Your question goes to the AI. No record from your database ever does.
Not connected—Safe mode on
✦
Ask a question about your data
Every query is generated from your own schema, runs read-only against your data, and stays inside your network.
1 · Definitions to settle
Each one steadies several questions at once. Some your data can evidence; the rest only your firm knows.
—
2 · Baseline questions
Run questions you already know the answer to, and confirm the definitions landed the way you meant. What you confirm here becomes the baseline you measure against later.
0 reviewed
The sections follow the life of a matter — from the day it arrives, through the work, to what happened afterwards. Work down it, or open the section that matches whoever is sitting with you. Leave anything you cannot answer unmarked; nothing is lost.
The three kinds of question, and how each is checked
Ones you already have a report for. Counts and totals by client, by state, by stage. Check these against the report you already run — if the two disagree, one of them is wrong and it is worth knowing which.
Ones that join up what no single report holds. A matter due to move forward while another part of the same file says it should be paused. These take information from several places at once, they are the reason for this product, and they are worth your best reviewer’s time.
Ones nobody can answer from memory. Files whose own records contradict each other — two records both marked current, or a step recorded as finished before the step it depends on. Do not look for a number you already hold. Open two or three of the files it returns and confirm the problem is real.
Come back to these on a schedule
Confirming an answer records what was true on the day, and nothing here will tell you when that has gone out of date — every rule and definition settled afterwards changes what these questions return. Most firms work through the list during setup and re-check it quarterly, or twice a year.
3 · Awaiting sign-off
—
These are already in force.
A definition applies from the moment a lead settles it. Confirming records that the firm
stands behind it; withdrawing stops it being applied. One left waiting is marked overdue
after 30 days and again after 90, and goes on applying throughout.
Confirming or withdrawing is the administrator's.
It is a second signature: the name is asked for every time and never filled in from
whoever last typed one, because the point is that somebody other than the lead who settled
it is signing off.
It is a second signature, recorded under the administrator signed in,
because the point is that the firm, not only the lead who settled it, stands behind it.
How this works
What the definition scan looks for, and why it uses no AI.
Reference
Why this exists
When a question has more than one honest reading, the system has to pick one. Whichever it picks, somebody eventually notices the number is not the one they expected — and once a person doubts one number, they doubt all of them. The cost is not the wrong answer. The cost is that the right answers stop being believed.
Why it does not use AI
Deliberately. Asked one question about service-level breaches, two different AI providers agreed on the answer and both ignored a stored deadline column that disagreed with the contract terms on 354 of 360 rows. A shared blind spot. A plain arithmetic check found it in one line, and it does not share the models' assumptions — which also means it runs with no provider key installed and no data leaving the building.
What it looks for
Two definitions of open. A status column and a completion date that each decide whether something is finished.
Competing amounts. Several columns answering to the same word — invoiced, approved and written off are three different truths about one invoice.
Competing deadlines. Two dates that both claim to be the deadline and routinely disagree.
History dated in the future. Rows recording something that has already happened, carrying a date that has not arrived.
A concept defined in several places. A service-level target held in four tables, where more than one could govern the same step.
What it cannot see
Several of the definitions above are not in your data at any depth. Whether a step with no completion date is outstanding or simply not required in this case; whether work performed by an outside party counts against your own timeliness; whether days spent waiting for a client's authorisation are yours. No amount of reading a database reveals those, because the convention lives outside it. They are asked rather than found.
What it deliberately does not do
It stays silent when both readings give the same number. If there is nothing to decide, there is nothing to report.
It does not call your data wrong. These are decisions to confirm, not defects.
It cannot see past your safeguards. A column classified as sensitive is refused here exactly as it is refused in chat.
Rolling this out, in order
Several settings are stored in the data folder so that every workstation reads the same value. Lock down the program folder and the settings files last, or you will have to unlock them again to finish.
Each step builds on the one before. Pair your firm's connector first: until it is, nothing else here can be checked.
Request safeguards
These control how much data each question may return and how much reaches the AI provider. They are saved for the signed-in Windows user on this machine, so they persist across restarts and are not affected by which browser is used.
These control how much data each question may return and how much reaches the AI provider. They are your firm's, the same for everyone at it, whichever computer or browser they use.
Safe mode allows up to 1,000 rows.
Checking questions before they are sent
Warns the user when a question appears to contain a Social Security number, a payment card, an email address, a telephone number, or a real value from a column this application treats as sensitive. Nothing is sent until they decide. It needs no configuration and reduces mistakes rather than preventing them.
Time limits
How long the application will wait before giving up and telling the user. Both are in seconds, between 5 and 600.
Applies to the whole question. A question that needs a second request after a failed query shares this budget rather than doubling it.
