6 tips – How to teach AI to work the way you want
Kirjoittaja:
Ella Lindroos
You always get an answer from AI, but the answer does not necessarily match your organisation’s instructions and operating principles. General language models simply do not know how your organisation works.
In this article we go through how to get AI to work in line with your own organisation’s operating principles and processes.
1. First, make sure AI has been given a high-quality, up-to-date context in which to operate
A general-purpose language model has been trained on a broad body of material, but language models do not automatically have the operating principles or processes of individual organisations available to them.
AI does not know who in your organisation decides on refunds, which instruction is in force, or which matters require a deviation report. When information is missing, the model fills the gap with the most probable answer.
You therefore have to guide the general language model to work the way your organisation works.
AI does not learn like a human. It becomes proficient through data and different kinds of instructions.
2. The guidance you give AI must meet these 4 requirements
An instruction alone is not enough if AI cannot know whether it is in force. A good instruction for AI is like a good instruction for a person: current and available.
Approach the guidance you offer for AI use in the same way as the guidance you offer people:
- Information is up to date: The instruction is the current, valid version, and old versions of the instruction have been clearly archived.
- An owner has been defined for the guidance: Every instruction has a responsible person who updates it when operations change or when the review interval has passed.
- The guidance has been approved by a human: The instruction has been reviewed and approved by a person, and you have access to the log of approvals and changes.
- The guidance is linked to the related processes: The instruction is linked to processes, roles and systems, so AI knows who the matter belongs to and where to find more information.
3. Three examples of sharing information with AI
There are 3 options for sharing information with AI:
- Attach the instructions and the necessary data separately to every prompt.
- Connect your AI tool to your document folder.
- Give AI access to the information in your management system.
| How it works | Benefit | Risk/opportunity | |
|---|---|---|---|
| 1. Attach the instructions to each question. | You copy the instruction to AI every time. | Quick to start, requires no tools. | Each user uses a different version, and the instruction goes out of date unnoticed. Information must be updated manually. |
| 2. Connect AI to a document folder. | AI looks for the answer in folders and the intranet. | The instructions already exist, and they do not need to be moved. | The folder contains contradictory and old versions, and nobody knows which one is in force. Information must be updated manually. |
| 3. A described and approved management system for AI use | AI reads the instructions from a place where they have been described, owned and approved. | One current version, clear responsibilities and traceability. | Requires that operations are described and the description is maintained, but maintaining the information can be automated with the help of AI. |
4. Raw data helps AI understand the big picture
Raw data is the quickest way to get started. When you give AI Excel exports, instructions, meeting minutes or email chains as they are, it gets a lot of information but does not know what to believe in it.
Raw data works well as source material: AI can turn it into a draft that a person checks, corrects and approves.
Only the approved and owned version should be made permanently available to AI.
Remember data protection, too: do not enter confidential data into a tool whose data protection terms you do not know.
5. How do you move from entering information manually to automated information sharing?
Entering information into AI manually does not make sense in the long term, because that way you do not scale the benefits of AI.
In addition, if you always have to copy the instruction to AI, the information is only as current as the person copying it manages to keep it, and each person using the information will probably use a different version.
Automated access to the information that guides work means, in practice, that AI reads the information directly from the place where it has been described, approved and kept up to date. In addition, at this point AI can be given the same access to the information, and to updating it, that the system in question offers to its human users.
6. A management system that meets the requirements of the age of AI works as AI’s source of information
A management system suited to the age of AI is your solution for automated information sharing.
The management system works as the organisation’s official source of information, from which people and agents read the same valid version through an interface, and nobody has to copy anything by hand.
This also gives you the opportunity to use AI’s help in updating information in a way where AI gives you its proposal, and only after your approval does the AI agent update the information in the management system.