Breadcrumbs

How the AI carries out a task

Along with its tools, the product gives the AI instructions: how to choose the way to do a task, and rules to follow in the conversation. This page describes them, so you know what to expect.

When your AI client connects, the product sends it a short version of these instructions. The full version comes from the product's <product>_get_started tool, for example sapio365_get_started or scanez_get_started, which the AI is told to call before any other tool of the product, and again whenever it is unsure of the state. That tool also tells the AI what the product holds at that moment and which restrictions are on. sapio365 and every EZ Suite product except signEZ give these instructions.

Rules every product gives

sapio365 and the EZ Suite products give the AI these rules.

  • It takes the shortest path. The AI prefers a tool that does the task directly, then an existing job (sapio365 only) or an automation script file you already have, and writes a new automation script only when nothing else can do the task. It counts the steps over the whole task, confirmations and lookups included.

  • It keeps a task in one automation script. When one step of a task needs an automation script, the steps before it go in that same automation script. This keeps answers fast and uses fewer AI tokens.

  • It does not guess names. Grid names, column IDs and, in sapio365, job keys come from the product or from the Ytria documentation server, not from memory.

  • It reads the documentation before writing an automation script. It reads the automation documentation before its first automation script of a conversation, and looks up every action and attribute an automation script uses.

  • It checks the state again. The product is a live application: the AI asks for the current state rather than trusting an earlier answer.

  • It searches a grid rather than writing an automation script. To select the rows matching a pattern, or to count them, the AI searches the grid as Find All (Ctrl+F3) does, then reads the rows it selected.

  • It exports only when you ask. The AI writes a grid to an Excel file only when you ask for a file; to show or analyze data, it reads the rows from the grid.

  • It finds out why something failed. The AI reads the automation trace of the run first, then the application trace around the error.

When you restrict the AI

Each restriction ticked in AI Settings adds a rule, so that the AI does not write an automation script that would stop on it.

  • Forbid saving data with AI: a Save or SaveSelected action, in an automation script or a job, ends in error and stops the run; the changes made before it stay pending. The AI puts no such action in an automation script it writes, says so before running one that holds one, and tells you that saving the changes is up to you.

  • Forbid reading data with AI: no grid content reaches the AI. It can still describe a grid, search it, count rows, and run automation scripts and, in sapio365, jobs, but it does not try to obtain cell values any other way. When a task needs the data itself, it tells you that reading it is up to you.

  • Forbid session load with AI (sapio365) or Forbid switch ID with AI (EZ Suite): the AI works with the session already loaded. A LoadSession or ConnectToCustomer action (sapio365) or a SwitchID action (EZ Suite) ends in error and stops the run, and the AI tells you that loading a session is up to you.

See Security, restrictions and your data.

Important: these rules guide the AI, but they do not replace your judgment: AI models can make mistakes. Keep your AI client's approval prompts on for actions, read what the AI says it will do before you accept, especially when data can change, and try a new automation script on a small selection first. To review every change before it is saved, tick Forbid saving data with AI.

In sapio365

The shortest path

sapio365 tells the AI to consider the options in this order:

  1. Built-in tools only: count what the dashboard counts, open a module with one of its views, show a panel, refresh a module, apply a saved view, save the grid as a new view, select the rows matching a pattern, read the selection, build a report, export. This holds only when every step of the task is a tool. A saved view that already shows the right columns, sorting, grouping and filters replaces any automation script that would rebuild them.

  2. An existing job: written and tested by Ytria or by you, it runs in one step, and you can run it again later from the Jobs list.

  3. An automation script file you already have: run as it is, rather than rewritten.

  4. Only then, one new automation script covering the whole task: written from the documentation, starting from the closest existing job, and checked against the job schema before it runs. Once it works, the AI offers to save it as a job, so the same request is quicker next time. When all an automation script did was rearrange the grid, saving the result as a view does the same: anyone can then apply it in one step.

For example, “How many guests do we have?” is a count the dashboard shows: one tool, with no module to open. “Show me the guest accounts” is Users opened with a saved view, with no automation script. With Users already open, “select the guest accounts there” is one search of the grid. “Open Users, select the non-guest accounts and give me their office location” is one automation script that opens Users and selects the accounts, after which the AI reads the selection back.

Modules and sessions

  • It asks which module you mean. When several modules are open and your request does not say which one it is about (for example two Users windows on two tenants), the AI asks before reading from or acting in any of them.