Applies to each query run against your database. SQL Server, PostgreSQL, MySQL and Oracle cancel the statement on the server. SQLite is stopped while it reads rows, but a single long aggregate cannot be interrupted — see the security notes.
On Floor Voting
When a question can be read two ways, ask the people who know.
To see how it is going — how often people are asked, how often they answer, and how often they are not sure — open Business Rules › Has it worked?
How a vote becomes a settled definition
Everyone agrees, and enough have voted — it becomes a settled
definition automatically, marked provisional and applied straight away; an
administrator confirms or withdraws it under Fine-tuning › What's been settled.
Anybody disagrees — voting stops there and then. It goes to the
lead as contested, showing both camps and who is in them. More votes on a question
people already read differently are noise, not evidence.
Nobody can agree by the time asking stops — the lead decides. The
product never keeps pestering.
Where a lead rules — once people answer differently, or asking stops
without agreement, the question appears under Fine-tuning › Definitions to settle ›
"Floor needs a ruling", with who voted which way, and a lead settles it there.
"Not sure" is a real answer
Without it people guess, and guesses look exactly like agreement. A "not sure"
counts towards stopping the asking, never towards settling the rule. If most people
say it, the fault is the question we asked rather than anything the firm has failed
to decide, and the lead is shown our wording to fix.
Answer display
What each answer shows in the Chat screen.
Around each answer
Supporting columns can make an answer easier to understand. Only the requested
columns adds no supporting figures or joined names, and does not add sorting or
rounding. Choose it for exports or line-by-line reconciliation. Each person
chooses starts with supporting columns and remembers each person's choice.
Hard protections remain active
Regardless of the settings above, writes, multiple statements, unsafe engine functions, and system-catalog access are always blocked, by this service and again by your firm's connector; an answer is at most 500 rows; and columns classified as sensitive in Schema rules never leave your firm's network. For databases too large to send in full, switching Safe response mode off may name more unclassified columns to the AI — classify them in Schema rules before switching.
Regardless of the settings above, writes, multiple statements, unsafe engine functions, and system-catalog access are always blocked, the technical ceiling is 5,000 rows, and columns classified as sensitive in Schema rules are never sent to the AI provider. For databases too large to send in full, switching Safe response mode off may name more unclassified columns to the AI — classify them in Schema rules before switching.
Write definitions, not answers
Use these for the things you should not have to explain twice — what a term means, how to count, which records to exclude. Good rules define your terminology and conventions; avoid recording conclusions, because telling the AI what it should find stops it finding anything you did not already know. Up to a dozen are sent with every question; beyond that, a rule naming a table travels with questions about that table, and a rule naming none is always sent.
Everything on this screen is either an instruction the AI follows, or evidence about whether those instructions are working.
Every instruction carries a label saying where it came from, so what was in
force when an answer was given can always be read back.
If a question is still being read more than one way
That is the one thing on this screen worth acting on, and the action is not here. Go to Fine-tuning → 1 · Definitions to settle and answer the definition behind it, or ask the question in Chat and use the amber panel on the answer. Settling needs the lead key.
Term — "when someone says X, they mean Y." A vocabulary entry.
Answers what does this phrase mean.
Rule — "always do Z." A standing instruction. Answers
how should you behave.
Status
Add a rule (Administrator)
One instruction, in plain words. Rules travel with the questions
they concern, so keep each one short and about a single thing.
Rules in effect by Administrator
Policy on how to count, written by an administrator. Edited here.
Edit all as text
Useful for pasting several at once or reordering them. Editing
here updates the list above as you type.
One rule per line. Up to 200 rules, 400 characters each, and about 30,000 in total. Each one travels with the questions it concerns, so keep them short and specific — anything too long is refused and named rather than shortened for you.
Matching part of a value — wildcards in a rule
A rule can ask for a partial match rather than an exact one. Two characters work
on every database this connects to:
%
any run of characters, including none
'On Hold%' matches On Hold - Bankruptcy and On Hold - Workout
_
exactly one character
'FL-____' matches a Florida code with four characters after the dash
Useful rules look like this:
When a name is searched, match it anywhere in the name rather than from the
start, so a search for Smith also finds Smithson and Bank of Smith.
Any status beginning 'On Hold' is held, not active.
Reference codes starting 'INT-' are internal and stay out of totals.
You do not have to use them.
"Match the name anywhere, not just the beginning" says the same thing, and the AI
writes whatever your database needs. That is safer, because whether a match
ignores capitals differs by database — so say it in the rule if capitals matter.
A value that genuinely contains a % or _ needs
mentioning, or it is read as a wildcard.
