User Guide

1. What AI CaseQuery does

Query results never leave this machine.

You type a question in ordinary words, such as "How many files were opened last month?". The AI writes a database query for that question. AI CaseQuery then runs the query on your firm's own database, read-only, and shows you the result. The AI writes the query, but it never sees the figures that come back. The sentence that describes your answer is written on your own computer from those figures.

AI CaseQuery can only read. It cannot change, add or delete anything in the database.

Some things do go to the AI: your question, and the names of the tables and columns it needs. Section 11 lists exactly what is sent, and what never is.

2. The screen at a glance

The sidebar

The status lines

At the bottom of the sidebar, four short lines tell you the state of things before you ask:

• Database. Which database you are connected to, or Database not connected.
• Read-only. Confirms the connection can only read, for example "SQL Server read-only".
• Speed. How long the next answer is likely to take. Answers in seconds means it will be quick. A line such as "Next answer: about a minute" means the first question after a pause takes longer. Nothing is wrong. Hover over the line to see why.
• Version. The version number, beginning with "v". Quote it when you report a problem.

The badges at the top

3. Asking a question

Writing a question that gets a good answer

Ask the way you would ask a colleague who knows the database but not what you have in mind. The clearer you are about what to count, which ones and over what period, the less the AI has to guess.

Starter questions

Before you ask anything, the chat screen shows a few example questions. Clicking one puts it in the question box. It is not sent. You can change the wording first or send it as it is. Starter questions are chosen in Settings, and they are saved for the Windows account that saved them, so on another computer or another Windows account you may see a different set.

Naming a table with @

If you know which table your question is about, you can say so. Type @ in the question box and a short list of your tables appears; keep typing to narrow it, use the arrow keys or the mouse, and press Enter or click to choose one. Nothing appears until you type @, so the question box stays a question box if you never want this.

Beside each table, under Often needed with it, are the tables that one is joined to — either because your database declares the link, or because queries that have run here have used it. Clicking one adds it to the question as well. That saves the common case of asking about invoices and forgetting that the case number lives on the matter.

Naming a table is a preference, not a restriction. If answering your question genuinely needs another table, it will still be used — and the answer says so, naming what it added. If nothing you typed matched a table, the answer says that too, and tells you what it read instead, so you are never left believing a question was narrowed when it was not.

Only tables this installation lets you see are offered. A table your firm has withheld does not appear in the list.

Sending your question

1. Type your question in the box at the bottom of the chat screen.
2. Press Enter, or click Run. To start a new line without sending, press Shift+Enter. The hint under the box says the same: Enter to send · Shift+Enter for a new line.
3. While it works, a progress line shows the stage it has reached and, after a few seconds, how long it has been running.

The button does not always say Run:

• Estimate & Run means your administrator wants everyone to see the likely cost before a question is sent (see below).
• Ask means the Ask Casey tick box is ticked. Your question goes to Casey, not to the database.

The estimate dialog

If your administrator has switched on Confirm token estimate, a dialog titled Review estimated token use appears before the question is sent. Nothing has reached the AI provider yet. It shows:

• Estimated input: roughly how much your question and its context will use
• Maximum output: the most the reply can use
• Estimated maximum total: the two together

View estimate details shows the breakdown. Click Okay, run request to send, or Cancel to stop. The figure is an approximate estimate made on your computer. "Tokens" are the units the AI provider charges by. A question your firm has already verified costs nothing, so no estimate appears for it.

A question like one already settled

If your wording is close to a question your firm has already verified, AI CaseQuery offers it first: Your DBA has already settled a question like this one. Your question and the settled one appear side by side. Choose Yes, that is my question to get the verified answer, or No, mine is different to have your own wording worked out as usual. You are the only one who knows what you meant, so read both before choosing.

When a question cannot be sent

Your administrator can have AI CaseQuery check each question before it leaves your computer. This is on by default. If your question seems to contain something your organisation does not allow to be sent, such as a Social Security number, a payment card number, an email address, a telephone number or a real value from a sensitive column, you will see This question cannot be sent.

