AI, in the terms of your firm
Twenty-two words people use about AI, explained the way an advice firm already works.
What are you actually using?
3 of 6 written
- ModelThe part of an AI product that writes. A model provider trains it, using it doesn't change it, and a newer model is a different model.
- AI-nativeSoftware where the AI works on the screen you have open and can do what you can do there, with your permissions and no more.
- AI-enabledSoftware where people do the work and the AI sits in particular places, like a side panel. The AI works on what you or the product gives it.
- Chat app vs AI inside a product · Model provider · Where client data goes
Why does it get things wrong?
1 of 4 written
- HallucinationA statement an AI makes up and presents as fact. It reads just like the accurate lines around it.
- Training · Non-determinism · Sycophancy
What sits around it?
8 terms
- HarnessThe software around an AI model. It decides what the model is given, what it may do, and which checks run on its work.
- PromptThe request you give an AI for one piece of work. If nothing says what the work is for, the AI guesses what matters.
- InstructionsThe standing directions a product sends the AI with every piece of work. The AI weighs them and usually, not always, follows them.
- ContextThe material an AI is given for one piece of work. Everything it knows about your client comes from it.
- MemoryWhat an AI product saves from earlier work and adds to later work. Many products keep it for the person using them, not for each client.
- ToolAn action a product lets its AI take, such as looking up a record or adding a task. Without one, the AI can only write a reply.
- AgentAn AI that works towards a goal by taking actions in your firm's systems. It chooses its own steps, so nobody can read them in advance.
- PermissionThe limits software sets on what an AI can read, change or do, and which actions wait for a person's approval. They can be wider than yours.
How do you check its work?
2 of 4 written
- Wrong-subject claimA statement where the AI copies a figure or fact correctly but gives it to the wrong person or policy. Checking the figure alone passes it.
- Human reviewA person checks the AI's work against its sources before the firm relies on it. Clicking approve is not the same thing.
- Source · Transcription error