Blog · Analysis · Last reviewed June 25, 2026

The Spreadsheet Becomes the Model Interface

Spreadsheets were already shadow models for institutions. AI does not merely make them easier to use. It makes the grid conversational, generative, and harder to govern unless the workbook keeps provenance, tests, and accountable review.

The critical line is accepted cell state: when a model-generated value, formula, cleanup, chart, summary, live AI function, or pasted AI output stops being advice and becomes part of the workbook that drives a decision.

The Grid Already Governed

The spreadsheet is one of the most powerful institutional interfaces ever built because it looks too ordinary to be treated as power.

A spreadsheet can be a budget, forecast, hiring plan, warehouse schedule, risk model, school roster, regulatory submission, grant tracker, clinical research table, campaign list, pricing tool, compliance register, or private government. It lets a person build a local world out of cells, formulas, filters, charts, macros, comments, colors, hidden tabs, copied assumptions, and small acts of judgment.

The file is not one thing. It is data store, calculation engine, user interface, work log, evidence artifact, and social contract. A high-impact workbook often says: these are the rows we recognize, these are the categories that count, these are the assumptions we accept, these are the people who may edit, and this is the number that will travel upward into the meeting, filing, invoice, roster, or decision.

That local world often becomes operational fact. A manager approves headcount because the workbook says the variance fits. A trader accepts a risk number because the sheet recalculated. A public agency allocates resources because the table ranked applicants. A nonprofit reports impact because the dashboard summarized rows. The spreadsheet does not merely represent the institution. It becomes one of the places where the institution thinks.

This is why AI in spreadsheets matters. It is tempting to treat Copilot in Excel or Gemini in Sheets as another productivity feature: help me write a formula, explain an error, make a chart, summarize rows, fill missing fields, or clean inconsistent data. Those features are useful. But the important shift is deeper. The office grid is becoming a model-mediated interface. The user no longer only writes formulas into cells. The user asks a language model to interpret the workbook, generate transformations, explain failures, and sometimes modify the structure of the sheet itself.

For this essay, an AI spreadsheet/model interface means any spreadsheet workflow where a model can read workbook context and produce something that may enter the workbook's decision path: a formula, generated value, filled column, classification, summary, lookup, chart, pivot, cleanup step, optimization, source pull, repair, or recommendation. The risk is not only that the model is wrong. The risk is that model output becomes ordinary cell state and then disappears into downstream formulas, reports, dashboards, filings, invoices, rosters, and managerial memory.

The spreadsheet was already a quiet decision engine. AI gives that engine a voice.

AI Enters the Cell

As of June 25, 2026, spreadsheet AI has several different surfaces. The first is the assistant around the workbook. Microsoft's current Copilot materials describe Excel support for formulas, explanations, summaries, insights, charts, PivotTables, formatting, filtering, and workbook edits. Microsoft 365 Copilot can also use content in Microsoft Graph, including work emails, chats, documents, meetings, and connected sources that the user is permitted to access. Microsoft says prompts, responses, and Graph data are not used to train foundation LLMs, but it also stores interaction data as Copilot activity history that can be managed through Microsoft 365 controls such as Purview. The governance question is therefore not only "can the feature read this workbook?" It is "which organizational context, permission boundary, retention rule, and audit path travels with the spreadsheet question?"

The second surface is direct workbook editing. Microsoft's support page for editing with Copilot in Excel describes a side-by-side experience that can build and edit workbooks with tables, charts, PivotTables, formulas, multi-step plans, model selection where available, and live workbook changes. The same support page also shows why mode boundaries matter: it describes source selection for web, work search, and federated connectors, while its other notes separately say editing with Copilot only works with the currently open workbook and that enterprise search or integration with external tools may not yet be supported in that editing path. A policy should not treat every Copilot-branded Excel surface as having the same read scope, write authority, or source record.