  • It knows which tools need an open module. A tool reads from or acts in a module window that is already open, the dashboard's counts aside. sapio365_build_quick_gridview opens a module with one of its views. An automation script opens the module it needs by itself, so the AI does not open it with a tool first.

  • It says what will happen before it acts. Running a job or an automation script may change tenant data: the AI describes what it will do first. Applying a view changes the display only, never tenant data.

  • It knows the state of your session. With no session loaded, the AI tells you and offers to load the session you name: it shows the Recent Sessions panel, selects that session there and loads it. When sapio365 asks you to sign in, you complete the sign-in. With Forbid session load with AI, the AI shows the Recent Sessions panel so you pick one yourself. It loads another session, for another tenant, user or role, only when you ask for it. With a session that is not elevated, it knows that reading works and that a refused change is probably due to that.

Data age

  • It tells you how old the data is. An open module shows what it loaded at its last refresh, and data older than 30 minutes counts as stale. Before answering from a stale module, or a question about the current state of something, the AI checks how old the data is. It refreshes stale data with a quick update without asking, and asks before a full reload. Whenever its answer comes from data older than 30 minutes, the dashboard's counts included, it says how old it is.

Prompts to try in sapio365

  • “Which tenant is sapio365 connected to, and is my session elevated?”

  • “Load my Contoso session.”

  • “How many guests do we have?”

  • “How old is the data in the Users window? Refresh it if needed.”

  • “Open Users and show me the guest accounts.”

  • “In the Users window, select the accounts whose user principal name ends with contoso.com.”

  • “How many rows are selected in the Groups window? Summarize them by type.”

  • “Which jobs can I run from here? What does the first one do?”

  • “Open Users, select the disabled accounts that still have a license and give me their department.”

  • “Save what you just did as a job called Disabled but licensed.”

  • “Group the users by department and save that as a view called Users by department.”

  • “Build a usage report for the last 30 days.”

  • “Export this grid to C:\Reports\users.xlsx and open it.”

  • “Why did my last job fail?”

  • “What is the difference between a job and a view in sapio365?” (answered from the Ytria documentation)

In the EZ Suite

aclEZ, actionBarEZ, agentEZ, consoleEZ, databaseEZ, designPropEZ, replicationEZ, scanEZ and viewEZ give the AI these instructions. signEZ gives none: the AI can use its tools, but has no signez_get_started tool and receives no instructions.

The EZ Suite has no jobs: the AI works with tools and automation scripts only. For the tools of each product, see The EZ Suite MCP servers.

The shortest path

Each EZ Suite product tells the AI to consider the options in this order:

  1. Built-in tools only: each tool does in one call what its button or menu does. This holds only when every step of the task is a tool.

  2. An automation script file you already have: the AI finds it among the automation files you ran last, and runs it as it is, rather than rewritten.

  3. Only then, one new automation script covering the whole task, written from the documentation. When you may want it again, the AI saves it to a file, so it outlives the conversation.

Rules the AI follows

  • It works as your Notes ID. The product works with the rights of the Notes ID of its session on each server and database, and no tool or automation script can do more. What the ID may not read is left out, and a change Notes refuses is most likely due to the ID's access.

  • It knows what is open. With no Notes session, the AI tells you that nothing works until there is one. With nothing open yet, it lists your favorite and recent databases or servers, and asks which one when your request does not name it. With Full Access Administration on, it says so before changing anything.

  • It checks what the main grid covers. In a product with a database tree, the main grid holds only the databases checked in the tree, as they were when loaded: before answering about a whole server, the AI checks how many databases are checked and how old the list is.

  • It follows each product's own rules.

    • aclEZ, databaseEZ and replicationEZ: changes stay pending until they are applied or saved, and the AI says what will change first.

    • agentEZ: saving signs the agents with the Notes ID of the session, which they then run as; the AI says which agents will change, and that they will be signed, first.

    • actionBarEZ, designPropEZ and viewEZ: a change to a design element inherited from a template is lost at the next design refresh; the AI says so before changing one.

    • consoleEZ: the AI never sends a console command; it gives you the command and the servers to send it to.

    • scanEZ: the AI starts from the tree, which drives what the grids show, and says when a category is not loaded or was loaded over 30 minutes ago.

Prompts to try in the EZ Suite

  • “Which database is open in scanEZ, and what access do I have in it?”

  • “In databaseEZ, select the databases whose title starts with Archive.”

  • “Run the automation script I ran last again.”

  • “Why did my last automation script fail?”