AI at Work
Google Sheets AI Functions for Work: Classify Text Without Mistaking Generated Labels for Facts
Preserve original feedback, explicit ranges and versioned proposals. Review synthetic classifications and count human-reviewed interpretations separately from verified facts.
Use Google Sheets AI functions to propose text labels, then review those labels against the original feedback before counting them. Pass the relevant cells explicitly, define a small category set with an uncertainty option, preserve the input and prompt version, and keep generated text separate from the reviewed category used in reporting. A clean count of unreviewed labels can still describe the wrong things.
This LandOffer workplace guide uses eight synthetic feedback comments to demonstrate a classification review and a completed version comparison. Every proposed AI output below is an author-written simulation, not a response generated by Google Sheets or Gemini. No live spreadsheet was created, edited, or tested. The category counts were independently checked offline; those checks establish arithmetic for this example, not model accuracy or spreadsheet execution.

Define the decision before choosing the labels
The fictional team wants a weekly review queue for comments about a small document-organizing tool. Analyst Valeria prepares the classification, and product owner Omar will decide which comments need investigation. The exercise is not a customer-satisfaction score, an incident count, or a product roadmap vote.
The team defines four mutually exclusive categories for this particular queue. Bug means the comment reports an observed failure. Request means it explicitly asks for a change. Praise means it expresses approval without a competing issue. Unclear covers insufficient detail, ambiguous intent, or multiple substantive topics that cannot fit one category under this rule.
These rules are a deliberate simplification. A different team might allow multiple tags or split mixed comments into linked issues. Here, forcing one label makes Unclear necessary. Keep mixed or ambiguous rows visible for a person to examine rather than silently forcing them into the largest category.
A label describes the reviewer’s interpretation of a comment. Even the reviewed Bug category does not verify that the software has a defect: it means the text reports a failure worth investigating. There is no production log, reproduction, or customer identity in this synthetic dataset. Omar must not turn two reviewed Bug comments into a claim that two product defects were confirmed.
Provide explicit context and check feature access
Google's AI function help page documents AI(prompt, [optional range]), with Gemini as an alternative function name. Responses are text; the function does not automatically receive the entire spreadsheet or other Drive files. Use the range argument for the comment being classified rather than assume a nearby definitions tab is included.
For this proposed layout, column A contains the stable ID, B the original comment, C the proposed label, D the reviewed label, and E the review note. The category definitions are written directly into the prompt. The following is an illustrative formula for C2; it was not executed:
=AI("Use only the supplied feedback. Return exactly one label: Bug for an observed failure; Request for an explicit change request; Praise for approval without another issue; Unclear for missing context, a question without an explicit change request, or multiple substantive topics. Do not verify product facts or infer unstated intent.", B2)
B2 is the supplied feedback range, not the entire table. Before extending the formula, an authorized user should inspect the next row's reference and test a small approved sample. Do not assume a long prompt forces the model to follow every instruction or prevents an unexpected label.
Google describes selecting AI-function cells and using Generate and Insert, with Refresh and Insert for new output. Eligible plans, administrator settings, and language affect availability. Consult the current supported-languages page for the AI function specifically; a side-panel feature's language list is not interchangeable. A September 2025 language expansion announcement is historical context, not a complete current entitlement list.
A filled classification review with original text retained
Valeria's input snapshot S1 contains the following eight invented comments. Prompt P1 is the broad initial instruction, “Classify each comment as Bug, Request, Praise, or Unclear.” Its proposed outputs are deliberately constructed to illustrate plausible review problems; they are not evidence that Gemini makes these particular errors.
Valeria reviews every row in this small example. The review is a complete eight-row inspection, not a statistically representative sample of a real customer population.
| ID and original synthetic feedback | P1 simulated label | Reviewed label | Review note and conflict |
|---|---|---|---|
| F1: “The Save button does nothing after I rename a folder.” | Bug | Bug | Reports an observed failure; root cause and reproducibility unverified |
| F2: “Please add dark mode.” | Request | Request | Explicit requested change; no claim that dark mode is objectively absent |
| F3: “Search feels fast now, thanks.” | Praise | Praise | Approval in this comment; not a measured latency result |
| F4: “Export stops halfway, and I wish I could schedule exports.” | Request | Unclear | Mixed failure and feature request; needs separate human triage under the one-label rule |
| F5: “Great, another empty download.” | Praise | Bug | Context indicates a reported failed output, despite the positive-sounding first word |
| F6: “Buttons.” | Bug | Unclear | No behavior, request, or judgment can be established from this text |
| F7: “Can I change the font?” | Request | Unclear | A question might need instructions; it does not explicitly request a product change |
| F8: “Add a keyboard shortcut for archiving.” | Request | Request | Explicit requested change; no promise that it will be built |
The P1 proposal has two Bug, four Request, two Praise, and zero Unclear labels. The reviewed set has two Bug, two Request, one Praise, and three Unclear labels. Both totals equal eight. The overall Bug count happens to stay at two even though F5 moves into Bug and F6 moves out. That unchanged total conceals two different row interpretations.
Valeria's report therefore includes both the counts and the conflict notes. She routes F4 for a later decision about splitting its topics, F6 for clarification if an approved follow-up route exists, and F7 for an answer before assuming a change request. These are proposed actions within the fictional workflow; no customer contact or product investigation was performed for this article.