External data is now part of the spreadsheet AI surface. Microsoft Support describes federated connectors as a way for Copilot to retrieve information from external services in real time without importing or storing that source data in the workbook, using the user's credentials and requiring explicit consent before data is sent to an external service. In June 2026, Microsoft's Excel Blog announced federated Copilot connectors for LSEG and Moody's in Excel. That is useful for analysts who need timely market or credit context, but it also raises the source-chain standard: a generated workbook value may depend on a licensed external source, a user credential, a connector policy, a timestamp, and a consent event that ordinary cell history does not show.

Microsoft also separately documents Word, Excel, and PowerPoint Agents that can create files from descriptions, use connected content depending on license and policy, and save generated workbooks to OneDrive. Those are not just help features. They move AI from adviser to file author.

The third surface is the model inside the grid. Microsoft's support page for the COPILOT function describes a cell formula that constructs a prompt from text and referenced ranges, sends that prompt and grid context to an AI model hosted on Azure, and returns model-generated output directly in the workbook. The documented use cases are semantic and generative: summarizing text, creating sample data, classifying or tagging content, generating short text, and looking up web information. Microsoft also documents practical boundaries: the function is limited to qualifying licenses and release programs, currently uses gpt-4.1-mini (2025-04-14), only has access to the prompt, referenced ranges, and web data rather than other workbook or enterprise data, can return different results over time as the model evolves, is limited in how many COPILOT functions can calculate in a time window, and cannot calculate in workbooks labelled Confidential or Highly Confidential. Its web-lookup source references are listed as coming soon, which matters for source-critical uses.

Microsoft draws a sharper substantive boundary too. The same support page says COPILOT can give incorrect responses and advises users to avoid it for numerical calculations requiring accuracy or reproducibility, workbook lookups that native functions should handle, and tasks with legal, regulatory, or compliance implications.

That warning is not a footnote. It names the institutional tension. Excel is where organizations do exactly the kinds of work that require accuracy, reproducibility, legal defensibility, and compliance memory. The AI function is framed as exploratory, but it is placed inside a tool whose social role is often official.

Google is moving from the other side of the office stack. In September 2025, Google announced that Gemini in Sheets could provide natural-language formula explanations, explain formula errors, generate corrected formulas in follow-up turns, and offer multiple formula options for complex tasks. On April 22, 2026, Google announced broader Gemini in Sheets capabilities for building and editing entire spreadsheets from natural-language prompts, including data retrieval, formatting, formulas, pivot tables, charts, side-panel editing, Workspace Intelligence context, and plans shown for user approval. On June 22, 2026, Google announced one-click formula-error troubleshooting in Sheets, with Gemini analyzing surrounding data structure, explaining the issue, and suggesting a corrected formula.

Google's help materials for building or editing entire spreadsheets also show the governance boundary. The user is told to review Gemini's plan and template outline, can add or clear sources, and can stop generation. The same page warns that feedback may be human-readable and says not to submit personal, confidential, or sensitive information in feedback. In its Workspace Experiments section, Google says prompts, generated content, referenced Workspace content, and feedback may be stored or used under the experiments terms, and warns users not to rely on experimental features as medical, legal, financial, or other professional advice. The product surface is not just "make a sheet." It is a source-selection, approval, feedback, and data-handling workflow.

Google also said Gemini in Sheets had reached a 70.48 percent success rate on the full SpreadsheetBench dataset, a benchmark for complex real-world spreadsheet manipulation. SpreadsheetBench itself is useful context. The benchmark is built from real-world forum-style spreadsheet questions and workbooks with messy structures: missing headers, multiple tables in one sheet, multiple sheets, non-standard layouts, and tasks that require robust generalization. But leaderboard claims are live artifacts, not fixed product truth. On the public leaderboard reviewed June 25, 2026, the original SpreadsheetBench V1 full table showed Qingqiu Agent at 83.11 percent, WPS AI (Seed 2.0) at 73.46 percent, and Gemini in Google Sheets at 70.48 percent, while the V1 verified subset listed Data Analysis Agent at 96.50 percent. That movement is the source-discipline lesson: a dated benchmark claim can be true and still stop being the current top claim within weeks.

