Communication with AI

For the AI to understand you, your instructions (prompts) must follow three rules:

1. Formulate the task clearly and unambiguously. Instead of “tell me about sales,” use “As a sales department manager and analyst with 30 years of experience, prepare a report on sales of the Alpha product in the southern region for the 3rd quarter of 2024 in the form of a bulleted list.”

2. Add context. The base model doesn't know your internal terms, data, or subject area, and can't know what exactly you mean, so be sure to provide it with all the information it needs to make a decision.
For example, instead of "Make a commercial proposal for rolled metal products," you should write a detailed prompt:
"Our company produces shaped rolled metal products: A500C rebar, P and U channels according to GOST 8240-97. Production in cities A and B. Next year, we plan to grow by 15% and launch sales in regions K, L, and M.
Prepare a marketing plan for entering these markets, taking into account logistics, seasonal demand, and consumption patterns in region X.
Specify sales channels, partner models, and offers relevant to target customers.

<data>

• Rebar A500C is a rod made of steel with a strength class of 500 MPa, suitable for welding.

• Channel P/U is a metal profile, P - with parallel shelves, U - reinforced.

• GOST 8240-97 — standard for rolled shapes.
  This prompt will allow the AI to prepare a more accurate answer.
  <data>»

3. The prompt should be short and to the point. Avoid ambiguity. Every detail in the prompt should help, not confuse.

The first prompt is rarely perfect. Analyze the result and iteratively improve the instruction. It is also recommended to create a system for testing prompts and automatically checking their effectiveness.

Read more materials on this topic in Compass CIO

From simple instructions to complex tasks

Once you have mastered the basic technique, also known as Zero-Shot, a one-word query, you can move on to more advanced techniques:

Few-Shot (training by examples): This is the easiest way to teach a model your format. You simply show it a few examples right in the prompt.

Example:

Text: "The client is delighted with the new interface." -> Category: Positive review.

Text: "After the update, the system constantly freezes." -> Category: Technical problem.

Text: "Can't find the report export button." -> Category: ?

The model will most likely answer: Technical problem.

Chain-of-Thought: For complex problems, make the AI “think out loud” by breaking the solution down into steps. This dramatically reduces errors.

Example: "Our company produces A500C rebar and U channels according to GOST 8240-97. We plan to enter new markets in the cities of K, L and M, located in region X. Think step by step and propose a marketing strategy. First, define the characteristics of the target market, then the logistical constraints, after that the key customer needs, and finally the offers and promotion channels."

Such reflections will allow the model to carefully prepare the ground for the final answer.

CoT techniques also help to understand how exactly AI made a particular decision. This can be critical in finance, law, or information security.

The main thing is data. How to teach AI to work with it?

The real value of AI assistants can be realized by giving them access to corporate data.

Document Management (RAG): Retrieval-Augmented Generation (RAG) technology enables AI to answer questions based on the content of your internal documents. The system finds the relevant fragment in the knowledge base and passes it to the model along with the question. This ensures accuracy, relevance and security, since the AI does not “fantasize”, but uses only verified facts.

Working with databases: In order for AI to analyze data from a CRM or ERP, it needs to be explained its structure (schema). After that, it can transform queries like “show me the top 5 most profitable customers for the last month” into real SQL queries that will first get the actual data from your system, and then the AI will prepare analytics and a response based on that data.

Security: The more data an AI assistant has access to, the higher the potential risk of data leakage or misuse if it is not properly secured and managed.
Therefore, in addition to technical expertise in industrial engineering, enterprises should develop strict security frameworks and governance policies for their AI assistants. It is recommended to use a role-based data access model that will prevent the AI assistant from accessing data beyond the rights of the user with whom the AI assistant is communicating.

The Scaling Problem and the Role of Platforms

Managing one prompt is easy. But when you have dozens or even hundreds of assistants for different departments, and each user starts to accumulate their own, favorite and most effective prompts for their role, manual management turns into chaos.

Specialized platforms such as GPTZATOR from the Lad group of IT companies can help here, solving key problems:

• Centralized management: Create a single library for storing and versioning assistants' "skills".
• Easy integration: Provide ready-made tools for connecting to the company's infrastructure, office products such as P7 or 1C, and to databases and documents (RAG).
• Simplifying development: Thanks to no-code/low-code interfaces, AI can be configured not only by developers, but also by business analysts or even employees at their workplace.
• Security: Enterprise-grade platforms allow you to deploy all components on-premise. This not only allows you to run AI models locally, but also ensures that your sensitive data never leaves the organization.

Using a platform like GPTZATOR allows you to take a systematic approach to implementing AI. You get a visual designer for creating skills and deploying them safely, which turns experiments into a manageable and scalable business process. Convenient interfaces for managing prompts and integrated corporate interfaces open up opportunities to create truly effective assistants.

From a toy to a working tool: conclusions

The difference between a basic chatbot and an effective AI assistant is not so much the power of the model, but the quality of the instructions and data you provide it.

The path to creating a smart assistant:

• Start with one specific task (for example, classifying support tickets).
• Master basic industrial engineering.
• Connect data via RAG.
• Use the platform to create a robust, useful enterprise service to systematically and securely deploy AI across your business.

This approach is what allows AI to transform from an obscure, trendy technology into a real asset that saves time and frees up your employees to solve more creative problems.

Note: GOST — State Standards used in the USSR and some post-Soviet countries.