AI & automation

Implementing AI in a company step by step — from process to pilot

Maciej Zaleski · Published · Updated · 7 min read

The best way to start implementing AI in a company is with one repeatable process whose output can be checked and compared with the work done so far. Only then should you choose a model, tool, and integration method. This sequence helps establish whether you need generative AI, conventional automation, or first to organize your data.

For management, the most important question is not “how do we use AI?” but “which operational problem is worth solving, and how will we recognize improvement?” The following plan leads from task selection through a pilot to a decision on further development. It does not assume that every trial should end in production deployment.

Implementing AI in a company starts with choosing a process

A good candidate is a regularly performed task based on available materials with an identifiable output. This could be classifying tickets, drafting a reply, or reading information from a document. Initially, avoid a process where a single error automatically creates a commitment that is difficult to reverse.

The process owner should describe its start, end, and exceptions. “Improve sales” is too broad a goal. “Prepare a draft summary of an inquiry and identify missing data” lets you define the input, expected result, and person responsible for approval. Only that kind of description is suitable for discussing a solution.

Work through real cases with employees rather than base the analysis solely on procedures. Informal checks, calls to another department, or manual corrections often reveal the actual difficulty. A model will not remove a problem caused by a lack of decision ownership.

Check whether AI is actually needed

If the task comes down to stable rules and structured fields, a simple script or integration may be easier to maintain. AI can be useful where natural language, varying document layouts, or contextual interpretation are involved. Both approaches can be combined in one process.

Before starting work, answer these questions:

  • Who uses the result and makes the final decision?
  • How are successful cases and exceptions handled today?
  • Do we have examples of accepted and incorrect outputs?
  • What data may be sent to the chosen environment?
  • Is there a safe route back to manual work?

Data and responsibility before choosing a model

A data review covers availability, quality, freshness, and permissions. Documents merely existing on a drive does not mean they are ready for use. Establish which versions are current, whether they contain confidential data, and who may see information from different systems.

Prepare examples representing normal work, difficult cases, and out-of-scope situations. Do not limit yourself to materials chosen because they look good in a presentation. Reserve some examples for final evaluation so you do not test the solution solely on data used while tuning the process.

Assign a data owner, a business owner, and a person responsible for technical maintenance. Company employees and an external partner may fill these roles, but responsibility boundaries should be clear. Without them, it is hard to determine who approves instruction changes, corrects a knowledge source, or responds to an incorrect result.

If personal data or regulatory issues arise, current GDPR, AI Act, and industry regulatory requirements should be verified with a specialist. Choosing a provider or installing in your own environment does not replace this analysis. Decisions on access, retention, and incident handling are also needed.

How to design a pilot

A pilot should test a specific hypothesis, such as whether useful draft replies can be prepared for a particular ticket type. Describe scope, exclusions, evaluation methods, and completion conditions. Also specify which activities remain manual even if they could technically be automated.

Initially, run the solution in observation mode. The system prepares an output but does not change production data or send messages without approval. This reveals its behavior on real traffic before it receives permission to take action.

Pilot area What to agree before launch
Scope One process and specified case types
Quality Criteria for correct output and unacceptable errors
Oversight Approver and escalation method
Cost Launch and maintenance cost components
Decision Conditions for continuing, improving or stopping

If you plan a tool that takes action in systems, also read how an AI agent in a company differs from an assistant generating text. Permission scope should follow process risk, not the tool's demo capabilities.

Measuring results and ROI without shortcuts

First measure how work is done before implementation. Include task execution time, waits between stages, corrections, and the share of cases needing additional review. Compare similar case types. Otherwise, a change in the mix of incoming tasks may be mistaken for a technology effect.

After launching the pilot, measure the entire process time, not just response generation. A quick draft is not a benefit if an employee must spend a long time finding its errors. Assess substantive correctness, completeness, usefulness, and security separately. An average score should not hide rare but serious mistakes.

Include analysis, integration, model usage, infrastructure, testing, security reviews, and supervisors' work in costs. For ROI, compare expected financial benefits with the solution's full cost. Treat recovered time as organizational potential until it is clear how it will be used in practice.

Prepare a conservative scenario and one based on broader use of the solution. Do not automatically assume results from simple cases will repeat across the entire company. A pilot should reduce uncertainty, not justify a purchasing decision already made.

Common mistakes when moving to production

The most common organizational problem is a lack of ownership after the experiment. Documents, forms, models, and user expectations change. Someone must run regression tests, analyze reports, and decide on updates. Simply handing the application to IT does not resolve business responsibility.

The second mistake is excessively broad data access. An assistant should use only information needed for the task and respect user permissions. Also separate preparing a proposal from executing it. Human approval makes sense when the person can see sources, changes, and potential consequences.

The third mistake is failing to handle outages and exceptions. Define behavior when the model is unavailable, data is missing, or an output fails validation rules. “I cannot determine this” can be better than an apparently complete answer. For document processes, we discuss the details in our article on automated document processing.

FAQ

Which department should start implementing AI?

No department is right for every company. Choose an area with an engaged process owner, available examples, and a way to assess results safely. A small, well-described process can be a better starting point than the area attracting the most internal attention.

Does all data need to be organized first?

You do not need to organize the entire company to test a specific hypothesis. However, you must prepare reliable data for the chosen scope and know what is missing. A pilot should not conceal quality problems by manually fixing every example outside the agreed process.

Is a ready-made model enough without training your own?

First test available models, appropriate instructions, and context supplied from company sources. Training or fine-tuning is not a mandatory stage. Test results and data, cost, and maintenance requirements should determine whether this work is needed.

When should a pilot be stopped?

When the solution fails agreed quality criteria, cannot be safely incorporated into the process, or requires oversight that cancels out its benefits. Stopping can be a valuable outcome: it reveals limitations and allows the budget to be directed toward a better-defined problem.

Who should approve production launch?

The decision should include the perspectives of the process owner, IT, security, and, depending on scope, the person responsible for legal matters. Approved error-handling and maintenance procedures are also needed. A positive reception of a demonstration alone is not sufficient grounds.

Summary

An effective approach to AI starts with a clearly defined task, prepared data, and an honest comparison with current work. A pilot tests quality, costs, and required oversight before the company expands automation. Willingness to change or stop an unsuitable idea is just as important as the technology.

If you want to organize the first step, explore our approach to AI & automation. During a conversation, we can walk through a selected process and establish what information is needed for a meaningful pilot assessment.

MZ

Maciej Zaleski

Managing Director, Go2ops

Related articles