Patterns of the Most Successful AI Projects

The patterns of the most typical and successful AI projects for business implementation are already known. They are:

• Data Extraction and
• Search Assistants

Data Extraction: This process involves converting unstructured or semi-structured information (such as PDF documents, scans, and photos of certificates, as well as client emails and call transcripts) into precise, structured data, ready to be loaded into our internal systems (ERP, CRM, HR databases, etc.). In essence, it automates the most tedious and expensive office work – manual data entry and verification – and serves as a direct replacement for the costly, slow, and error-prone manual labor that hundreds of our employees are still engaged in.

Search Assistants / Intelligent Search: Imagine we could provide every employee with a personal analyst who has read all of the company’s documents, including the internal Wiki, knowledge base, technical documentation, reports, regulations, and project correspondence from the past decade. Search Assistants are the evolution of internal corporate search from a primitive “keyword search” to a system that understands the user’s query and generates a direct, meaningful answer based on the learned materials, not just a list of links. It transforms our disparate knowledge bases (internal Wiki, SharePoint, technical documentation, regulations, project archives, etc.) into a single, interactive company-wide knowledge base. Technically, this is implemented through the Retrieval-Augmented Generation (RAG) approach.

How exactly?

A modern LLM is taken, the use of Structured Output techniques is mandatory, and a Custom Chain-of-Thought in the form of a Checklist is written into the scheme.

Structured Output: We don’t just ask the AI to answer; we force it to fill out a clear template (a table, a form). This ensures that the data will always be in the correct format for our CRM and ERP systems, without any issues.

Chain-of-Thought / Checklist: We teach the AI to think step-by-step, like our best employee. It not only provides a result but also follows an internal checklist, which dramatically increases accuracy and reliability, especially in complex tasks such as the legal review of a contract.

We will cover the techniques for improving AI response accuracy – Structured Output and Custom Chain-of-Thought – in more detail in upcoming publications.

And WOW – it’s done! This pragmatic approach enables the implementation of even large projects that appear impressive and have a significant impact on business metrics, demonstrating their high efficiency. This can also be a strategy for achieving quick wins.

Patterns of Less Attractive AI Projects (but which can still be brought to real numbers)

Currently, complex AI systems, such as copilots or multi-component AI agents that utilize RAG, are less attractive to businesses due to high resource requirements, lengthy development times, increased risks, and lower financial returns. Companies are adopting a pragmatic approach: mastering a broad range of tasks using reliable and secure methods. Development is constrained not by a lack of ideas but by a shortage of specialists capable of comfortably working with AI systems, which creates a bottleneck in implementation.

Examples of successful (and relatively easy) AI projects to launch:

• Automating the search for errors in incoming purchase orders;
• Matching component nomenclatures between suppliers (to monitor the market and sell faster);
• Automatically reconciling data with the corresponding order in the ERP;
• Checking the terms of a supplier’s contract against our company’s “gold standard”;
• Preliminary assessment of a candidate’s compliance with job requirements;
• Automatic updating of cargo status in a unified system based on the processing of all incoming logistics documents, from which it extracts container numbers, shipping dates, and descriptions of goods;
• An internal chatbot in the corporate messenger that instantly answers typical HR and accounting questions, referencing current policies and documents: How do I arrange a business trip? What is our limit for medical insurance? Where can I find the vacation request template?
• An internal chatbot in the corporate messenger that instantly answers typical questions about the company’s product or service, referencing current policies and documents: Can we guarantee drug A or service B? What are the recommended services for diagnosis A?
• A medical assistant for patient reception;
• A generator of short marketing and sales materials, with links to historical materials to speed up preparation for deals, improve the quality of proposals, and increase sales: Give me 3 arguments for our product against Competitor X for a retail client, and find a relevant case study.
• Etc.

What to know: Risks and limitations: accuracy, security, “hallucinations”

• Accuracy: 100% accuracy cannot be guaranteed. It is crucial to determine an acceptable level (e.g., 95%) and have a process to verify the remaining 5%.
• Data Security: When using cloud-based LLMs (such as GPT), ensuring data confidentiality is crucial. This can be achieved through an API with robust security policies or by deploying models in a private environment.
• “Hallucinations” in Search Assistants: One of the properties of AI is to generate statistically correct text. Sometimes this can lead to hallucinations (the AI confidently states something that does not exist). How to deal with them? Modern RAG systems can cite sources, enabling users to verify the answer.

Questions for the next AI strategy session:

1. Which specific department in our company spends the most man-hours on manually transferring data from documents into corporate systems?
2. If we could ask one question and instantly get an answer from all our company’s documents, what would that question be?
3. What is the estimated financial cost of a single error made during manual data entry in our procurement or invoicing process?
4. How much time, on average, does a new employee spend searching for basic information in their first three months, and how could an intelligent assistant impact this?
5. What is our most critical knowledge base (e.g., legal, technical, client-related) where access is currently a bottleneck dependent on a few key experts?
6. Which departments spend the most hidden resources on correcting errors, manual verification, searching for information, and reconciling data between systems?

Pilot Project (Proof of Concept – PoC)

Before implementing on a full scale, it is worth conducting a quick (2-4 weeks) and inexpensive pilot project on a very narrow task.

Why is this important for a manager?

• Minimization of risks: Allows testing a hypothesis without significant investment.
• Quick validation: Determines whether AI can effectively handle our specific documents and data.
• Building trust: A successful pilot is the best argument for allocating a budget for a full-scale project.

Example: Before automating the entire procurement department, select 100 invoices from three key suppliers and attempt to extract data from them in a pilot mode. Evaluate the accuracy and time. This will cost a minimum but provide maximum understanding.

How to measure success: specific KPIs

For Data Extraction projects, our KPIs are:

• Reduction in processing time per document: from 10 minutes to 30 seconds.
• Reduction in the number of errors: from 5% to less than 1%.
• Direct savings in the payroll fund: X man-hours per month.

For Search Assistants projects, your KPIs are:

• Reduction in time spent searching for information: e.g., the average employee saves 2 hours per week.
• Reduction in the number of requests to key experts: by 40%.
• Acceleration of onboarding: new employees find answers independently, reducing adaptation time by 30%.

Who is responsible: The Pilot Team

For a successful pilot, you don’t need a large AI department. It is enough to assemble a small, agile team:

• Process Owner: An employee or department head who suffers the most from the current problem. They are familiar with all the nuances and will be the primary stakeholders.
• Technical Specialist / Partner: An internal IT specialist or an external contractor who implements the pilot.
• End Users (2-3 people): Employees who will test the solution and provide feedback.

The path to effective AI use can be clear and pragmatic. Instead of risky innovations with no clear idea of the future result, we can focus on the systematic implementation of two proven project patterns that are guaranteed to bring financial benefits and increase our operational efficiency.

Our first step is not to look for an AI specialist, but to answer the strategic questions listed above. We choose the ONE most painful process. We formulate a hypothesis for a pilot project. And with this task, we turn to the technical team or a contractor. We start small, measure the result, and then scale the success.

‍

‍