AI activity
Other languages
The case's record of how AI was used in it, with who did what, when and with which model, to read in Claira or download as a PDF or an Excel workbook and keep with the matter.
AI activity
When a client, opposing counsel or a court asks how AI was used in a review, the answer has to be specific: which documents were read, with which prompt, by which model, who ran it and when, and which version of the prompt the team settled on, with the comments that explain it. AI activity gives that answer for each Discover case, in one place, and lets you download it to keep with the matter.
It lists every bulk scan, the single scans each person ran each day, every prompt version and the comments on it, every Insight and revision, every Case Context change, every early case assessment, every Agent Mode chat, the fields the Agent created, feedback, the moments Claira switched to another model, and every report downloaded from the case.
When to use it
- Before a run. Once a prompt is ready and marked Final, download AI activity to file the prompt's history, with the comments that explain each change, before the prompt is run across the documents. The Excel workbook holds the text of every version; the PDF prints the Final version and the versions that carry comments in full, and lists the others.
- During the review. Check which version of a prompt each bulk scan ran, and whether it was the Final one.
- When the case closes. Download the record before the case's data is deleted, so the matter keeps it after Claira no longer does.
Opening AI activity
- Open Claira in the case.
- Open the mode menu, choose More, then AI activity.
The record is put together when you open it, from what Claira already keeps about the case, so it is always current. The line at the foot of the view, As of followed by a date and time, says when it was read. Click the refresh icon (Refresh) at the top of the view to read it again, for example after a scan finishes.
Two ways to read it
Six figures at the top sum up the case, whichever layout you choose: Bulk scans, Documents read by AI, Prompt versions, Insights, Agent Mode chats and People, followed by the Period from the first activity to the last. Under the period, a line such as Times are shown in Eastern Time (UTC-05:00). names the time zone of every time in the view: your own, the one your browser uses. Nothing changes in what Claira stores; a colleague in another time zone sees the same activity in theirs.
Documents read by AI counts each document once, however many scans read it. A document in a bulk scan counts once the AI has processed it, whether or not an answer was written to its field. The line under the figure splits it between bulk scans and single scans; a document read both ways is counted in each, so the two can add up to more than the total. When the case has single scans, a line under Bulk scans, such as Single scans: 14, gives their number.
Below the figures, the switch shows the layout you are in, Timeline or Summary. AI activity opens in Timeline the first time. Click the switch to choose the other; Claira remembers your choice in this browser.
Timeline
Everything that happened, newest first and grouped by day. People are named by their Nuix Discover username, on screen and in both downloads. Text quoted under a line (a comment, the prompt an Insight was generated from, a revision's instruction) shows its first three lines on screen; the downloads give it in full. Each line says what happened, who did it and when, and adds the details that matter for that kind of activity:
- Bulk scans: the scan's reference (such as SC012) and name, its status, how many documents were written out of how many were selected, how many are partial (at least one field was not saved) and how many failed, the model, the prompt version it ran, and Ran the Final version or Did not run the Final version. A scan that did not run the Final version carries an amber warning sign, so it is easy to spot (see Prompt versions and the Final version). Skipped documents and the destination field are not on the line: they are in the PDF and the workbook. A scan started from an Agent Mode chat also carries From Agent Mode, under the name of the person whose chat it was, followed by how it started: The Agent started the scan itself under Give permissions, or the name of the person who approved it on the Agent's card (see Tasks the Agent started). From Agent Mode with nothing after it means the scan was started before Claira kept that detail, so the record does not say whether a person approved it. A scan whose prompt came from a template carries From a template. These details are on the timeline on screen, in the PDF's Timeline and in the Details column of the workbook's Timeline sheet; the PDF's Bulk scans entries and the workbook's Bulk scans sheet do not carry them.
- Single scans: one line per person, model and day, with how many scans were run and on how many documents.
- Prompts: each version saved, inserted or used to scan, the version marked Final, and each comment.
- Insights: when each Insight was created and revised, the model, the prompt it was generated from and each revision's instruction, and whether it was proposed in Agent Mode.
- Case Context, early case assessment and Agent Mode: each context created, edited, updated by AI, approved or deleted; each assessment run, with its document count; each chat, with how many questions were asked and which models answered.
- Feedback, model switches and downloads: responses marked good or poor and requests to Revise with Agent; each time a model could not answer and another answered in its place, with the feature and the reason; each report downloaded from the case, and its format, an early case assessment and the list of a task's Documents needing attention (a CSV file) among them; an Insight downloaded without its source documents says Without source documents. When no model could answer at all, the line carries the amber warning sign.
