Business ·
AI enters a business through one process
AI implementation in business often begins with a model, a subscription or a supplier. Technology does not know where a company loses time, money and attention. The first step is to find one process with identifiable participants, repeated actions and a visible cost of delay or error.
Buying access does not implement anything
Employees may receive a corporate AI chat and continue working almost as before. One person writes emails faster and another prepares a presentation, yet the company's process remains unchanged. Individual use of a tool and implementation in a business are different scales of work.
Implementation begins when the sequence of actions changes. Information arrives in a defined form, the system performs part of the work, a person reviews the result, and the next action is recorded in the environment the company already uses. Without this loop, AI is another window beside email, spreadsheets and the CRM.
The question of which model to choose can therefore wait. First, the work that needs to change must become visible.
Examine the process before automating it
Every process has an input, participants, decisions, exceptions and an outcome. An enquiry may arrive through a form, be clarified by a manager, turned into a proposal, negotiated and assigned a status. If these steps remain implicit, AI automates an imagined version instead of the real work.
Workarounds matter: messenger threads, private spreadsheets, verbal approvals and tasks completed by a single experienced employee. They often contain both the reason for a delay and the knowledge without which an automation will fail.
This examination does not require months of research. Follow several real cases from beginning to end and mark where people wait, copy, search, retell and double-check.
Choose the first process by four properties
A good candidate repeats often enough, has a clear outcome, uses accessible data and allows human review. Sorting incoming requests, preparing a draft reply, extracting facts from documents or summarising changes are better starting points than a rare strategic decision with a high cost of error.
Define the measure in advance. Processing time, the share of corrected outputs, missed enquiries or report preparation time can be compared before and after a pilot. “The team finds it more convenient” does not show whether the business changed.
The first process does not need to deliver the greatest possible saving. It needs to test the method of implementation: data quality, integration, control and the team's readiness to work through the new loop.
A pilot must work on real cases
A demonstration shows a possibility. A pilot tests that possibility inside a particular process. It needs ordinary cases, difficult exceptions and a predefined boundary at which work returns to a person.
Keep the source data, the system's response, the employee's correction and the final action. This history reveals whether an error came from the model, an incomplete instruction or contradictory rules within the organisation itself.
A pilot ends with a decision: expand, revise or stop. An endless experiment without a success criterion simply postpones the difficult choice until next month.
A person remains part of the system
Control is more than an approval button. The employee must know what to inspect, which errors are critical and where to send a case when the available information is insufficient. The more serious the consequences, the clearer this role must be.
It helps to separate suggestions from actions. AI may classify an enquiry, draft a response or propose a change. Sending a message to a customer, changing a price, making a payment or deleting data requires a distinct level of authority and observation.
This loop does not make the system less automated. It makes the system suitable for a business in which an error has an owner and consequences.
Scale after a measurable result
If a pilot reduces time while preserving quality, the next question is resilience. Who owns the instruction? How is knowledge updated? What happens when the model is unavailable? Where is the action log stored? The answers turn an experiment into a working product.
Only then should the system take on new cases or an adjacent process. Copying a solution across a company too early is dangerous: differences in data and rules quickly multiply errors that were still visible in a small pilot.
Within the Method, I treat AI implementation as the design of a new form of work. Technology becomes a material, while the Concept defines the change in process, the human role and the result the business can verify.
AI for a specific process
If you are considering AI implementation, we can start with the process, available data, automation boundaries and a way to assess the result.