First step
One process to start: how to choose your first AI deployment
Once the decision to deploy AI has been made, the most common question is: where do we start? We propose a deliberately narrow scope: one process, one measure and one person responsible. Below, we explain how to find such a process in your company.
Why only one?
Five parallel deployments require five teams to be involved and five data scopes to be agreed. Responsibility for the result is easily diluted. When the board asks about the effect after six months, it receives a presentation instead of one number showing the change.
With one process, the data scope is small, so preparation is quick. One person is responsible for the process, so there is no need to wait for decisions. The result can be stated in numbers and compared with the situation before the deployment. The one-process limit applies only at the outset. After the first measured result, further processes can begin in parallel at a pace set by the company. The agreed access and measurement rules can be applied to the next processes.
The marks of a good candidate
Both traits have to be present.
The first is a large number of similar cases. The process should repeat many times a month, often enough for the saving to justify the deployment cost. We calculate the threshold from company data, because it depends on the value of each case and the work involved.
The second is a measurable result. There must be a number the deployment is meant to change: the time taken to handle one case, response time to an enquiry or the error rate in retyping. Without that number, it is impossible to say whether the deployment changed anything.
What a good first process looks like
Documents arrive in a mailbox or folder, and a person reads them and retypes the data into a system. This is how orders, purchase invoices, complaints and requests for quotation are handled, among other processes. The initial effect is easiest to measure in processes like these.
The start, step by step
First, we measure the baseline: the current value of the metric. We then agree the data scope and read-only access. We discuss every field and explain why it is needed. Next, our AI layer begins to read cases and send the results to people for approval.
At the end of the period set out in the agreement, usually two or three months, we compare the result with the baseline. The company decides what happens next. If the agreed threshold is reached, the scope can be expanded. If it is not, the work ends at that stage and the next stage does not begin. The cost of the check is known in advance. The fee for work already done remains payable, but the company does not pay for the next stage.
What not to pick first?
At the outset, we do not choose processes involving decisions about people because they require the greatest oversight and involve the most sensitive data. We also advise against processes without a person responsible, because there would be nobody to approve the results. Processes the company is currently redesigning are also a poor choice. With a moving target like that, the result cannot be assessed clearly afterwards.
The process that came to mind while you were reading is usually a good candidate. Send a short description through the contact form or to kontakt@uruchom.ai. State who currently runs it and how many documents pass through it. Within one working day, we will send three questions about the process and propose a 45-minute call. We will prepare the first-stage outline after that call. It will assess whether the process is suitable as a first step.
Back to all posts