Summary
The Overview comes first: a short account of how AI was used in the case, in three parts, How it unfolded, What AI was used to analyze and What came of it. It says how many people took part, over what period, in what order and to what effect. Claira writes it the first time anyone opens the Summary (a download from the view never writes one), from what everyone in the case can see in the record: dates, people, counts, task names, prompts, the kinds of Insights run, the names of Case Contexts and of fields the Agent created, and the titles of Agent Mode chats shared with the case team. It never uses document content, field values, the conversation in an Agent Mode chat or the title of a private chat.
The Overview is saved with the case, so everyone sees the same one. The line under its heading says when Claira wrote it and whose visit or request it was written for; it does not name the model that wrote it. If there has been activity since, the Overview says so: click Update to have Claira write it again from the record as it stands, and everyone then sees the new one. It is written by AI, so check it against the timeline before you rely on it. It cannot be written or updated while you use a local model. Writing or updating the Overview uses no tokens.
People who used AI and Models used come next. Below them, one section for each feature, with its count, lists what was done in it. Under Prompts, each prompt shows its number of versions and comments, how many bulk scans used it, who marked it Final, and the start of its text. Each section has a Show in the timeline button, which switches to the timeline showing only that kind of activity.
Click a section's title to collapse it, for example to get past a long list of bulk scans; a collapsed section shows the dates it covers, and Show in the timeline still works. Claira remembers which sections you collapsed in this browser. The downloads always contain every section in full.
Filtering
On the timeline, type in Search activity to keep the lines that contain every word you type: a scan's reference such as SC012, a prompt's name, words from a comment, or a person. Use the filter icon to narrow the timeline further:
- Activity: one or more kinds of activity, such as Bulk scans and Prompts. Only the kinds the case has are offered.
- Select users to filter: what one or more people did. It appears when more than one person is named in the record.
- Date: the last 7, 30 or 90 days.
Search and the filters are greyed out in Summary. They change what you see; the downloads always contain the whole record.
Opening what a line describes
Click a line to open what it describes: the bulk scan's task details, the prompt version in Prompt History, the Insight, the Agent Mode chat, the Case Context or the early case assessment. A comment on a bulk scan opens that scan, and so does a model switch that happened during one. A line you can open ends with an arrow. A line with nothing to open, such as a day of single scans, feedback, a download, a Prompt Generator draft, a field the Agent created, a deleted Case Context or chat, or another person's private chat, is not clickable.
Prompt versions and the Final version
Versions are grouped into prompts the way Prompt History groups them: automatically, by how alike their text is. Each prompt is numbered in the order it was first used, so Prompt 3, version 2 means the same version on screen and in the PDF and the workbook downloaded from AI activity. The spreadsheet exported from Prompt History numbers prompts in the order of its own list, and AI activity shows a prompt's number and text, not the title it has in Prompt History. To find a prompt in both, go by its wording, not its number (see Version and prompt numbers across downloads).
Ran the Final version and Did not run the Final version compare the version a bulk scan ran with the version marked Final when AI activity is read, not when the scan ran. The record holds the Final mark as it stands now: if someone moves the mark to another version, the verdict on earlier scans follows it, and the earlier mark is no longer listed. If you file the record as your account of a review, download it once the Final mark is settled.
A scan gets no verdict when its prompt has no version marked Final. In the PDF and in the workbook's Bulk scans sheet, the Ran the Final version row reads Yes, No (in amber in the PDF) or, without a verdict, None. A scan Claira cannot match to a version in Prompt History shows no prompt version at all.
Downloading the record
Download AI activity, at the top of the view beside Refresh, saves the whole record as a file to keep with the matter, whichever layout, search or filters are on screen. Downloading uses no tokens.
Click it, then choose the format:
- PDF (.pdf) is the file to file with the matter and to send as it is. It does not list documents one by one: only the workbook does, on its Documents sheet.
- Excel (.xlsx) is the file to sort, filter and search, with one sheet per part of the record.
Both are built in your browser from the record on screen, the one the As of line dates, so click Refresh first if work has happened in the case since you opened the view. Each file gives both moments: when the record was read and when the file was exported. Both carry the case's saved Overview when it has one. While the case has no saved Overview, for example because nobody has opened the Summary yet, the files have none either. The download menu has no option to leave a saved Overview out (see Before the record leaves your team). They are named for the day you download them: Claira-AI-Activity-<case>-<YYYYMMDD>.pdf or .xlsx.
