THE KEY ANSWER
Start with a repeatable process with a clearly defined outcome, available examples, and a person responsible for quality. First, measure the current work. Only then decide whether you need AI, integration, or the organization of rules.
Where to look for the first process?
Ask three people from one team to write down the tasks they return to every day. Do not ask yet what they want to automate. Ask where the data comes from, where it needs to be re-entered, and who is waiting for the next step. This way, you will find work invisible in official procedures: checking attachments, looking up customer numbers, correcting similar offers.
Choose a process that starts with a specific event and ends with a verifiable result. “Improve sales” is too broad a goal. “Prepare a draft offer from an incoming inquiry” provides boundaries. Separately, record the standard flow, exceptions, and situations where someone must make a decision. If every case requires a new interpretation of company rules, you first need those rules.
Context and references: DORA: State of AI-assisted Software Development 2025
AI, rules, or integration?
Check what causes the difficulty. Transferring an order number between two systems usually requires integration. Reading intent from an ambiguous email may justify a language model. Calculating a price according to a price list should remain in verifiable business logic. One process may contain all these elements, without handing over all responsibility to AI.
Demonstration example: an inquiry arrives in the inbox, the model organizes the requirements, the application checks the customer in the CRM, and the calculator computes the price. The salesperson sees missing information and the source of each value. A human approves the dispatch. This boundary makes it easier to find errors: it is known whether text understanding, data, or calculation failed.
How to measure the starting point?
Collect a sample covering several normal days and at least one more difficult period. Measure active work time, waiting time between steps, the number of corrections, and the final result. The average alone may hide cases that block the department for several hours. Therefore, also record the type of case and the cause of the delay.
Compare similar tasks, not a simple pilot week with a seasonal peak of the old process. Add to the post-implementation time the checking of results, corrections, and handling of failures. Recovered minutes do not automatically mean cash savings. Determine what work the team will use the recovered capacity for: faster response, a larger number of offers, or organizing customer relationships.
When does a pilot deserve further development?
Before launch, list the conditions for three decisions: develop, improve, or stop. Good quality of a single presentation is not enough. You need stable performance on new examples, acceptable control effort, and a person who will take over the process after implementation.
Initially, the system can run in the background and prepare results without affecting the customer. The team will compare them with their own work. Only after such verification should limited traffic be enabled. Establish a method for returning to manual handling and a queue for exceptions. If the pilot does not meet the conditions, narrowing the task is a valuable decision. There is no need to save the investment by adding more features.
WHERE TO START
Bring this into your project.
- Indicate the start, end, and owner of the process.
- Collect typical, difficult, and incomplete examples.
- Calculate time including control and corrections.
- Establish continuation criteria and a manual handling plan.
Choose one thing your process is missing today. It's a useful topic for your first conversation with the team.
QUESTIONS AND ANSWERS
Frequently asked questions.
Can a small company start with AI?
Yes, if it has repeatable work with sufficient volume. Company size is less important than the availability of examples, clear rules, and the ability to use recovered time. With only a few cases per month, a simpler improvement may be more reasonable.
Do you need to have a ready specification?
No. To start, a process description, sample materials, and a conversation with the person performing the task are sufficient. The specification should arise from observing actual work, not from a list of trendy features.
Sources and context
- DORA: State of AI-assisted Software Development 2025 ↗
DORA describes AI as an amplifier of the existing strengths and weaknesses of an organization. The above method of selecting a process is a practical proposal by ALGOV.
Prepared by the ALGOV team. Current as of September 8, 2026. Examples describe possible scenarios, not results from client projects. How we create our guides.