Nothing was sent to the AI provider. The dialog lists what was recognised. Click Ask a new question and your question goes back in the box for you to reword. Ask without the specific value, for example by name, date or status instead of an identifier. The application can usually still answer.

Other reasons a question will not go

• A message says to connect a database, or to enter an API key or choose a model, in Settings. The installation is not fully set up. Tell your administrator. You cannot fix this yourself.
• A message says you have reached the installation's daily limit. Your installation limits how many questions can be asked per day. The count resets the next day, and an administrator can raise the limit.

4. Reading an answer

An answer is built from several parts. Which parts you see depends on the question, on your Explanation level (section 9) and on what your administrator has switched on. Anything warning that an answer may be wrong is shown at every level, except the line saying which business rules were not used, which shows from Reasoning up.

The answer in words

A short sentence or two saying what was found. When the result has figures, a plain description of them follows. That description is written on your computer from the results, not by the AI.

The Query result table

The rows that came back, under a heading that reads "Query result table — read locally, never sent", followed by how many rows are displayed. When the table is the answer, such as a breakdown by month, it opens straight away. When a sentence is the answer, the table is folded and you can click to open it. Long tables show 30 rows at a time, with a button to show more.

• View as chart draws the figures as a chart above the table. It is only offered when the result has a shape a chart can show honestly. Click Hide chart to put it away.
• Export CSV saves the rows as a spreadsheet file. Each export is recorded by the installation with your question, the number of rows and the column names. Your administrator can turn exporting off, in which case you will be told so.

Your administrator may turn the result table off for everyone. The answer in words is still shown.

How this was measured

From the Standard level up, a short list shows the choices the AI made in working out your answer. For example, it might say which date it used or which records it left out. Read this list: a figure can be calculated perfectly and still measure something your firm does not mean. Each item has a Not right? link (see section 6).

The Verified query mark

Beside the answer's title you may see Verified query, followed by a name and a date. It means the answer came from a query someone at your firm checked and settled. No AI call was made, and the same checked query is run every time you ask, so the figure changes only when your data does. If nobody has re-checked it for a long time, the mark says how many days it has been.

Rules that were not used

At the Reasoning level and above, a folded line may say something like "3 of 12 business rules were not used for this question". Those rules describe tables your question does not touch, so they could not have applied. They are listed so that if one should have applied, you can spot it and report it. At Plain and Standard the line is not shown.

This answer depends on a definition nobody has settled

This folded line means the answer depends on a word or measure your firm has not yet defined. It often shows how much that matters, such as "58 rows one way, 44 the other". Open it to see the question and the possible readings.

Choosing one of the readings settles the definition for everyone, so it needs the leads' key. If you do not hold it, click Neither — escalate to Lead. That sends the question, with your wording, to the top of the list a lead works through. If you are a lead, see the Lead Guide: Fine-tuning.

When it asks what you meant

Sometimes the AI stops and asks which of two or more readings you meant, instead of guessing. It offers a few choices. Clicking one puts it in the question box without sending it, so you can adjust the wording, or ignore the choices and ask your own way.

Follow-up suggestions

If your administrator has switched on Suggest follow-up questions, two or three related questions appear under each answer. Clicking one puts it in the box, ready to edit or send.

The reasoning and the SQL

At the Reasoning level, a folded section called How it read your question shows the AI's reasoning and assumptions. At Expert — SQL it is called Generated SQL and reasoning and also contains the query itself. Most people never need these.

5. When to look twice

AI CaseQuery checks how each answer was worked out and tells you when something deserves a second look. These notices do not say the figure is wrong. They say what could make it wrong, so you can decide.