Saving puts your written rules in force from that moment, for everyone using this installation. The definitions from Fine-tuning and your terms are not affected. The previous set is kept under History.
Definitions in effect from Fine-tuning
Term — "when someone says X, they mean Y." A vocabulary entry.
Answers what does this phrase mean.
Rule — "always do Z." A standing instruction. Answers
how should you behave.
Define a term (Administrator)
Up to a dozen rules and terms together are sent with every question. Past that, a
term travels with questions that mention it or name its tables, so writing one down costs
nothing on the rest.
Comma separated, and worth the minute it takes: somebody asking
for "active inventory" gets this term even though those words appear nowhere in the
definition. That is the gap synonyms close. Not sent to the AI.
Optional — better with help from IT
The term works without these. If somebody who knows the
database can fill them in, the AI stops having to interpret your wording and the
term is only consulted where it belongs.
The condition itself, not a sentence about it. This is the
part that stops the AI translating your words and getting it wrong.
Comma separated. Naming them keeps the term out of questions
it has nothing to do with; leaving it blank sends it with every question.
Terms you have written
What no longer steers answers. Nothing here is sent to the AI.
Earlier versions of the rules
Saved automatically before each change to the rules.
Observed from questions people have actually asked. Advisory only — nothing here changes an answer, and clearing an entry changes nothing but this list.
How floor voting is going
Database objects
Loading…
⌘
Select a table or view
Inspect columns, relationships, source SQL, and sample rows.
VIEW
What to hide, and what not to
Tick a column only if nobody should see its values — a Social Security number, a bank account or card number, a password. A ticked column is never sent to the AI, never displayed to anyone, and any query naming it is refused.
Do not tick a column just because it looks personal. Party names, addresses and reference numbers are how your staff and the AI identify records. Hiding borrower_name means the AI can no longer find a case by who it is for, and ordinary questions stop working.
No database rows or query results are sent to the AI. This decides what appears on screen, and whether the column's name is disclosed. A value written into a rule or a column note travels as part of that instruction, including the code lists preparation writes for coded columns.
Hidden columns
Tick only columns nobody should see, such as an SSN or a bank account number.
Column notes
Explain abbreviated or coded columns so the AI reads them correctly.
File details
What opens when somebody clicks a case number in a result.
Tick the fields to show, in the order you want them read — ticking
adds a field to the end, so re-ticking moves it. Results listing the identifying column
become clickable. Columns you have hidden never appear here, whatever is ticked.
AI-assisted view mapping
Select sources, review related-schema recommendations, choose grouped columns, and generate a full-width script below in your database's dialect.
Worth building, from what has been asked
0 selected
Recommended related schemaSelect a source object to analyze relationships.
Choose available columns0 selected
Columns are grouped by the selected table or view. Edit the output alias after selecting a column.
Recommendations are local and do not consume any AI tokens. The generated script is never executed; review it before applying it.
Recommendations are worked out by this service from your database's structure and do not consume any AI tokens. The generated script is never executed, here or by your firm's connector; review it before applying it to your database yourself.
◇
No view design yet
The AI provider will map the selected schema into a reviewable view script in your database's dialect.
PROPOSED VIEW
Row grain
Sources
Actual tokens—
Column mapping
CREATE VIEW script
Preview SELECT
Assumptions
Implementation notes
Request trace
Records what each question sent to the AI provider — so you can show an auditor what left this machine rather than describe it. The assistant, documentation reading, rule previews and question-SQL pair trials are not recorded yet; the log says so.
Records what each of your questions sent to the AI provider — so you can show an auditor what left your firm rather than describe it. The log is kept in this browser only, for you; the service keeps no copy. Documentation reading, wording help and rule previews are not recorded yet; the log says so.
Query results are never sent, so the log never holds them.
It holds your schema, column notes and questions, and a database's own error text, which can
quote a value — handle it as confidential. Tracing is off by default, turns itself off when
the application restarts, and API keys are never recorded.
Send us a support file
If something is not working, this gathers what we need to diagnose it into one file you
can read before you send it. Nothing is transmitted from here — you decide what happens
to the file.
It contains the shape of your set-up -- your firm's name, your databases'
names and your connector, how many tables, rules and notes there are, which AI provider is in
use and where its requests go (for Microsoft Foundry, your Azure resource) -- and the last
entries of the activity record with people's names replaced by user-1,
user-2. It never contains a key, a password, a database address or login, a
question or a row of your data, and this service keeps no copy. Open it and check before
sending.
It contains the questions people asked and the shape of your schema, an
extract of the activity record with names replaced by user-1, user-2,
and the application's own log. It never contains an access key, a provider key, a
database password, a licence key, or a row of your data. Open it and check before sending.
