When a company decides to implement AI, contractors usually offer three formats: consulting, custom development and ready-made products. They solve different problems and have different costs. This article explains what each format provides, how long it usually takes and when to choose it.
Consulting and strategy
Consulting is useful when a company understands that AI could help but does not know which process to start with or what the benefit might be. The consultant’s task is to find areas where AI can create a measurable effect and prepare a realistic plan that takes the company’s processes, data and IT systems into account.
The result is a list of priority initiatives, a solution architecture and an implementation roadmap. This work usually takes at least a month and a half.
How the assessment works
It is useful to divide the assessment into three stages of roughly two weeks each.
- Business processes. Interviews with managers and process owners, followed by an analysis of key operations, metrics and labour costs. The result is a process map, a list of possible AI scenarios and a preliminary estimate of the impact.
- Data and systems. An audit of data sources, their quality and availability, as well as the systems, APIs and integration constraints. The result is a map of systems and data, integration requirements and an assessment of the company’s readiness.
- Portfolio and roadmap. Initiatives are compared by impact, complexity and risk, and pilots are designed for the strongest candidates. The result is a prioritised list, pilot concepts, stages, timelines and a preliminary budget.
The second stage is often underestimated. A scenario may look profitable on paper, but if the required data does not exist or a system cannot provide it through an API, the project will run into integration work.
Custom development
Custom development is needed when a company has a specific problem and no ready-made product fits it. Typical areas include:
- AI agents and business-process automation;
- corporate knowledge bases with document search (RAG, retrieving information from a database and generating an answer from it);
- machine learning and recommender systems;
- computer vision and image processing;
- integration with internal systems and data sources.
The first working version usually appears within a few weeks. The solution then moves from a pilot to production use. It is worth agreeing in advance who will support the system after launch: the company’s team or a contractor under a service-level agreement.
Ready-made solutions
The market offers ready-made products for common tasks. They are configured for the company’s processes and connected to its systems. Common examples include:
- customer-service assistant: answers customers in messengers and on the website using a knowledge base, accepts requests, performs actions in the CRM and passes a complex conversation to an operator together with its history;
- sales quality control: transcribes managers’ calls and messages, checks them against scripts and checklists, and shows violations and missed opportunities for each employee;
- visual search for online stores: finds products from a photograph and recommends similar or complementary items.
Connecting such a product takes from several days, and the pilot runs on the company’s data. The product can be hosted in the cloud or in the customer’s infrastructure, which matters to companies with strict data-security requirements.
Where AI is already producing results
Five scenarios that most often reach a working result:
- Call monitoring. Previously, a sales manager listened to calls selectively. The system retrieves recordings from the CRM, transcribes the conversation, separates the manager’s and customer’s lines, and checks every call against a checklist. The manager sees the full picture and can give feedback on specific fragments.
- A knowledge base for managers. Prices, terms and document requirements are scattered across different places, so the speed of a customer response depends on the manager’s experience. An assistant searches the knowledge base and reference materials and answers with a source link. Managers rely less on experts and use consistent wording.
- Management reporting. Revenue, procurement and expense data sits in different systems, and preparing a profit-and-loss report takes days. The system brings the data into one structure, calculates metrics and explains deviations.
- Document processing. Employees manually find details in scans and transfer them into forms. The system recognises documents, extracts the data and prepares a filled-in document for a specialist to check. Preparation time and transcription errors are reduced.
- Production safety monitoring. Manual camera review misses some violations. The system analyses the video stream, checks for helmets and entry into restricted areas, and records every violation.
How to choose a format
- If you do not know where to start: assess processes and data.
- If the task is typical (customer service, call monitoring or catalogue search): start with a pilot of a ready-made solution.
- If the task is specific to the company: start custom development with one process.
The formats work well together. An assessment shows which tasks can be covered by ready-made products and which require custom development.
Whatever format the company chooses, I recommend starting with one process and one metric that can be measured before and after launch. I described how to find such a process in the article “How a business owner can find AI opportunities in their company”.
This text is based on an analysis of the business of my colleagues at practico.ai.
If you are not sure where to start, fill in the AI implementation brief to structure the task and choose the next step.