Before aggregating, Valeria checks the review-state column in her working record as well as the label. A blank human decision is pending work, not agreement with the proposal. An out-of-list response such as “probably a bug” is retained as returned text and sent to review rather than automatically converted into Bug. Keep that mapping explicit so the report does not mix reviewed decisions with untouched proposals.
Carry the original IDs into exports and sorted views so reviewers can reconnect each decision to its source comment. If a reviewer sorts the feedback but pastes a label column in the old order, correct-looking categories can attach to the wrong comments. Reconcile by the stable ID and input snapshot, not by row position alone. This is a data-handling check independent of which model proposed the label.
Count reviewed categories with ordinary arithmetic
Use a deterministic calculation for the numeric summary after reviewing the category column. Google's COUNTIF documentation describes counting cells that match a criterion. If the reviewed labels occupy D2:D9, this illustrative formula counts the reviewed Bug labels:
=COUNTIF(D2:D9,"Bug")
The example count is two. Do not ask an AI text function to estimate the number of bugs and treat its narrative as an exact spreadsheet calculation. Also do not nest the classification function inside a conventional formula; Google documents that embedded AI functions are unsupported. Generate, inspect, and review text first, then count the separate reviewed column.
Reconcile the denominator before presenting a chart. The four reviewed category counts sum to eight, matching eight retained IDs. No row is dropped for being ambiguous. If a production workflow excludes duplicates or empty comments, record that rule and its affected IDs separately; otherwise a changed denominator may look like a changed customer trend.
An exact count still depends on the label definitions and reviewed decisions supplied to it. Nor does an exact eight-row total tell Omar how frequent an issue is among all customers. This constructed queue has no real customer sampling process, volume baseline, or measured classification reliability.

A completed prompt and version comparison
Valeria replaces the vague P1 instruction with P2, the explicit rule-based prompt shown earlier. To keep the comparison interpretable, input snapshot S1 remains unchanged. The following P2 outputs are again author-written simulations, not a regeneration performed in Sheets.
| Version and changed case | Prior proposal | Simulated P2 proposal | Reviewed disposition |
|---|---|---|---|
| P2 on S1, F4 mixed export comment | Request | Unclear | Keep Unclear; Omar still needs to decide whether to split the topics |
| P2 on S1, F5 empty download | Praise | Bug | Keep Bug as a reported failure, not a verified defect |
| P2 on S1, F6 “Buttons.” | Bug | Unclear | Keep Unclear; missing context remains missing |
| P2 on S1, F7 font question | Request | Request | Reject the proposed category; reviewed label remains Unclear under this queue's explicit rule |
| P2 on S1, F1, F2, F3, F8 | Bug, Request, Praise, Request | Same four proposals | Reviewed labels unchanged; these examples do not establish general agreement |
The complete P2 proposal set is two Bug, three Request, one Praise, and two Unclear. The human-reviewed summary remains two, two, one, and three respectively because F7's proposal is not accepted. F7 remains a visible conflict between P2's question rule and its simulated Request label; a longer prompt has not resolved that example.
Valeria records S1, P1, P2, the proposal columns, the reviewed decision, and the reason for each override. In a real approved run, she would also record the generation time and actual returned text. Here the generation-time field is “not run”; there is no actual Gemini response log to attach.
Refresh deliberately when inputs or rules change
A new output should not overwrite the evidence supporting an earlier report. Preserve the reviewed snapshot before a planned refresh, compare changed rows, and decide which reviewed labels need to be reopened. The prompt, input version, proposed output, and human decision can change independently.
Google's AI-function guide notes that changes to cells concatenated into a prompt do not produce a refresh prompt; the user must determine when to refresh. Do not treat a source edit as proof that existing generated text has recalculated. In this example, the supplied B2 range makes the intended per-row context clear, but the analyst still verifies the generated output's input version before reporting it.
Suppose the author of F6 later supplies the different sentence “Buttons disappear on a dark background.” That would create input snapshot S2 and require a new review for F6. The S1 summary must remain labeled S1; do not quietly replace the text while retaining the old Unclear decision and generation record. No S2 classification or count is claimed here.
For a larger queue, inspect ambiguous comments, unexpected labels, important categories, and a documented sample of the remainder. Record what was reviewed and what was not. This eight-row exercise cannot supply a production sampling threshold or an accuracy percentage; choose review coverage appropriate to the actual decision and its consequences.
Report the interpretation and its limits together
The completed fictional decision is to keep the reviewed S1 queue for Omar's triage discussion, with three Unclear comments visible and F7's proposed Request label overridden. Neither simulated proposal version is approved as an unattended classification system. Valeria retains the S1 record and unresolved review notes; Omar owns the next triage discussion, further product investigation, and any approved category-rule change.
Before using AI labels in a work report, retain the original feedback, category definitions, explicit input range, input and prompt versions, proposal text, human overrides, and count denominator. State whether a number counts reported concerns, reviewed interpretations, or independently verified events. Generate new proposals only within the organization's approved data-use process. More workplace guides are available through LandOffer.
Sources and Further Reading
- Google: Use the AI function in Google Sheets — current syntax, supplied context, generation and limitations; checked October 25, 2026.
- Google: Supported languages for Google Workspace with Gemini — feature-specific current language lists; checked October 25, 2026.
- Google Workspace Updates: AI function language expansion — September 23, 2025 historical announcement.
- Google: COUNTIF — conditional counting syntax; checked October 25, 2026.