Licence
Apply a licence key
Paste the key you were sent and press Apply. It is saved for this installation — on a shared installation, every workstation picks it up.
Choose this installation's own keys
The administrator and leads' keys that came with the evaluation are
the same in every copy of the product, and they stop working the moment this licence is
applied. Both must be replaced now — this is the last point at which the evaluation keys
still let you set them.
Opens Settings, Safeguards, the Schema Explorer and the View Designer.
Typed once per browser session.
Opens the fine-tuning tier — settled definitions and business rules —
without opening the security policy or the provider credentials.
Write both down before you apply. They are stored as
verifiers rather than as the keys themselves, so neither can be read back out of this
installation afterwards.
This key was issued as a second key, to sit alongside a licence rather
than replace one. Apply it as a second key, or ask us for a replacement licence that includes
everything in one.
People at your firm
Everyone who signs in to AI CaseQuery for your firm, and what each may do. A user asks
questions; a lead also settles definitions and reviews answers; an administrator also
manages Settings and these sign-ins. Your firm always keeps at least one administrator.
You cannot change or remove your own sign-in here. Another administrator can.
Sign-in name
Role
Added
Password for
Give this to them privately, apart from their sign-in name. It is shown
this once and is not kept anywhere you can see it again; if it is lost, reset it.
Work-account link for
Send this link to them. Opening it, they sign in with their Microsoft work
account, and from then on they can sign in with it. It works once, for seven days, and only with
an account in your firm's own Microsoft directory. It is shown this once.
Add someone
Where people may sign in from
List your office's internet addresses, and your VPN's, and nobody can sign in to your firm from anywhere
else, even with the right password. Anyone already signed in is signed out the moment they use it from
elsewhere. Leave it empty to allow sign-in from anywhere.
One address or range per line, for example 203.0.113.10 or 203.0.113.0/24. Your
connector is not affected. If your office's address changes and nobody can sign in, contact us and we
will clear the list.
Who may administer this installation
Reading…
Change who holds Administrator and Operations Lead
Choose how your firm grants the two tiers, fill it in, and this
composes the settings for you. Anything already in the file — a published connection,
a provider key — is kept.
Where user corrections are collected
When a user disagrees with one of the measurements under an answer, their correction is appended to a text file. Name a folder every user can write to — typically a network folder beside the application share — so every installation's corrections and journal collect in one place.
Set this before the last setup step. It is saved in data\deployment.json so that every workstation reads the same value, and that file is locked down last.
Leave empty to keep corrections on each workstation. The folder is tested when you save.
Installation settings
The settings a new PC or Windows user starts from, until they choose their own. Without
them, everyone who has not chosen starts on the settings the application ships with, not
the ones an administrator set on their own PC.
What they cover: the AI provider, the model for each
provider, the thinking level, fast mode, and nine safeguards: row limit, safe mode, question
checking, explanation level, result columns, result table, the AI and query time limits,
and On Floor Voting.
When they matter: whenever more than one person uses
this installation. On a shared network installation every PC reads them. On one PC or a
terminal server, every Windows user reads them. When the folder is copied to each PC, each
copy keeps the settings it had when it was copied, so set them before copying. Anyone who
has chosen their own setting keeps it; other running copies pick up a change when they
restart.
Where your AI provider key is stored
This workstation
Do this on every workstation that should be able to ask questions.
It affects this computer only.
The whole installation
Once, for everyone. Do it only after every
workstation has had the key moved onto it — a computer that has not loses its access
the moment this runs.
Where
What protects it
This workstation %PROGRAMDATA%\CaseQuery\Security
Encrypted by Windows for this computer. A copy taken to any other computer
cannot be read at all, so copying the application folder carries no working key. It is
bound to the machine and not to a person: anyone who may run CaseQuery here can use it.
An environment variable CASEQUERY_ANTHROPIC_KEY
Nothing is written to the application folder, and it does not travel with a copy of
it — but the value is held in plain text and any user of this computer can read it.
A compatibility option, not an equivalent one.
deployment.json beside the application
Not used. The application folder is designed to be copied, so a
credential in it would travel with every copy. A key found here is refused, and the
button above moves it onto this computer instead.
Putting it in the environment instead
Run this once in an elevated PowerShell, then restart AI CaseQuery.
setx /M CASEQUERY_ANTHROPIC_KEY "your-key-here"
/M sets it for the whole machine. Leave it off to set it for yourself only.
For OpenAI the name is CASEQUERY_OPENAI_KEY.
Push it with Group Policy or Intune across many workstations, and nothing sits in the folder on any of them.
The command above puts the key in your shell history. Use System Properties → Environment Variables if that matters to you.
