Part 5 · Worked examples

Working with Ask

Ask is the read-only assistant in a review’s header. It answers questions about the review’s record and its data, and shows where every figure came from. This example walks through thirteen steps on the public demo review, from opening Ask to starting a fresh conversation.

Manual v1.0 · Last changed 9 November 2026

Early access
This works today for teams that have asked for it; ask us to turn it on for your workspace. What it does and records may still change.

Use it to see what to ask, what each answer shows, and how to tell where a figure came from. The figures are real results from the demo review. For what Ask answers and how its answers are marked, read 2.3 Audit trail and history; for the assistant as a whole, read 2.4 What the AI does by itself.

Public demo850 records8 included studies · 19 reports13 steps
What this example covers
The public demo review: a personal review that reproduces a published meta-analysis of neoadjuvant chemoimmunotherapy for resectable non-small-cell lung cancer, with 850 records. By its PRISMA flow, 8 studies (19 reports) are included. The questions were asked on 2 November 2026. A personal review has one person and no sign-offs, so a question about a team, such as how well two screeners agreed, is answered as not applicable. Ask’s wording varies and the review can change, so your answers may differ from these; their figures always come from the review.
Follow along in the demo
Open the demo, choose Sign-off status in the sidebar and click Ask. The demo answers a limited number of questions on each visit and shows how many are left; steps 2 to 11 ask nine. A conversation lasts only as long as the browser tab, and people appear as labels. There are no thumbs and no New conversation, so steps 12 and 13 need a review of your own. In a review of your own, Ask on Synthesis opens that page’s own assistant instead, as it does on Overview and Analysis if you can run analyses.

The steps at a glance

StepWhereYou ask or do
1. Open AskSign-off statusClick Ask in the header
2. Ask about the recordSign-off statusWhy were records excluded at full text?
3. Check where the figures came fromSign-off statusThe same answer’s card
4. See where people overruled the AISign-off statusWhere did a person’s decision differ from the AI’s?
5. Ask for a number that needs arithmeticSign-off statusHow many reports assessed at full text were not excluded?
6. Follow up on an earlier answerSign-off statusRemind me of the full-text override rate you gave earlier.
7. Narrow the questionSign-off statusWhich full-text decisions quote a passage that was not found in the document?
8. Ask who did what, and whenSign-off statusWho changed extracted values most recently?
9. Ask about the page you are onExtractionWhat was extracted for this study and what still awaits review?
10. Quote a saved resultExtractionWhat is the saved pooled hazard ratio for event-free survival, with its confidence interval and GRADE certainty?
11. Ask it to change somethingExtractionMark this study’s overall survival value as validated.
12. Say whether an answer was rightUnder an answerThumbs up or thumbs down
13. Start a fresh conversationAsk’s headerNew conversation

Step 1 of 13

Open Ask

Click Ask at the right of the review’s header. Ask opens beside the page and stays open as you move between the review’s pages, so an answer that is still arriving is not lost.

  • A new conversation offers starting questions that suit the page. On Sign-off status they ask about the record: the screening funnel, screener activity and the review’s timeline. Click one to ask it, or type your own question.
  • The Read-only badge in Ask’s header means what it says: Ask never changes the review.
www.axelium.aiSign-off status
  1. 1. Ask, in the review’s header.
  2. 2. The same question box answers from the record too: who did what, when, who confirmed it.
Fig 1Ask opened on Sign-off status with its starting questions. Captured 9 Nov 2026

Step 2 of 13

Ask about the record

You ask

Why were records excluded at full text?

The record is everything the review keeps about who did what, and when: screening decisions and their quotes, approvals, searches, documents and extraction.

The answer opens with a line that names the part of the record Ask read, here Screening funnel, marked Completed. Then come the words, with a table where one helps. Ask answered that the recorded PRISMA flow shows 43 records excluded at full text, for these reasons:

Recorded exclusion reasonRecords
Wrong study design27
Wrong comparator15
Not the primary report1

To see the reason and the quoted passages behind one record’s decision, ask about that record by its title.

Step 3 of 13

Check where the figures came from

Under the words, a card holds every figure the answer used, each labelled with its source. Its heading says where the answer came from, here From the review’s record, and its footer says when the figures were read and what they were counted from.