Worth checking

An amber folded panel under the answer, with a heading such as "Worth checking: …". Quieter notices read A note about this answer. Open the panel to see what was noticed. Common examples:

• The same question has been read in different ways on different occasions. The panel lists the readings, marks the one used for the answer above, and shows how often each has been used.
• A total may have counted some records more than once.

"Asked twice — and read two different ways"

Some answers are checked by planning the question twice, independently. You will mostly see this on questions opened from Fine-tuning. If the two attempts disagreed, a panel headed Asked twice — and read two different ways says how they differed. The answer shown is the first of the two. If the attempts agreed, a single quiet line says so under the answer.

Help me sort this out…

Most Worth checking items have this button. It opens a short guided dialog of four steps. Nothing is decided until the third step.

1. What happened: what was noticed, and the readings involved.
2. What it means: why it matters, and what usually settles it. If Casey is available and you still do not follow, click I still do not follow — explain it to me. Casey is sent your question and the readings, never any rows.
3. Your call: choose one of three:
   • Yes — I can say what we mean. You write the definition and it applies from the next question on. This needs the leads' key, and you will be asked for it if you have not entered it.
   • No — pass it to whoever settles these. Opens the correction box described in section 6, with the readings you were looking at already included.
   • Nothing for now. The answer stands, and the question will be raised again next time it comes up.
4. Write it down: only if you chose Yes. You write the definition and check it before it is recorded in your name.

6. When an answer looks wrong

If one of the items under How this was measured is not what your firm means, correct it where you see it.

1. Click Not right? beside the item. The Correct this measurement box opens and quotes the item.
2. Under What went wrong?, pick any that apply, for example It answered a different question, The number or records are wrong or Something important is missing. This is optional.
3. In What should it say?, write what the measurement should have been, in your own words. This is required, and it is the part that gets read. Example: "Work logs with no end date are still in progress, not backlog."
4. Enter Your name so whoever picks it up knows who to ask. The browser remembers it for next time.
5. Click Send for review. The link under the answer changes to Sent for review.

Where it goes. Your correction goes to the top of Fine-tuning › Definitions to settle, where a lead can settle it. A copy is also written to a file for whoever supports your installation. That copy holds your question, the measurement, what you picked and wrote, the query that produced the answer, your name, the date, and your computer's name with the version and database it was using. The correction is not sent to the AI, and it does not change the answer already on your screen.

Draft a rule… turns a correction into a standing instruction. That is an administrator's action. If you are not an administrator, clicking it asks for the Administrator key, so leave it alone.

7. When you are asked what a word means

Words like "active", "open" or "outstanding" can mean different things to different people. The people doing the work usually know what the firm means, so AI CaseQuery can ask them.

When the question appears

A short question appears under the answer, above the table, when both of these are true:

• your question uses a word such as active, inactive, open, closed, current, currently, still, outstanding, pending, unresolved, incomplete, ongoing, retired or live; and
• to answer it, the AI counted exactly one value in a status-type column that holds between 2 and 6 different values.

It is not asked about answers from a Verified query. It is also not asked when the answer used filters such as "anything except", "one of" or "like", when it filtered on numbers or yes/no flags, or when it filtered on more than one status column.

Answering it

The question reads, for example: Counted "open" as Open. Is that right? There is a button for each value in that column, with the one that was used marked "— yes", and a Not sure button.

• Click the value your firm means. If the one used is right, click that one.
• If you do not know, click Not sure. It is a real answer. A guess looks exactly like agreement, and "not sure" never counts towards settling a definition.

After you answer, the question is replaced by a line saying what happened. For example: "Recorded. 3 more people answering the same way would settle it." Your vote does not change the answer on your screen.

Who sees your vote

A line under the question tells you before you answer:

• If your firm signs in with Microsoft or Windows accounts, your name and your answer are visible to whoever settles definitions.
• If your installation uses a shared access key, your answer is recorded without a name, and votes without a name cannot settle a definition on their own.