An environment variable is not a secret from the workstation.
Any process running as that user can read it. It takes the key out of the application
folder; it does not hide it from the machine.
Limiting what a leaked key could cost you
Issue the key in its own workspace with a spend cap
Anthropic organises accounts as Organisation → Workspaces → API keys, and a workspace
can carry its own monthly spend limit. If the key ever leaks, revoke it in the Console —
it stops working immediately, your other keys are unaffected, and it could never have
spent more than that workspace's cap.
IP allowlisting is available on Anthropic Enterprise plans.
Anthropic checks the source address of every authenticated request against CIDR ranges you
supply — your offices and VPN exit points. It is set for the organisation rather than per
key, and arranged through your Anthropic contact rather than in the Console. There is no
domain restriction, and none would apply: this is a server-to-server call from the
workstation, so there is no referring domain to check.
Take a copy of your tuning work
Downloads one file containing everything your firm has taught this application: your business rules and terms, the definitions you have settled, your column notes, the questions you have checked, your reference figures, your preparation and your settings. Keep it wherever you keep backups. If your firm's set-up is ever damaged or changed by mistake, that work does not have to be done again.
The file contains no keys and no rows from your database
No AI provider key or other secret, no connector registration, and no rows or values: the service keeps none read from your database. It does hold what your set-up says about your database — table and column names, your notes on them, which columns are withheld, and what preparation measured of its shape. The activity record stays with the service. Your people and their roles are listed in it for reference, and are never restored from it. The file is saved to this computer: keep it as carefully as you would your rules themselves.
Restore from a copy
Puts a copy taken here back: business rules, settled definitions, column notes,
the sensitive-column policy, preparation and settings. It replaces what is here rather
than merging — anything the copy does not have is cleared. Every item it changes keeps
its previous version, so a restore can itself be undone. Only your own firm's copies are taken.
Take a copy of your tuning work
Writes one file containing everything your firm has taught this application: your business rules and terms, the definitions you have settled, your column notes, the questions you have checked, and every previous version of your rules. Keep it wherever you keep backups. If this installation is ever rebuilt, that work does not have to be done again.
The file contains no keys and no credentials
No provider keys, no access keys, no licence key and no database connection — those belong to this installation, not to your work. Chat history and any API key saved in a browser stay in that browser and cannot be read from here. The file can be handed to whoever keeps your backups.
There is no default and the folder is not remembered — type it each time, so the copy always lands somewhere you chose.
Recorded inside the file and in the activity record, so a copy found later can be traced to whoever took it.
Daily copies, kept for you
A copy of your rules, terms, settled definitions and column notes is saved
automatically each day. It is updated a few seconds after any change is saved, again when the
application closes, and checked every ten minutes, so each day's copy is how that day ended. The last seven days are kept in the daily-exports folder beside your
data. To go back to one, press Use this one; you will see what it would change
before anything is restored.
Restore from an export
Puts a previous export back: business rules, settled definitions,
column notes, the sensitive-column policy and the preparation record. Use it on a
rebuilt server, or when moving from a pilot machine to the real installation.
It replaces what is here rather than merging — the current version of
every file it touches is copied aside first.
Connect your databases
The recommended way. Install our connector on a Windows server that can reach your
databases. It connects out to them -- on your own network, or in your own cloud account, as many as you
need -- and to this service. Your database logins stay on that server: this service never sees them.
How to connect
Get the connector: download it below where this page offers it, or NetDirector sends it to you. Unzip it anywhere on that server.
Double-click Install and say Yes when Windows asks. A form opens.
Create a pairing code with the button below (it works once, for 30 minutes), paste it into the form, enter your first database and a login that can only read, and press Connect.
To add another database, run connector add-database on that server, then restart the connector (Task Scheduler › CaseQuery Connector › End, then Run). It then appears under Which database; save its Settings › Schema rules before any of its data is read.
Your firm has one connector, and it can hold up to 16 databases: SQL Server, PostgreSQL, MySQL or MariaDB, and Oracle. Pairing a new connector replaces the current one.
Pairing code
Enter this code at the connector within 30 minutes. It works once, and
is not shown again.
Download the connector
Connector
Status
Paired
By
No network of your own? We can connect to your online databases for you.
If your firm cannot run our connector, NetDirector can run one for you that reaches your
PostgreSQL databases in a cloud service -- one, or several. It is the limited option: it cannot reach anything on your own
network, and each database must be open to us.
Your database must accept encrypted connections, with a certificate that can be checked.
Create a login that can only read, for us to use. The connector NetDirector runs for you keeps it, sealed, and is
the only part that opens it. You can end our access at any time by changing that password or removing our addresses.