FigureValueSource
Assessed at full text62PRISMA 2020
Excluded at full text43PRISMA 2020
Excluded at full text: wrong study design27Full-text screening
Excluded at full text: wrong comparator15Full-text screening
Excluded at full text: not the primary report1Full-text screening
Included8PRISMA 2020
  • Data as of 2 Nov 2026, 23:50 is when Ask read the record for this answer, shown in your own time zone.
  • A note under the figures names what they were counted from: here, the PRISMA 2020 summary, with the day it was saved.
  • The thumbs under the card are for your verdict on the answer; see step 12.

Step 4 of 13

See where people overruled the AI

You ask

Where did a person’s decision differ from the AI’s?

Ask split the differences by screening phase, and by whether a person reversed a firm AI verdict or decided a record the AI had left unsure.

Screening phasePeople’s decisionsDiffered from the AIOverride rate
Title and abstract4747100%
Full text442965.9%
Total917683.5%

Of the 76 differences, 9 reversed a firm AI verdict and 67 were decisions on records the AI had left unsure.

An override rate is not an error rate
An override counts a person’s decision on a record the AI marked unsure. Here most differences are of that kind, so the override rate says little about how often the AI was wrong. The card carries the same caveat under its figures, so it travels with them.

Step 5 of 13

Ask for a number that needs arithmetic

You ask

How many reports assessed at full text were not excluded?

Some answers need arithmetic. Ask answered 19: 62 reports assessed minus 43 excluded, and added that 19 counts reports, not included studies.

A number Ask works out from the card’s figures is underlined in the answer, with a small calculator mark, and the card shows the sum: Computed 19 is computed from the figures (62 − 43).

Step 6 of 13

Follow up on an earlier answer

You ask

Remind me of the full-text override rate you gave earlier.

Ask remembers the conversation. It can answer a follow-up from figures it has already shown, without reading the record again. Here it answered 65.9%: 29 of 44 people’s decisions at full text differed from the AI’s, 22 of them on records the AI had left unsure and 7 reversing a firm verdict.

  • Each reused figure is marked Carried over, with the time it was read.
  • The card’s heading reads From this conversation, and its footer From earlier answers in this conversation.
  • For the figures as they are now, ask the question afresh.

Step 7 of 13

Narrow the question

You ask

Which full-text decisions quote a passage that was not found in the document?

Ask listed each decision with its study and the criterion the passage was quoted for. When a question is about part of the review, the card says what its figures were narrowed to.

  • Here the card’s heading reads From the review’s record · narrowed to quotes not found in the document, and its figures describe only those quotes: 13 quoted passages, none of them found.
  • An unmatched passage is a reason to check the decision, not proof that it was wrong.

Step 8 of 13

Ask who did what, and when

You ask

Who changed extracted values most recently?

Questions about activity are answered from the review’s audit trail and its approvals. On 2 November the answer named who last edited an extracted value and when (in UTC), and gave the latest validation as a separate act: an edit changes a value, a validation confirms it. Its card was narrowed to entries about extracted values.

  • The review has changed since 2 November, so the people and times you see will differ.
  • A member of the review sees people by name, as Sign-off status shows them. A person outside the review appears as a label, as everyone does in the demo, and no email address is shown.

Step 9 of 13

Ask about the page you are on

You ask

What was extracted for this study and what still awaits review?

Ask knows which page you are on. When the page shows one study, as Extraction does, a Looking at bar under Ask’s header names it, and “this study” means that one. Clear the bar with its cross to ask about the whole review.

www.axelium.aiExtraction
  1. 1. Ask, in the review’s header.
  2. 2. On the extraction page it knows which study you are looking at.
Fig 2Ask opened on Extraction, with the Looking at bar naming the study in view: here another study of the review than the one below. Captured 9 Nov 2026

Here the study was the 2023 report Perioperative Pembrolizumab for Early-Stage Non-Small-Cell Lung Cancer. All four endpoints had extracted values, three of them validated:

EndpointTimepointReview status
Event-free survival (or progression-free survival)Longest follow-upValidated; ready for synthesis
Overall survival (OS)Longest follow-upAutomatically extracted; awaiting review
Pathological complete response (pCR)Post-neoadjuvant, at surgeryValidated; ready for synthesis
Major pathological response (MPR)Post-neoadjuvant, at surgeryValidated; ready for synthesis
  • An answer from the review’s data, such as extraction, says Answered from the review’s data, with the time.
  • The answer keeps a value awaiting review apart from one the paper does not report.

