Step 1. Numbers before the decision
Before any decision, an agent collects numbers from your systems: how many leads there were, what the conversion is and what the process costs now. If the numbers do not exist, that is the first finding: the process is not measured, and it is too early to decide.
Step 2. A pilot instead of an argument
A hypothesis is tested on one process with a metric. For example: answering a lead in one minute instead of one hour will raise the share of leads that reach a deal. The pilot runs 14 days, the metric is written in advance, and the result does not depend on who argued louder.
Step 3. A decision from the report
At the end of the pilot you get a report with “before” and “after” numbers and three options: stop, keep on support, or scale. If there is no payback, that is also a result, and the report records it.
Which decisions suit a pilot
Pilots suit decisions about processes that have a flow and a number: hire a person or use an agent, change the order of answering leads, open a new request channel. Pilots do not suit decisions that cannot be tested in 14 days on one process, such as a strategy change, entry into a new region or a large purchase. For these, AI prepares the numbers, and the test needs another method.
How to phrase the question for a pilot
The decision must read as a choice: hire a manager or use an agent, keep the process or change it. The choice needs a number that shows after 14 days which option won. If you cannot name such a number, the question is not ready to test.
What data you need for data driven decisions
Data driven decisions in a small business need no data warehouse and no analyst on staff. Three sources that you already have are enough: your accounting system (1C), your CRM and your bank. The agent reads them with a separate account and collects numbers on leads, deals, cash and debts. You do not need to change systems.
An example of a decision from a pilot
The question: hire a second manager for leads, or use an agent? Numbers before: leads per day, leads with no answer for over an hour, and the cost of a manager with taxes. Pilot: an agent takes leads for 14 days, and the metrics are reply time and the share of leads in the CRM. Decision: from the report, with a clear price for each option.
Where AI makes mistakes and how to allow for them
An agent can misread a question or take the wrong number. So every report names its data source, disputed cases go to a person, and in the first weeks the numbers are checked by hand. The responsibility for the decision stays with the executive.