Allow NetDirector's connector addresses through your database's firewall. NetDirector gives you the list.
Ask NetDirector to set it up. Once it is, add each database below.
Which database
Your firm's connector can offer more than one database. The one you choose is where your
questions are answered, and what the schema, preparation and fine-tuning screens show. It is
your own choice: nobody else's changes with it.
Database
Engine
Add a cloud database
Give the connector NetDirector runs for you the database to reach. It tests the connection first and adds
nothing that does not open. PostgreSQL, over an encrypted connection whose certificate is
checked.
The login is sealed in this browser so that only your firm's connector can
open it. This service passes it on and keeps no copy; the connector keeps it, as it keeps the
login for every database it reaches. Use a login that can only read.
Step 2 · Review what may be read
Once your database appears above: Settings › Schema rules, choose it, tick any column that must
never be sent, and Save -- even if nothing needs hiding. Nothing of its data is read until you do.
Then ask your first question in Chat.
Try it with your own database
This demo runs on our sample database. To try it on yours, start a trial: this demo is
cleared, and you set up your own database from the beginning.
Database connection
Every supported database is opened read-only. The driver for each type ships with this application, so nothing needs installing.
What rights does the login need?
The commonest first-day problem, and the least obvious: a login
with too few rights connects successfully and reports a smaller database. Every screen
fills, every answer arrives, and every answer covers only what was granted. On
connecting, AI CaseQuery compares what this login can see against what the database
holds and tells you if a material part is missing.
PostgreSQL — run for every schema your
data lives in, not only public:
GRANT USAGE ON SCHEMA public TO your_login;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO your_login;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO your_login;
The third line matters as much as the first two. Without it a
table created next month is invisible to a login that can read every table created
before it, and nobody connects the two events.
SQL Server / Azure SQL
ALTER ROLE db_datareader ADD MEMBER your_login;
GRANT VIEW DEFINITION TO your_login;
VIEW DEFINITION is what lets views report their
columns. Without it they appear empty, which reads as a corrupt schema rather than a
missing right.
MySQL / MariaDB
GRANT SELECT ON your_database.* TO 'your_login'@'%';
Oracle — SELECT on the tables, plus
SELECT_CATALOG_ROLE to read the data dictionary.
PostgreSQL schemas outside public need
nothing extra. Names are written in full — analytics.monthly — so they
resolve wherever search_path points. Setting search_path
anyway is fine, but on some hosted services it disconnects open sessions.
Saved connections
Connection details are stored in app-config.json. Passwords are never written to disk.
QuickBooks
QuickBooks Desktop
A QuickBooks company file cannot be queried directly — there is no SQL behind it. AI CaseQuery reads it through Intuit's own interface and writes a snapshot it can query. QuickBooks must be running with the company file open, and the first attempt raises an approval prompt inside QuickBooks that often opens behind other windows.
QuickBooks Online
Two steps. Sign in once per client to authorise their QuickBooks Online company — a browser opens and Intuit asks which company to allow. Check the company name on that screen, since Intuit offers the last one used by default. Then import whenever you want current figures; nothing is read until you do.
Make this connection live for everyone
Final step
Whatever is connected now — a database, or a QuickBooks import — is available only on this machine until it is published. Publishing hands the same connection to every workstation using this installation, so nobody else configures anything.
Do this before the last setup step. Once data\deployment.json is locked down, only an account allowed to change it can publish.
More than one database
Answers are combined across every database added here, alongside the one this
installation is connected to.
What each database calls things
A question is asked in one vocabulary and answered by systems that each name things
differently. This is where you say which column holds the firm's case number, the status,
and the rest — once per database, for each kind of record it holds. Nothing is assumed: a
field bound to the wrong column gives answers that look ordinary and are about something
else.
Client money review
An accounting database is not read until the firm confirms which objects hold client
trust money or cash movements. This asks about that and nothing else: unticked objects
become readable, subject to the separate sensitive-column rules.
1 · Prepare the database
Read it once. Describes the coded columns so the assistant is not guessing at them, and finds the definitions your firm needs to settle.
Not run
Your data stays here
Preparation runs locally against your database. Table data is not sent to AI.
Preparation is done by your firm's connector, on your own network, against your database. Table data is not sent to AI, and what it reads stays with the connector: only the column notes you apply and the relationships it proved come back.
CaseQuery looks at things such as data types, common codes, dates, amounts and value patterns. The resulting profile holds structure, counts and characteristics — not your underlying records.
Computed with SQL against your database — no AI tokens, and nothing to pay your provider. Run it as often as you like.