Step 10 of 13

Quote a saved result

You ask

What is the saved pooled hazard ratio for event-free survival, with its confidence interval and GRADE certainty?

Ask reports a saved analysis as it was saved: the estimate, its interval, the number of studies, heterogeneity, the GRADE rating, and which data the analysis used.

Saved resultValue
Hazard ratio, event-free survival at longest follow-up0.579
95% confidence interval0.504 to 0.664
Studies6, using validated data only
HeterogeneityI² = 15.6%
GRADE certaintyHigh, finalised
  • Ask quotes saved results and never recalculates. A new analysis, a subgroup analysis, a meta-regression or a check that leaves named trials out is not an outcome’s saved result; run or open those on the Analysis page. A GRADE rating is quoted only beside the analysis it rated.
  • When an outcome has more than one saved analysis, say which one you mean.

Step 11 of 13

Ask it to change something

You ask

Mark this study’s overall survival value as validated.

Ask is read-only. Instead of acting, it said where the change is made: on the Extraction page, by opening the study’s Overall survival (OS) value, reviewing it and validating it.

  • An answer that needed no figures says No figures: answered without reading the record.

Step 12 of 13

Say whether an answer was right

Point at any answer and a row of controls appears under it: Copy, Retry on the latest answer, a thumbs up and a thumbs down.

  • Thumbs down asks why: Wrong, a figure or a fact is not right, or Unclear, the answer is hard to follow.
  • After either, a box opens for a short note, which you send with Send.
  • The answer then reads Marked as wrong, Marked as unclear or Marked as helpful; Change gives another verdict in its place.
  • A verdict changes nothing in the review.

Step 13 of 13

Start a fresh conversation

Click New conversation at the top of Ask. It asks first, Start a new conversation?, because the questions and answers so far are cleared and cannot be opened again; nothing in the review changes. Choose Start new conversation, and Ask starts again with its starting questions.

  • Answers already listed under Answers in Figures and tables stay there.
  • The button is greyed out while there is nothing to clear or an answer is still arriving.
  • Start fresh before a new line of questions, so that a follow-up does not lean on figures from an earlier topic.

Reading the marks

Every answer says where its figures came from. The card’s heading reads From the review’s record, From the review’s data or From this conversation. These are the other lines and marks to look for:

What the answer showsWhat it means
Data as of 2 Nov 2026, 23:50When Ask read the record (step 3).
Answered from the review’s data at 23:54Read from the review’s data, such as extraction or saved analyses (step 9).
From earlier answers in this conversationFigures reused from earlier answers (step 6).
No figures: answered without reading the recordNothing was read, such as for a declined request (step 11).
narrowed to …The figures cover only that part of the review (step 7).
Computed: 19 is computed from the figures (62 − 43).Worked out from the card’s figures (step 5).
Carried over · read 23:33A reused figure, and when it was read (step 6).
UnverifiedA number that no lookup returned and that is not worked out from the figures. It is underlined in the warning colour: check it before you rely on it. Dates, times, years and the numbers in a trial’s name are never marked.

Other questions to ask

  • Which databases were searched, and did any search fail?
  • On which day did a person make the most screening decisions?
  • Were any attached papers refused as the wrong paper?
  • How much has been extracted and validated for each outcome?
  • Which reports included at full text have no validated extracted value yet?
  • Which items did the AI act on that no person has confirmed yet?
  • Who signed off each stage of this review, and when?
  • What has changed in this review recently, and who did it?
  • How long did each stage of this review take?
Asking well
Ask one thing at a time, name the stage or the outcome you mean, and open the page a question is about before you say “this study”.

In a review of your own, the same questions can be asked in the Overview's box and on the Analysis page, if you can run analyses. The statistics assistant answers them there with the same card. The demo turns both off.

Learn more:2.3 Audit trail and history · data as ofLearn more:2.3 Audit trail and history · your verdictLearn more:2.4 What the AI does by itself · Ask about your data