What happens with the votes

• When enough different, identified people give the same answer (four, unless your administrator changed it), that reading becomes a settled definition straight away and applies to everyone's answers. It stays provisional until an administrator confirms it.
• If people answer differently, or enough people have answered without agreeing, asking stops and a lead decides.

If you would rather not be asked

If your administrator has left this to each person, the question includes a Stop asking me link. Click it and you will not be asked again on this browser. The vote in front of you can still be cast. To be asked again, tick Ask me what words mean beside the question box (see section 9). If your administrator has switched on Everyone is asked, the link does not appear.

The "Counted … as …" line

Sometimes you will see only a line such as: Counted "active" as Active. There is nothing to click. It tells you which reading your answer used. You see it when you have switched asking off, when the column holds 7 or more values, or when your administrator has switched on Show which reading each answer used without voting. If the reading is not what your firm means, report it with Not right? as in section 6.

8. More than one database

Tick All databases beside the question box to ask one question of every connected database at once. The answer is labelled All databases. Where it is safe to add them up, it gives a combined total. It always shows where the figures came from, database by database.

• If a combined total could count the same record twice, it shows the figure for each database and not a total.
• If a database stopped at the row limit, its figure reads "at least …".
• If the question cannot be answered across the databases, the reply explains why.

Answers across databases are simpler than ordinary answers. There is no estimate dialog and no Worth checking panel, and your Columns and Ask me what words mean choices are hidden while the box is ticked. Untick it to go back to asking your usual database.

9. Your own choices

Next to the question box, marked Yours only, are choices that affect only your screen. They are remembered in this browser. Each one appears only if your administrator has left it open to you.

Your chat history

Conversations are saved automatically in this browser only. They are not kept on a server, and they will not appear in a different browser or on another computer.

• Click Chat history in the sidebar to see Past chats. Click one to reopen it.
• Click New conversation at the top of the list, or New chat in the sidebar, to start again.
• Click the × beside a chat to delete that one.
• Delete all removes every saved conversation from this browser. You are asked to confirm first. Deleted history cannot be recovered.

Large result tables are saved with only a shortened preview of their rows.

What your administrator decides

These are set in Settings, which needs the administrator key, so you cannot change them yourself. They are saved for the Windows account that saved them, on that computer (see the note below), so on a shared computer or a copied installation your administrator sets them where you work.

• The starting Explanation level, and whether everyone is held to it
• Whether answers show the result table (Show result table)
• Whether follow-up questions are suggested (Suggest follow-up questions)
• Result columns: supporting columns, only the requested columns, or your choice
• Whether you see an estimate before each question (Confirm token estimate)
• Whether questions are checked for private details before sending (Check questions before sending)
• Safe response mode and the most rows an answer may return
• Whether people are asked what words mean, and whether they can opt out
• Whether Casey is offered on the chat screen

10. Asking Casey

Casey is "Your context-aware CaseQuery assistant". Casey explains how AI CaseQuery works: what a warning means, what a rule is, how to word a question. Casey does not answer questions about your data, never runs a query and changes nothing. To ask Casey, click Ask Casey in the sidebar, press Ctrl+K, or tick the Ask Casey box beside the question box, where the button then reads Ask. The sidebar button and Ctrl+K open the Casey panel, which keeps one conversation as you move between screens; a question asked with the tick box is answered in the chat instead, and does not join the panel's conversation. Casey never sees your query results. Casey is only available if your administrator offers it, and questions to Casey use your firm's AI provider account. For everything else, including what Casey can read and its limits, see Working with Casey.

11. What is sent to the AI, and what never is

Query results never leave this machine.

What is sent to your firm's own AI provider

• Your question
• The names of the tables and columns needed to answer it
• Your firm's notes, terms and rules
• Within one conversation, your earlier questions and the queries and reasoning the AI itself wrote for them, so a follow-up makes sense. A query can contain a value you asked about. The answers shown to you are not sent.

