Studio AI Assistant

AI Assistant

The AI assistant (Plus) reads your work and makes suggestions. It drafts a questionnaire from a brief, rewords a question, proposes answer options, reviews the wording of a whole questionnaire and reviews an analysis flow. This page covers each feature, how to turn the assistant on, what it sends where, and how its allowance of credits works.

Coding open answers is not one of them. AI coding of open answers is switched off on this platform, because it sent what respondents wrote to the model provider. You code open answers yourself, by hand and by word rules, in the codeframe editor — no assistant, no credits. See Coding open answers is switched off.

---

Suggestions, not edits

The assistant never changes your questionnaire or your flows by itself. Everything it returns is a proposal shown next to what you have:

• nothing changes until you press Use this, Use this wording, Replace options or Use this draft;
• an applied suggestion is an ordinary edit: it can be undone, and it becomes part of your study only when you Save;
• answer codes stay yours. The assistant proposes labels only, and options you already have keep their codes;
• its findings are labeled as opinions ("suggestions, not rules"), separate from the engine's checks, which are facts about the document.

---

Turning it on

The assistant is included from Plus and is off until the organization's owner turns it on. Turning it on is a consent decision:

By turning this on you agree that questionnaire and analysis text from this organization may be sent to a third-party model provider, currently DeepSeek, which processes it in China. What is sent, and what never is, is set out in the Privacy Policy and the Terms of Use.

1. Open Organization → Settings → Integrations.
2. On the AI assistant card ("Reviews the wording of a questionnaire and the soundness of an analysis, on request. Off until you turn it on."), click Turn the assistant on ("Turning on…").
3. The card now reads "On · turned on by on ." Turn the assistant off reverses it at any time.

Only the owner can do this. Admins and members see "Only the organization owner can make this decision, because it sends your team's work to a provider in another country." On Free the card reads "The assistant is a Plus feature". See Integrations.

What is sent, and what is not

Feature	Sent to the provider
Review wording	the questionnaire title and description, page names and titles, and for each question its id, type, text, hint, whether it is required, the scale of its variable, and its answer labels; plus the engine's own findings, so they are not repeated
Reword, Suggest options	that one question (id, type, text, hint, required, scale, answer labels), the survey's title and description, and your note
Draft from a brief	your brief, the working title and the language you named
Flow review	the flow's title and description, its steps with their settings and connections, and the codebook (variable names, labels, scales)

Never sent: your theme, scripts, quotas, files, contact lists, or anything a respondent wrote. (AI coding of open answers, the one feature that sent respondents' answers, is switched off on this platform — which is why the consent above asks only about questionnaire and analysis text.)

---

Draft from a brief

Available in a project with no questionnaire yet: the Builder's empty state ("No questionnaire yet") shows Draft from a brief next to Create questionnaire.

The Draft from a brief dialog: "Describe who you are asking and what you want to learn. The assistant proposes questions; you read them, change them and save them — nothing is saved for you." followed by "About N of your R remaining credits."

Field	Notes
Brief	up to 8,000 characters: who the respondents are, what you want to learn
Language ("optional — the brief's own language by default")	e.g. English, German

Draft questions ("Drafting…") returns "<title> — N questions across P page(s). Read it in the Builder before you save." with a list of pages and questions. Try another brief goes back; Use this draft opens the ordinary Save dialog, so the draft becomes your first Save. The organization's house style is applied on the way, as it is for Create questionnaire (see below).

What a draft contains:

• up to 12 pages and 40 questions, grouped into short pages;
• only Single choice, Multiple choice, Likert scale (3 to 11 points, 5 by default), Number and multi-line Open text questions; no matrix, ranking, MaxDiff or conjoint;
• choice questions with 2 to 12 options, coded 1, 2, 3…, and a codebook entry for every question;
• readable variable names chosen by the assistant (visit_frequency, would_recommend): lower case, letters, digits and _, at most 30 characters before any suffix. A name used twice gets _2, _3…, and a question without a usable name is numbered by its position (q7). Each question's Id is the same as its variable name;
• a hint under a question when the assistant writes one (up to 300 characters);
• no theme settings of its own. Use this draft puts your organization's house style under it, as Create questionnaire does: colors, type, logo, institution, privacy link and the rest, custom CSS included (both the assistant and custom CSS come with Plus). See The organization's house style.

If the brief gave the assistant nothing to work with: "The assistant made nothing usable of this brief. Say who the respondents are and what you want to learn from them, and try again — the attempt was still charged." with Back to the brief.