SpreadsheetBench 2 pushes further into end-to-end business spreadsheet workflows such as financial modeling, debugging, and visualization. Its overview listed a much lower top overall score, 34.89 percent, for those harder professional workflows, with financial-modeling, debugging, and visualization scores reported separately under a specific agent scaffold. That gap is the useful governance signal. A spreadsheet agent can look strong on a manipulation benchmark and still be fragile on multi-step financial models, repairs, charts, audit requirements, retention rules, and professional duties.

The direction is clear even where the scores are imperfect. The spreadsheet assistant is not only answering questions about a sheet. It is becoming a candidate operator of the sheet.

The Old Risk

AI does not enter a clean environment. Spreadsheets already have a long error history.

Raymond Panko's survey of spreadsheet error research concluded that spreadsheet errors are common and non-trivial, and that the only technique clearly demonstrated to reduce errors was cell-by-cell code inspection. Later work by Panko continued to emphasize a basic human-factors point: spreadsheet developers are often overconfident, and ordinary review practices miss errors.

The famous public cases matter because they show how a humble office artifact can sit under major institutional decisions. The U.S. Senate Permanent Subcommittee on Investigations report on JPMorgan Chase's 2012 "London Whale" losses described a value-at-risk model whose computation used spreadsheets and manual processes that were considered error-prone and not easily scalable. The report said key trading data were uploaded manually, spreadsheet-based calculations had insufficient controls, formula and code changes were frequent, and calculation errors lowered the VaR results. It was not a story about one bad cell. It was a story about governance, incentives, infrastructure, and model control.

Bank regulators have long understood the broader problem of model risk, though the current guidance must be read carefully. On April 17, 2026, the Federal Reserve, OCC, and FDIC issued SR 26-2, revised supervisory guidance on model risk management that supersedes SR 11-7 and SR 21-8. The guidance is expected to be most relevant to banking organizations with more than $30 billion in assets, emphasizes risk-based model governance, materiality, inventory, documentation, validation, monitoring, effective challenge, and vendor oversight, and says it is not a set of enforceable standards by itself. It also narrows its formal scope: simple arithmetic calculations such as those found within spreadsheets are excluded from the definition of model, and generative AI and agentic AI are described as novel and rapidly evolving systems outside the guidance's scope. That carveout should not be misread as safety. The same guidance says organizational risk management practices should guide governance and controls for tools, processes, or systems outside scope.

That is the practical lesson for AI spreadsheets. A workbook can sit outside a formal model inventory and still materially shape pricing, staffing, risk, research, compliance, or public-service decisions. If AI helps create or alter that workbook, validation and governance still matter even when the spreadsheet is not a "model" in one regulator's technical definition.

The AI spreadsheet inherits this old risk and adds a new layer. A formula can be inspected. A cell reference can be traced. A macro can be reviewed. A model-generated classification, summary, explanation, repair, or data transformation may be harder to reproduce, harder to validate, and easier to mistake for understanding because it arrives in fluent language.

The Promotion Boundary

The most important governance line is not whether a model helped somewhere near a spreadsheet. It is whether model output is promoted into the workbook's official logic.

There are at least five states. A side-panel explanation can remain advisory. A generated formula can be pasted into a cell and become deterministic workbook logic. A model-filled category or summary can become accepted cell state without any visible formula. A COPILOT-style cell can remain live, probabilistic, and model-dependent every time it calculates. A Gemini-style edit can restructure the sheet after the user approves a plan. Each state needs a different record. "AI helped with the spreadsheet" is too vague to govern.

Promotion is therefore a governance event, not a UI click. The record should show the state before and after acceptance, the model surface that produced the change, the source range or retrieved source, the tests or reconciliation checks applied, the reviewer, and whether the output remains live or has been frozen as an ordinary value.

A high-impact workbook should therefore separate draft AI assistance from accepted workbook state. If a generated formula, classification, filled value, chart, cleaned column, source pull, or summary enters a report, regulatory file, personnel process, customer decision, financial model, clinical table, or public-service workflow, the promotion event should be reviewable: what changed, which source ranges were used, whether external, Workspace, or Microsoft Graph context was pulled in, who accepted the change, what tests passed, and which downstream tabs or files now depend on it.

