AI implementation
AI is meant to shorten a specific process. First we check where it pays off, then we launch the solution so that it still works in the third month.
“Everyone is implementing AI. Where does it make sense for us, and how do we launch it so that it still works a month later?”
Note for decision-makers
- Assessed processes, the first of them selected and a solution working in it, with before-and-after measurement and a person responsible on the company's side.
- It is clear which process hurts, or there is an “AI implementation” offer on the table and nobody knows whether it is a project or a licence sale.
- The goal is to “have AI”, with no process identified and no measure of success. Or the process has not yet been described: in that case we start with process analysis.
- What issues do people in this team deal with most often, and who is responsible for this process, by name?
Stage 1: AI audit, or where it pays off
We start with a list of issues, not a list of tools. We assess each process against four criteria: how often it recurs, how much the cases differ, whether the result can be checked and how much a mistake costs.
Sometimes the right answer is a simple script. If a task is repetitive and unambiguous, a program will do it faster, cheaper and more predictably than AI. We recommend AI where interpretation is needed.
The model itself is the cheapest part of the project. Most of the work goes into access to data: integrations, permissions, putting the data in order. The audit shows this cost before the budget decision is made.
Result of this stage: a list of assessed processes, the first candidate for implementation, a measurement of its current state and a list of places where AI makes no sense. If the process has already been selected and measured, we skip this stage.
Stage 2: implementation
When we consider a solution implemented
If any of these conditions is missing, it is a demo, not an implementation.
Access to tools. The solution works in the systems where the work happens and has access to what it needs to finish the task.
Closed loop. It sees the effect of its own action and can correct it.
Completion criterion. It is clear when the task is done, and clear what happens when it cannot be done.
Three principles
The solution must work where people already work. In e-mail, in the ticketing system, in the spreadsheet. A tool that has to be opened separately stops being used after a few weeks.
Permissions grow gradually. Read access first, write access later. Irreversible actions, such as sending a message to a customer, a payment or deleting data, always require confirmation by a person.
Without a set of test cases, there is no way to know whether a fix helped. Before the solution sees real cases, we check it against a set of cases with an agreed correct result.
How the work proceeds
Audit: list of issues, assessment of processes, selection of the first one and measurement of the state before.
Design: what data, what permissions and how we check the result.
A set of test cases taken from the company's real cases.
Launch in one team.
Gradual extension of permissions.
Measurement after implementation and comparison with the state before.
What you can count on
An assessment of cost-effectiveness based on measuring the process. The result shows which processes are worth automating and which will not recover the cost of implementation.
What the price depends on
The number of teams and processes in the audit. The complexity of the selected process and the number of systems to connect. The state of the data. Availability of a technical team on the company's side.
Let's talk about your case.
The shortest route is one sentence about what you are trying to decide. We will get back to you on whether and how we can help.
Tell us which process needs improving and who is responsible for it.
Write to us →