On Pro and Corporate, and during the Pro trial, drafting uses a larger model. Plus uses the standard one.

---

Review wording

Builder → Validation → Assistant: "wording and response options, read by a language model — suggestions, not rules". It sits under the engine's checks on purpose: the engine reports facts, the assistant offers opinions.

• Before a review: "Not read yet. It looks for leading or double-barreled questions, unbalanced scales and missing opt-outs — the things the strict lint above cannot see. Costs about N of your R remaining credits."
• Review wording ("Reading…") reads the questionnaire on screen, unsaved edits included.
• Each finding shows a code, one sentence and often "Suggested: ". Use this wording applies it to the question ("Applied — save to keep it."). Clicking a finding opens its question.
• Nothing wrong: "Nothing to flag across N questions."
• The review describes the version it read, so it disappears as soon as you edit.
• At most 40 findings per review.

Code	Meaning
LEADING	the wording pushes the respondent toward one answer
DOUBLE_BARRELLED	one question asks about two things, so an answer is ambiguous
ASSUMES	the question presupposes something that may not be true of the respondent
MISSING_OPT_OUT	no honest answer for someone to whom the question does not apply, who does not know, or would rather not say
UNBALANCED_SCALE	more options on one side of the scale than the other
OVERLAPPING_OPTIONS	answer options overlap, or leave a gap
SCALE_MISMATCH	the answer labels do not fit what the question asks
JARGON	a term respondents are unlikely to understand the same way
SENSITIVE	a sensitive question without a way to decline
OVERLONG	a sentence long or convoluted enough to be misread

---

Reword and Suggest options

In the Inspector, a question's Question section has:

• Reword, which proposes up to three alternative wordings, each with a few words on what it fixes. Use this replaces the question text. If the wording is already fine: "Nothing to change — the assistant would leave this wording as it is."
• Suggest options (single choice, multiple choice and ranking only), which proposes up to 12 answer options in reading order, with a one-line note. Replace options applies them ("Options you already have keep their codes."), including options a question shows from its variable's value labels. A proposed label that matches an existing one (ignoring case) keeps that option's code, new labels get the lowest free codes, and options not in the proposal are removed, with their value labels. If nothing comes back: "The assistant proposed no options for this question."
• What to fix (optional): a short note for either button, such as "too formal" or "add a don't-know option". Keep it under 300 characters.

Both buttons show "Thinking…" while they work, and are disabled until the question has text.

---

Flow review

On Flows, each flow in the list has a review link (tooltip "Have the assistant read this pipeline for analysis mistakes"). It reads the flow's saved version and opens a panel under the list: "Assistant · · whether the analysis is sound, read by a language model — suggestions, not rules · N credits, R left", with Close.

• While reading: "Reading the pipeline…".
• Findings show a code, a sentence and often "Suggested: …", with the step they concern. The assistant knows each node's choices, so a suggestion can name the one to pick — a Test of kruskal with Post-hoc dunn, a Crosstab with Fisher's exact test for thin cells. There is nothing to apply; change the flow yourself.
• "Nothing to flag across N steps.", or "This flow has no steps to read yet."
• The flow must pass its structural check first. If it does not, the review is refused ("fix what the structural check reports first — …") and the panel reads "Nothing came back. The structural check has to pass first."
• Clicking review again closes the panel. A second read costs credits again.

Code	Meaning
WEIGHTS_UNUSED	weights are computed but never applied
WEIGHTS_INCONSISTENT	some analyses on the data are weighted and others are not
MISSING_NOT_HANDLED	a statistic counts missing codes (refusals, don't-knows) as real values
TEST_SCALE_MISMATCH	a test or statistic does not suit the variable's scale
NO_DEDUP	duplicates are not removed, so a respondent may count twice
PARTIALS_INCLUDED	incomplete or implausibly fast responses are not filtered out
SPARSE_CROSSTAB	a crosstab on a variable with so many categories that cells will be too thin
ORDER	a step comes after one that already invalidated it (filtering after weighting, recoding after the analysis)
DEAD_END	a node's result is never used
UNREPORTED	a result reaches neither the report nor a Live tile

See Analysis Flows.

---

Coding open answers is switched off

The assistant used to build coding schemes for open answers: Code open answers… on a flow's Code open answers node had it read the answers to one open-text question, propose themes and assign each answer to one. That meant sending the text respondents wrote — not your team's own text — to the model provider, DeepSeek, which processes it in China. It is switched off on this platform for now:

• the Code open answers… and Code more answers… buttons and their dialog are gone;
• a request to start a coding job through the API is refused before anything is read or charged: "AI coding of open answers is switched off on this platform.", and a job queued before the switch ends as failed with the same words, without reading an answer;
• codeframes the assistant built earlier stay in their projects, and the Code open answers node applies them as before, without any model.