This is where audit tooling matters. Microsoft Purview documentation for Copilot and AI applications describes audit records that can include the host application, accessed resources, sensitivity-label identifiers, plugins, model transparency fields, prompt/response message references, and policy details. That does not automatically solve workbook governance, but it shows that relevant evidence fields can exist in the platform layer. Critical spreadsheet policy should connect application logs, workbook version history, source ranges, and human sign-off instead of treating the final grid as self-explanatory.

The record also has to name the mode. A side-panel answer, a direct workbook edit, a live model-backed cell, a file-generating agent, a web lookup, and a connector-assisted retrieval are different delegations. If the audit record says only that "Copilot" or "Gemini" was used, the institution may know the brand while losing the authority boundary.

From Formula to Judgment

The spreadsheet's traditional power came from a promise of computability. If the assumptions are visible and the formulas are correct, the result follows. That promise was always partial because workbooks contain hidden assumptions, copied errors, stale links, ambiguous categories, and human choices disguised as arithmetic. Still, formulas gave the grid a kind of mechanical accountability.

AI changes the unit of work. Instead of "calculate this number from these cells," the user can ask: classify this feedback, identify outliers, explain this trend, generate an executive summary, clean this column, infer categories, suggest next steps, build a budget, repair the model, produce a chart, or tell me what matters.

Those are not only spreadsheet operations. They are acts of judgment. Classification decides what kind of thing a row is. Summarization decides what can be ignored. Outlier detection decides what deserves attention. Data cleaning decides which messiness is error and which is evidence. Chart generation decides what shape the story takes. Formula repair decides which logic counts as intended.

This does not make AI assistance illegitimate. It may help many users understand formulas they previously copied blindly. It may reduce some spreadsheet errors by explaining broken references, mismatched ranges, text-formatted dates, and missing assumptions. It may make spreadsheet work less dependent on one local power user who knows all the hidden formulas.

But it also creates a new authority gradient. The model can become the senior analyst inside the workbook. It can speak with confidence where the user has uncertainty. It can make a suggested formula look more legitimate because it explains itself. It can produce a chart that feels like insight before anyone has checked whether the data, labels, filters, and assumptions deserve that presentation. That is a spreadsheet version of automation bias: users may over-trust a fluent system precisely where they most need to inspect the logic.

Shadow Model Governance

Many institutions already struggle with end-user computing: business units building critical tools in spreadsheets, scripts, notebooks, local databases, or SaaS automations outside ordinary software governance. These tools exist because official systems are too slow, too rigid, too expensive, or too distant from operational knowledge. Shadow tools are often where real work happens.

AI makes shadow model governance harder because it lowers the cost of building and modifying these tools. A person who cannot write a complex formula may be able to ask for one. A person who cannot debug a workbook may ask the assistant to repair it. A team that would have waited for an analyst may generate a working model inside a shared spreadsheet. The barrier to action falls before the institution has updated its inventory, review, or approval process.

This is not only a security issue. It is a knowledge-governance issue. A workbook can encode business logic that no official system contains. When AI helps create, alter, or explain that workbook, the institution needs to know what changed, why it changed, what source material was used, who approved it, whether the output is reproducible, and whether the AI-generated part is still appropriate after data, policy, or market conditions shift.

The related workplace problem is shadow AI: institutional work routed through tools, accounts, models, or task paths that the organization has not inventoried or governed. AI spreadsheets can be sanctioned and still behave like shadow AI when the workbook, feature, connector, or task use is outside the approved data boundary. The site's essay on enterprise connector permission maps makes the same point from the opposite direction: the AI surface is only as governable as the map of what it can read, write, retrieve, and remember.

The earlier essay The AI Register Becomes Public Memory argued that organizations need inventories of AI systems before accountability can begin. AI spreadsheets show why the inventory problem is difficult. The system may not look like an AI product. It may look like the same shared workbook the team has always used, now with AI-generated formulas, AI-filled fields, AI-created summaries, and an assistant sitting in the side panel.