Three exceptions, named

1. Short lists of codes from the preparation scan travel as column notes. For example, a note may list the codes a status column uses, so the AI knows what they are.
2. When you ask the assistant about a tuning item, it is sent that item's counts and values from the preparation scan.
3. When a query fails and is retried, the query that failed is sent back so it can be corrected — and a query can contain a value you typed. What is not sent is the database's own error message: the AI is told only what kind of error it was.

What is never sent

• The rows and figures your query returns
• Columns your administrator has hidden
• The answer text, which is composed on your own computer
• Your saved chat history as such: it is kept only in your browser. (Within a conversation, earlier questions and the AI's own queries and reasoning are sent with a follow-up, as listed above; answers never are.)
• Corrections you send for review, which go to your firm, not to the AI

12. Common questions

Can AI CaseQuery change anything in our database?

No. It connects read-only, and anything other than reading is always blocked.

Why did I get a different figure when I asked the same question again?

Each answer is worked out afresh from your wording, and a question that can be read more than one way may be read differently next time. Check How this was measured and any Worth checking panel. Answers marked Verified query are the exception: they run the same checked query every time.

Why is the first answer slow?

After a pause, the first question can take about a minute, and later ones come back much faster. The speed line in the sidebar tells you which to expect.

I can't find the Detail choice.

Your administrator has set one level for everyone, so there is nothing for you to change.

Why is there no table under my answer?

Either your administrator has turned the result table off, or the question returned no rows. If the table is there but folded, click its heading to open it.

Can I see the query it ran?

Yes, if your Explanation choice is open to you. Set it to Expert — SQL.

Does sending a correction change my answer?

No. It goes to your firm's settling list and to a file for whoever supports the installation. Once a lead settles it, future answers change.

I cleared my browser and my chat history has gone.

History is kept only in the browser. It cannot be recovered from the server, because it was never stored there.

Why was I asked for a key?

You clicked something that needs a lead or an administrator, such as settling a definition or drafting a rule. If you do not hold that key, cancel, and use the option that passes the question on instead.

Does it cost anything to ask?

Questions use your firm's own AI provider account, and your installation may set a daily limit. Answers from a Verified query cost nothing.

If something looks wrong

Whenever you report a problem, include the version number shown at the bottom of the sidebar.

Glossary

Administrator

The person who sets up and runs the installation. Only they can open Settings, save business rules, and confirm or withdraw settled definitions.

AI provider

The company whose AI model writes the queries. Your firm uses its own account, and questions are charged to it.

Business rule

A standing instruction the AI applies to every question it concerns, written by an administrator. You can read them under Business Rules.

Casey

The assistant for questions about how the product works. Casey never sees query results and never runs a query.

Column note

A short description of a column, sent to the AI so it understands the column. It can include a short list of the codes the column uses.

Explanation level

How much of each answer you see: Plain, Standard, Reasoning or Expert.

Requested only

An answer that returns only the columns asked for, with no supporting columns, sorting or rounding.

Hidden column

A column an administrator has marked as never to be sent to the AI.

Lead

An Operations Lead: someone at your firm who settles definitions and checks baseline answers.

On Floor Voting

An optional feature that asks the people using the product what a word such as "open" or "active" means, and settles it when enough of them agree.

Preparation scan

A reading of the database that an administrator runs before anyone asks questions, so the AI is given useful descriptions of it.

Provisional

A settled definition that is already in use but has not yet been confirmed by an administrator.

Query

The instruction to the database that fetches the answer. It is written by the AI and run read-only against your firm's database.

Reading

One way of understanding a question or a word, for example counting "open" files by status or by close date.

Settled definition

A decision, recorded in someone's name, about what a word or measure means at your firm. It applies to everyone's answers.

Token

The unit AI providers charge by. More text sent or received means more tokens.

Verified query

A question whose query someone at your firm has checked and settled. It is answered without an AI call by running the same checked query every time.

Worth checking

An amber notice under an answer pointing out something about how it was worked out that deserves a second look.