What the PDF contains
- The header: the case, who exported the record and when, the period it covers and the time zone its times are in.
- The Overview, when the case has one, with the line that says when Claira wrote it and for whom.
- The summary: the key figures, the people involved and the models used.
- Bulk scans: one entry per scan, with who ran it, when it started and finished, its status, its model, how many documents it had and how many were written, partial, failed and skipped, its destination field, the prompt version it ran, whether that was the Final version, and its comments.
- Prompts: for each prompt used in a scan, marked Final or commented, the full text of its Final version (or its latest), who marked it Final and when, each version that carries comments with those comments, and a list of all its versions. Prompts that were saved but never used, marked Final or commented are counted, and listed in the Excel file.
- Insights, with the prompt each was generated from and every revision's instruction.
- Case Context, Agent Mode and Single scans.
- Other activity (early case assessments, the Prompt Generator, feedback, model switches and the fields the Agent created), then Downloads.
- The full Timeline.
- About this record: what the record holds and what it never holds, that it stands as of the moment it was read, which Agent Mode chats keep their titles, and As of with the moment the record was read. That moment can be earlier than Exported on in the header, when the view was open for a while before you downloaded.
A part with nothing in it is left out. Every page carries the report's name, the case, the export date and the page number, like every other Claira report. Every time in the PDF is in your time zone, the same as on screen. The header names the zone and its offset from UTC, for example Eastern Time (UTC-05:00), and the Timeline repeats it, so a person in another time zone can read the file; a file a colleague downloads names their zone. When the record crosses a daylight saving change, both offsets are named. Single scans are counted by day in UTC, so a row whose scans fall on two dates in your time zone gives both.
What the Excel workbook contains
The workbook opens on a Summary sheet: the case, who exported it, when (Exported at), when the record was read (Record read at), the time zone, the period, the people, the models, the Overview when the case has one (one row for each of its three parts) and every figure. Two rows tell you whether the file is complete: Incomplete sections names any part of the record that is missing or cut, including any value longer than an Excel cell holds (the PDF has it in full), and Documents sheet says whether every document is listed. The sheet ends with About this record, the same text as in the PDF. Then one sheet each for the Timeline, Bulk scans, Prompts (every version of every prompt, with its full text), Comments, Insights, Case Context, Agent Mode, Single scans, Other activity and Downloads, and last the Documents sheet. A part with nothing in it has no sheet. Times are written in your time zone, as on screen and in the PDF, each with its offset from UTC, such as 2026-11-10T04:30:00-05:00, and the Time zone row of the Summary sheet names the zone, so the file reads the same wherever it is opened.
The Documents sheet
The Documents sheet answers "was this document read by AI, and what happened to it?". It lists every document of every bulk scan, one row per document per scan: the scan's reference, the document's Main ID and Document ID, the Outcome, for a failure the Reason code, and when it was processed. Single scans are counted, not listed document by document, so the sheet covers bulk scans only.
| Outcome | Meaning |
|---|---|
| Written | The answer was written to the destination field. |
| Partial | In a Multi-Code scan, at least one field was saved and at least one was not. The bulk scan's task details name the fields; see Partial results and retrying documents. Once a retry saves the missing fields, the document reads Written. |
| No answer written | The model answered, but there was no value to write. |
| Failed | The document could not be processed; the reason code says why. See Error codes. |
| Skipped | The document was skipped, for example because it had no usable text. |
| Not processed | The scan has not reached the document: it stopped first, or it is still running. |
The sheet holds identifiers only, never the answer: the answer is in the document's field in Nuix Discover. The sheet lists up to 250,000 documents, so the file can be built in your browser; when a case's bulk scans read more, the Summary sheet says how many documents are not listed. Reading the documents of a large case takes a moment. While it runs, the download button shows Reading scanned documents with two numbers, such as 2 of 6: they count bulk scans, not documents. If the documents of some scans cannot be read, the file is still saved, Claira tells you so, and the Summary sheet names those scans.
Download records
Every download of AI activity is recorded for your organization's compliance records, like the other exports: who downloaded it, from which case, in which format and when. It also appears under Downloads in the next AI activity you open. The record is written before the file is saved; if it cannot be written, the download is cancelled and Claira asks you to try again.