Failure Modes

The first failure mode is semantic leakage into official numbers. A model-generated classification, summary, or lookup quietly feeds downstream calculations, reports, dashboards, or decisions as if it were a deterministic value.

The second is explanation as false assurance. A formula explanation sounds coherent, so the user trusts the formula. The explanation may describe a plausible intent without proving that the formula matches the business rule, data shape, or edge cases.

The third is repair without responsibility. An assistant fixes a broken workbook, but the institution does not record what changed, what assumption was selected, what alternative was rejected, or who owns the repaired logic.

The fourth is benchmark overreach. A spreadsheet agent performs well on a public dataset, so users infer readiness for financial reporting, regulatory submissions, clinical research, payroll, benefits, or other contexts where the real risk is not only task completion but auditability, source permission, business-rule fit, and reproducibility.

The fifth is provenance collapse. A generated value, copied formula, cleaned column, or AI-written summary is pasted into ordinary cells without retaining the prompt class, source range, model surface, reviewer, or reason for acceptance.

The sixth is shadow automation. AI lowers the skill threshold for building business-critical spreadsheets faster than governance can discover them. The organization gains local productivity and loses institutional memory.

The seventh is data-boundary confusion. A spreadsheet may contain customer records, employee data, protected health information, confidential projections, or regulated financial information. Users need to understand which AI features can read which cells, which files, which organizational data, which web sources, and what labels or policies block use.

The eighth is connector spillover. A workbook question may cause the assistant to retrieve context from files, emails, chats, saved sheets, web results, or approved external connectors that the user is technically allowed to access but did not intend to mix into the workbook's source chain. A one-time source approval in a chat session can become invisible once the result is pasted into the grid.

The ninth is workslop in grid form. A plausible dashboard, clean table, or polished executive summary saves the sender time while shifting verification, correction, and trust repair downstream. That is the spreadsheet version of workslop and the trust tax.

The tenth is human de-skilling. If the assistant becomes the usual route to formula design, debugging, and analysis, users may become less able to inspect the logic that governs their own work. Convenience can hollow out the very competence needed for oversight.

The eleventh is live-output drift. A model-backed cell or assistant-generated edit may depend on model version, feature rollout, retrieval behavior, usage limits, policy blocks, or web context. If the workbook later recalculates or is reopened under different conditions, the institution may not be able to reproduce the output that supported an earlier decision.

The twelfth is audit mismatch. Platform logs may know that an AI interaction occurred, while the workbook itself may not show which cells changed or why. The compliance record and the spreadsheet record have to be joined before either one can explain the decision.

The thirteenth is record substitution. A workbook created or repaired by an assistant may become the official record even though the source chain, prompts, rejected options, generated assumptions, and reviewer judgment live somewhere else or nowhere at all.

The Governance Standard

A serious AI spreadsheet policy should begin by treating high-impact workbooks as models, not as casual documents.

First, classify critical spreadsheets. Any workbook that affects money movement, risk reporting, hiring, benefits, legal compliance, healthcare, research findings, public services, or customer treatment should have an owner, purpose, version history, review cadence, and risk rating.

Second, distinguish AI surfaces. A side-panel assistant, a cell function, an agent that can edit the workbook, a file-creation agent, a connector to organizational data, and a web lookup are different control problems. Policy should name which surfaces are allowed for which data classes and task classes, and whether the output may remain live or must be converted into reviewed values.

Third, separate exploratory AI from official logic. AI can help draft, explain, prototype, and debug, but official calculations and regulated outputs should use reproducible formulas, documented transformations, and reviewable code paths.

Fourth, record AI-assisted changes. When an assistant generates or repairs formulas, fills fields, changes structure, imports context, or creates summaries used downstream, the workbook should preserve what changed, by whom, with what prompt or instruction class, which model surface was used, which source ranges, source approvals, connector names, or external sources were used, and under what review. The practical pattern is an agent-log receipt scaled down to the workbook.

