Hello. I am in charge of Cinnamon AI public relations.

AI products are attracting increasing attention in response to the recent focus on improving business efficiency and work style reform, but on the other hand, when it comes to progressing a project, there are times when a PoC (proof of concept) is carried out, and accuracy remains unresolved even after implementation. It is not well known that it is unique compared to conventional system implementation, such as continuing to learn as it progresses, and there are cases where it is difficult to consider when considering it.

In this article, we will first explain the differences between an AI project and the introduction of an existing system, and then provide an overview of the data and PoC required for learning, which is a process unique to AI development.

Differences between AI and traditional system implementation

There are three major differences between AI and traditional system implementation projects:

①Policy

AI development:

We conduct POC (proof of concept) based on the established hypothesis and evaluate it for development. Projects that have achieved the requirements and target accuracy will proceed to the actual development and implementation step.

The following article introduces the reasons why consideration ends at PoC.

“Three reasons why AI implementation projects fail”

Depending on the scale of the project and scope of requirements, evaluations may be conducted several times, such as 2nd PoC.

Traditional system development:

Each system has a certain degree of certainty in its solution to the client's problems, so the project progresses according to the existing development roadmap. Basically, what the client wants to do and what the system can do match, so although we may introduce it on a trial basis, we do not conduct PoC in most cases.

②Quality approach

AI development:

AI is replacing tasks that should be performed perfectly, such as reviewing contracts and reading invoices, but technically there is still no AI that can process tasks with an accuracy of 100%. not. Therefore, in order to proceed with development from the perspective of whether the expected system has been achieved, it is necessary to review the work flow design.

Traditional system development:

Quality is guaranteed according to SLA (Service Level Agreement).

③How to proceed with development

AI development: Proceed with projects using agile development.

Agile development is a project development method for system and software development, in which development is carried out by repeating implementation and testing in small units, without dividing the system into large units. It is called agile because it takes less time to develop than traditional development methods.

Quote source:https://hnavi.co.jp/knowledge/blog/agile_software_development/

Traditional system development: We proceed with projects using waterfall development.

The waterfall method is a method in which system development is divided into the following stages: basic planning, external design, internal design, program design, programming, and testing. Since it is assumed that there is no return to the previous process, it is called a waterfall, likening it to the flow of water that does not return from downstream to upstream.

Quote source:http://www.okapiproject.com/computer/leran_comp/sysdev/sd_01_003_0.htm

What data is required for AI development?

Next, we will explain learning using data, which is one of the processes unique to AI implementation projects.

 In traditional system implementation projects, the vendor mainly performs the development and installation, and the amount of work on the client side is not very large.

On the other hand, in most AI development projects, the client side must prepare the training data for learning (vendors often do not have a large amount of diverse training data in-house). Also, once the necessary training data is prepared, the next step is annotation work that gives meaning to each item in the data. In this way, learning AI development often takes a lot of man-hours, and frequent exchanges occur between the vendor and client sides.

Although it is technically possible to amplify sample data from a small amount of training data, a successful AI development project will not be possible unless the client and vendor cooperate.

What is PoC in AI development?

PoC is the process of verifying the feasibility and effectiveness of a new idea or theory before full-scale development begins.

Unlike traditional system implementation, AI implementation projects are difficult to see the requirements and qualitative results in advance. If you proceed with the actual development, you will incur a huge loss if the project goes awry. In order to reduce such accidents, we move on to actual development after making detailed adjustments, such as whether the direction of the project is correct and whether there are any points that should be reconsidered.

summary

We also explained the positioning of PoC while comparing AI introduction with conventional system introduction.

I hope this will be helpful for your introduction. At Cinnamon AI, we tune the AI engine for each customer, so PoC is positioned as early development. We also provide consulting to maximize the resolution of customer issues through the introduction of AI, so please feel free to contact us. For inquiries, please click here.

 In addition, in order to respond to the voices of people who have considered AI-OCR, but have stopped at POC and are not making any progress in their consideration, we regularly hold seminars.

(Written by: Morita)