What it never contains
AI activity names who used AI, when and with which model, and it quotes the text of prompts, comments, Insight prompts and revision instructions. Most of that text was typed by people. A prompt one of Claira's AI tools wrote is marked with who wrote it (such as Written by the Prompt Generator), and a comment the Agent added is marked From Agent Mode. The one part Claira writes about the record itself is the Overview, which is labelled as such wherever it appears. It never contains:
- the AI's answers about documents, Insight reports, Case Context text or early case assessment results;
- field values or document content;
- the conversation in an Agent Mode chat.
Those stay where they are (the answers in Nuix Discover, the reports in Claira), each with its own download.
Private chats
Agent Mode chats you own, and chats shared with the case team, are listed with their titles, on screen and in the files you download (see Before the record leaves your team). Another person's private chat is listed as Private Agent Mode chat, with its owner, its dates, how many questions were asked and which models answered, so the record shows that the chat happened without revealing what it was about. It cannot be opened. A deleted chat is listed as Deleted Agent Mode chat, without its title.
How a review was validated
AI activity does not measure or record how a review's results were validated: it holds no sample, recall or precision figure. To keep that with the record, write it as a comment on the bulk scan, in the task's details under Comments, or on the prompt version in Prompt History. Comments are quoted in AI activity, on the timeline and in the workbook's Comments sheet, and a task's comments also print in the Bulk Task History Report.
Before the record leaves your team
Read the record through before you send it to anyone outside the review team: a client, another party or a court.
- It quotes what your team typed. The PDF and the workbook carry the full text of prompts, comments, Insight prompts and revision instructions. A prompt can name your lawyers and their clients, describe each side's position or list the issues you are looking for. No download leaves that text out, so decide who may receive it.
- It lists your own private chats by title. The file is written for the person who downloads it: your own private Agent Mode chats appear with their titles, marked Private, under Agent Mode and in the Timeline. Another person's private chats appear without their titles. A download from the Admin Portal lists every private chat without its title.
- The verdict on a scan follows the Final mark as it stood when the record was read. Ran the Final version compares each scan with the version marked Final at that moment, not when the scan ran, and the file does not say so itself (see Prompt versions and the Final version). None in that row means the scan's prompt had no version marked Final, or the scan could not be matched to a prompt version, not that nothing ran. Say so in your cover note if the reader needs it.
- The Overview is written by AI. A file downloaded in Claira carries the case's saved Overview when it has one. Check it against the timeline before you rely on it. For a copy without the Overview, ask an organization admin to download AI activity from the case's page in the Admin Portal: that download leaves it out.
When part of the record is missing
If Claira cannot read part of the record, the view says which part, shows everything else, and the PDF and the workbook say the same. A case with an exceptionally large amount of activity may also reach the limit of what one record holds; the oldest entries of that part are then left out, and the view and the downloads say so.
Who can see it
Everyone working in the case can open AI activity and download it. If your organization enforces roles, the permission View and download a case's AI activity decides who can; every role has it unless an organization admin takes it away. Without it, AI activity shows a message that your role does not allow this action: ask your organization admin for that permission.
Keeping the record when a case closes
Deleting a case's data deletes everything AI activity is read from, so the record cannot be recovered afterwards. Download it first. In the Deactivate Case dialog, choosing Deactivate now and delete all case data shows a reminder with Download AI activity right there, and so does the dialog that deletes a deactivated case's data. Deactivate only keeps the data and shows no reminder.
A file downloaded from either dialog carries the case's saved Overview. When the case has none, Claira writes one for the file first, and the dialog says Claira is writing the case's Overview for this file. This can take a minute. Nobody has read that Overview on screen, so read it in the file before you rely on it. If it cannot be written, for example while you use a local model, the record is downloaded without it.
Organization admins can also read and download AI activity from the case's page in the Admin Portal, with the same figures and sections and the same PDF and Excel files, except the Overview, which the Admin Portal does not show or download. There, times are in UTC on screen and in both files, and the files say so. This works for a deactivated case too, until its data is deleted, so the record can be kept even when nobody opens the case in Discover again. In the Admin Portal, private chats are listed without their titles, and opening AI activity is recorded in the Activity feed.
Related
- Exports: every report you can download from Claira.
- Prompt History: prompt versions, Final marks and comments.
- Deleting a case's data: what is deleted, and when.
Need help? Contact us at support@claira.to.
Was this page helpful?
Continue reading
Need more help?
Contact our support team at support@claira.to — we are here to help.