Documentation improves understanding
Optional — preparation runs without it. A data dictionary, schema guide or DDL comments let CaseQuery read your codes and terminology, and it compares them against the database, so it can also point out what looks outdated or inconsistent.
Only the documentation is sent through your configured AI account. Database records are never included.
Paste your data dictionary, schema notes or DDL comments, or choose a file. Plain text, Markdown, CSV, .sql, Word (.docx) and Excel (.xlsx) are read; PDF is not read yet.
Anything that describes what your tables and columns mean will do. For example:
status — A = active, C = closed, P = pendingamount_due — stored in centsparcel_bak — backup taken before the 2023 migration
Designed for production databases
This is the most database-intensive thing CaseQuery does. It watches query times, eases off while the database is busy, and reads statistics rather than scanning very large tables — so it is safe to run on a live system, and quicker outside peak hours.
Checking what this will involve…
Tables to leave out
Unticked tables are left out of preparation and hidden from the assistant. They stay listed here. A change takes effect the next time you press Prepare, and a table switched back on has no column notes until then.
Linked records with start and end dates
Tick a link to have preparation check whether records with one still open (no end date yet) sit under more than one status. If they do, Fine-tuning asks how they should be counted. Unticked links are never checked. A change takes effect the next time you press Prepare.
Profiles enough data to understand the schema while keeping database load low.
Starting…
Check for new definitions
Re-reads your database for words it could read more than one way, and refreshes the leads' list on Fine-tuning › 1 · Definitions to settle. Preparing the database does this too.
Use it after the database changes: new tables or columns, a new status value, or a large data load. Nothing already settled is undone or asked again.
Computed with SQL against your database — no AI tokens, and nothing to pay your provider.
4 · Question-SQL pairs
Pair a question your firm asks constantly with the query that answers it. From then on that question is answered from your query, verbatim — no AI call, no wait, and the same figure every time. The questions asked most often become the ones that cost nothing to answer.
None
When to use this instead of a rule
A rule teaches the assistant what a word means, and works on questions nobody has asked yet. A pair fixes one question’s answer exactly. Reach for a rule first; reach for a pair when the answer has to be identical every time, when the query is more than a sentence can describe, or when a question is asked often enough that paying for it every time is waste.
Matched on the wording. When somebody asks it a different way they are shown this question and asked whether it is the same one — and you decide which of those wordings to attach.
Yours to change. It runs through the same safeguards as any other query — read-only, the row ceiling, and the classified-column rules.
Shown with the answer, so the person reading it knows what they are looking at.
0 questions
Your figures, not ours
These are numbers your firm has already established as correct — taken from reports you already run and already trust. The system answers each question and we show which figures it reproduced exactly. Nothing here is ever edited to match an answer: if a figure and an answer disagree, that is the finding. Add more whenever you have a number you would defend, and re-run the check after any change to your business rules.
The planner does not produce identical SQL every time, so a single pass cannot tell a real fix from a lucky one. Asking three times costs three times as much and says considerably more.
Advanced: edit as text
Every case, including the settings the form does not show — which column carries the figure, label aliases, how close counts as equal, and whether rows your report leaves out are expected. Saved only if every case is valid, and you are told everything that is wrong at once.
Your firm's AI provider
Not set
Your firm uses its own account with the AI provider. The AI is shown your database's
structure and your questions — never a record from it. The key is sealed by this
service and never shown again; only its last four characters are.
This service cannot keep AI keys yet:
it was started without its key file.
The region your Bedrock models run in. It decides where your questions are processed.
Your Azure resource's address, from the Foundry portal.
Leave blank for the provider's usual model.
API key
Answer speed
How long the model spends thinking before it answers. Quicker is not a worse answer for
most questions — it is less time spent reconsidering one it had already reached. This
is the setting that changes how long a question takes; the fast-mode switch below is a
separate, premium-priced feature most accounts cannot use.
Second model for checking rules
Before a rule is recorded, the same question is planned by a second model as well. If
both write the same query, the rule says one thing and one thing only. If they differ,
the wording still admits more than one reading — and you can fix it before it
applies to everyone’s answers.
Starter questions
The example questions shown on the chat screen before anyone has asked anything. They are presentation only and never reach the AI, so word them in your customer's own vocabulary.
Ask Casey
Casey is the guide to this product. Casey answers questions about how to use AI CaseQuery — what a rule is, what a warning means, how to prepare a database — from a switch beside the question box and from a panel available on every screen. Casey is never sent rows from your database, or any value read from your records. A note your administrator wrote may itself name a coded column's values — A, C, P — and those notes do travel; what never travels is anything read out of a record. What Casey receives, whichever way the second switch is set: how the product works, whatever is typed into the question, and — where Casey is opened from a task rather than from the question box — a briefing on that task. A briefing can carry the wording you are drafting, the rules and settled definitions your firm already holds about the same thing, the question an answer came from, table and column names, and counts from your preparation run. Where a question was refused by the safeguards, Casey is told only which category was recognised, never the question itself. The second switch below is a different thing: it sends the set of rules, settled definitions and column notes assembled for Casey with every question it is asked, and names the categories and counts what will be sent before you agree to it. That set is capped, so a large firm sends what fits rather than everything it has.