Open answers are coded in the codeframe editor instead: by hand, and by word rules that also code the answers collected later. It runs in Studio, sends nothing anywhere, is on every plan and spends no credits. See Coding Open Answers.

---

Plans and credits

The assistant spends credits. One credit is roughly 1,000 tokens of text. Allowances are per organization:

Plan	Credits per month	Credits per day	Requests per person, per hour
Free	not available	—	—
Plus	8,000	2,000	30
Pro	50,000	8,000	100
Corporate	300,000	30,000	300
Pro trial (unpaid)	500, once (does not renew)	200	10

• The monthly allowance resets at the start of each calendar month, and the daily one at midnight UTC.
• A trial's 500 credits are for the organization's whole trial. When they are gone, the assistant stops until you subscribe.
• The requests-per-hour limit counts each person's assistant requests of all kinds.

Typical costs:

Feature	Typical credits
Review wording (a questionnaire of ordinary length)	about 11
Reword, Suggest options	about 2 each
Draft from a brief	about 34 on Plus, about 115 on Pro and Corporate (larger model)
Flow review	about 7

You are charged for what the provider actually processed. Before sending, Studio checks that your remaining allowance covers a cautious estimate of the largest possible reply, so near the end of an allowance a request can be refused even though its typical cost would fit. A refused request costs nothing. A completed one is charged even if its result is empty.

Remaining credits appear in the Validation → Assistant hint, in the Draft from a brief dialog and in the flow review's header.

---

When the assistant is unavailable

When the assistant cannot be used, the Reword, Suggest options and Draft from a brief buttons are hidden. Validation → Assistant says why:

Reason	Message
plan	"The assistant is part of Plus and above."
not turned on	"Off for this organization. An owner can turn it on in Settings — that also confirms questionnaire text may be sent to the model provider."
paused by support	"Paused for this organization. Contact support."
platform limit	"Temporarily unavailable. Try again later."

Error messages

A failed request shows a toast such as "The assistant could not review this" (or "…could not rewrite this question", "…could not propose options", "…could not draft this", "…could not review this flow"), followed by the reason:

Message	What to do
"the assistant is not included in the plan — upgrade to Plus", for example "…in the Free plan…"	upgrade
"the assistant is off for this organization — an owner can turn it on in Settings, which also confirms that questionnaire text may be sent to the model provider"	ask the owner
"the assistant is paused for this organization — contact support"	contact support
"the assistant is temporarily unavailable — please try again later"	try later
"this organization's assistant daily allowance cannot cover this request (X of Y credits used) — it resets at midnight UTC; try a smaller document if needed"	wait, or send less
"this organization's assistant monthly allowance cannot cover this request (X of Y credits used) — it resets next month, or support can raise it"	wait, upgrade, or contact support
"this organization's assistant trial allowance cannot cover this request (X of 500 credits used) — subscribe to keep using it"	subscribe
"the assistant is limited to N requests an hour per person — try again shortly"	wait a little
"this questionnaire is too long for the assistant to review in one pass" (similarly "this question is too long to send", "this brief is too long to send", "this pipeline is too large for the assistant to review in one pass")	shorten it
"there are no questions to review yet" / "the question has no text yet" / "this flow has no steps to review yet"	add content first
"the assistant's draft could not be turned into a questionnaire — nothing was saved"	try again

---

Good practice

• Treat findings as a second reader's notes: accept what improves the question, ignore the rest. The engine's checks decide validity; the assistant does not.
• Review wording before you pilot. Rewording a question mid-field changes what the data means.
• Use What to fix to steer a rewrite. "Make it neutral" or "simpler for teenagers" works better than a bare request.
• After Replace options, check the codebook: removed options disappear from the question and from its value labels, and new ones take new codes. On a multiple-choice question in the wide layout, each new option also gets its own 0/1 variable, and a removed option's variable goes.
• After Use this draft, check the theme (it starts from your organization's house style, if there is one) and read every question as a respondent would.
• Tell your ethics board and your respondents' consent text what you send to the provider, if your institution requires it.

See also

• Testing Your Survey
• Coding Open Answers
• Analysis Flows
• Plans, Trial and Billing
• Organizations and Team

---

← Testing Your Survey · Studio contents · Publishing and Environments →