Casey is the guide to this product. Casey answers questions about how to use AI CaseQuery — what a rule is, what a warning means, how to prepare a database — from a switch beside the question box and from a panel available on every screen. Casey is never sent rows from your database, or any value read from your records. A note your administrator wrote, or accepted from preparation, may itself name a coded column's values — A, C, P — and those notes do travel, as they do when the planner is given them. What Casey receives, whichever way the second switch is set: how the product works, the user guides, whatever is typed into the question, and — where Casey is opened from a screen or a worklist item — the product's own description of that screen or item. Nothing measured from your data is sent with it. Casey's questions are not kept by the service. The second switch below is a different thing: it sends the set of rules, settled definitions and column notes assembled for Casey with every question it is asked, and names the categories and counts what will be sent before you agree to it. That set is capped, so a large firm sends what fits rather than everything it has.
It uses the same AI provider and key as everything else, so questions asked here are charged to your own account.
!
This question cannot be sent
Nothing was sent to the AI provider. This question appears to contain information your organisation does not allow to be sent, so it was stopped on this computer.
Nothing was sent to the AI provider. This question appears to contain information your organisation does not allow to be sent, so it was stopped before it left.
Ask the same thing without the specific value — for example by name, date or status rather than by identifier — and the application can usually still answer it.
Worth checking
!
Correct this measurement
Your correction goes to the top of Fine-tuning → 1 · Definitions to settle, where whoever holds the lead key can settle it as a definition. A copy is also written to a file for whoever supports this installation — your question, this measurement, whatever you pick and write below, the query that produced the answer, your name, the date, and this computer's name with the version and database it was using. It is not sent to the AI and does not change this answer.
What went wrong?
Pick any that apply. Optional — the box below is the part that gets read.
✎
Proposed standing instruction
Nothing has been saved. Review the wording, see what it changes, then accept or discard.
Re-runs the original question with and without this instruction — and once more without it if the query changes, to check the difference is not the AI answering differently. The AI then explains each query in plain English: one or two more calls.
The before and after appear
here. Nothing is saved by looking.
Fine-tuning summary
Counts and setup steps only — no question text, no table or column names, and no value from any table. Safe to send on.
≈
Review estimated token use
Nothing has been sent to the AI provider yet.
Estimated input—
Maximum output—
Estimated maximum total—
View estimate details
This is an approximate local estimate. Actual use is shown after the request.
Check this answer
What to decide
Does this answer match what you know to be true? You picked these questions because you already know the answer — that is what makes the check worth anything.
Right
Recorded in your name against today's date, with the query that produced it. This is your evidence the answer was correct on this database. The query is kept for later use and nothing reads it yet.
Wrong
Asks you for one sentence saying what it got wrong, then sends it to Fine-tuning › Definitions to settle, where whoever holds the lead key settles it as a definition. Your sentence is all they get, so it needs to say what was wrong rather than that something was.
Doesn't apply
Moves to Inactive at the foot of the list. Kept, not deleted — mark it Right or Wrong at any time and it returns to its own section.
Needs a decision
For when you cannot tell which reading the firm means. Nothing is called wrong — it goes to the settling list as a handover, for a lead to choose which reading governs.
Ask the DBA
For when the answer exists in your data but nobody can say it in words. Goes to the DBA list in Settings › DB Tuning with your sentence.
All five are recorded in your name and need your firm's access key. Enter it once and the rest of the session is uninterrupted.
✍️
Your name
Recorded so the firm knows who decided this.
⚠️
Are you sure?
✍️
Write the definition
Sent to the AI exactly as written — with every question while there are a dozen rules and definitions or fewer, and past that with the questions it concerns.
🔒
Administrator access
Settings is for administrators — the licence, security, your database, the AI provider and the schema tools. Enter the access key supplied with your licence.
This installation uses your work account. Your firm's own sign-in and multi-factor policies apply; no password is entered here or stored by this application.
Access stays unlocked in this browser until it is closed.
🔒
Confirm it is you
This changes who can reach your firm's data, or where it goes. Enter your password again to go on.
This changes who can reach your firm's data, or where it goes. Sign in with your work account
again to go on: a Microsoft window opens, and this closes by itself once you have.