Fifth, validate against business rules, not only examples. A formula that works on visible rows may fail on edge cases. Review should include test cases, range checks, source reconciliation, hidden-sheet inspection, sign-off from someone who understands the operational domain, and a record of known limitations.

Sixth, preserve cell provenance. Generated cells, pasted AI outputs, and converted values should be distinguishable from directly sourced data and deterministic formulas. If that is not technically possible in the application, critical workbooks need a change log, protected notes, separate generated-output tabs, or external review records.

Seventh, gate promotion into official use. A generated or repaired cell should not become official simply because it is visible in the grid. Critical workbooks need a promotion step: reviewer, reason, tests, source ranges, affected downstream formulas, rollback path, and a rule for whether live model-backed cells may recalculate after acceptance.

Eighth, restrict high-stakes AI functions by label and context. Confidential, regulated, or compliance-sensitive workbooks should not allow casual generative calls simply because the feature exists in the application. Sensitivity labels, data-loss rules, connector permissions, source-selection controls, and approved-use policies should reinforce each other.

Ninth, require source discipline. A model-generated summary should not become evidence unless the source range, source file, retrieval path, or external reference is visible enough for another person to check. For source-critical work, the pattern should look more like a data sheet than a chat bubble.

Tenth, preserve user competence. AI explanations should be used to teach inspection, not replace it. A user who cannot explain a critical workbook should not be the only approver of AI changes to it.

Eleventh, inventory the office layer. AI governance that only tracks purchased AI products will miss the everyday grid where decisions are prepared. Registers, audits, and model-risk programs need a path for AI-enabled spreadsheets, not only formal machine-learning systems. The related control is an AI audit trail that follows consequential outputs into the systems that later rely on them.

Twelfth, test the human workflow. A spreadsheet AI review is incomplete if it tests only the model. It should test whether users notice wrong formulas, reject weak summaries, inspect sources, understand sensitivity labels, and know when to escalate. That connects workbook governance to human oversight, not only model scoring.

Thirteenth, archive the decision version. If a workbook supports an invoice, filing, grant score, risk limit, hiring list, research conclusion, or public-service decision, the organization should preserve the exact accepted workbook, source extracts, AI-assisted change record, reviewer sign-off, and calculation state. A live AI cell, unresolved external query, or editable draft is not an archival record.

Fourteenth, audit the assurance claim. When a team says an AI spreadsheet is "reviewed," "validated," or "audited," the claim should name the scope, evidence, tests, reviewer independence, limitations, and consequences. Otherwise spreadsheet governance becomes a miniature version of AI audit compliance theater.

Fifteenth, keep a workbook manifest for critical files. High-impact AI-enabled workbooks should have a compact manifest: owner, purpose, data sources, connectors, sensitivity labels, AI features enabled, live model-backed cells, accepted generated ranges, validation tests, review status, archive rule, and known limitations. That connects spreadsheet governance to AI data provenance, data cascades, and model drift.

What This Changes

The spreadsheet is a small simulation machine. It turns a workplace into rows and columns, lets assumptions propagate, and gives the result the authority of recalculation. It is one of the places where recursive reality became ordinary before anyone called it that.

AI does not replace the spreadsheet's older power. It amplifies it. The grid becomes conversational. The formulas explain themselves. The messy table asks to be cleaned. The dashboard narrates its own significance. The assistant turns partial records into fluent managerial language. A local model of the world becomes easier to create, easier to believe, and harder to inspect at the moment of use.

The danger is not that every AI spreadsheet will be wrong. The danger is that wrongness becomes socially smoother. The workbook no longer fails with an obvious error code. It produces a plausible category, a plausible summary, a plausible chart, a plausible repair, a plausible recommendation. The institution sees less friction and calls it intelligence.

The better path is not to ban intelligence from the grid. It is to remember what the grid already is: an interface where representation becomes decision. AI can help users understand and govern that interface, but only if the institution refuses to let fluency substitute for validation.

The spreadsheet became powerful because it let ordinary workers build models without asking permission. That democratic force is real. So is the risk. In the AI office, the question is no longer only who can edit the cells. It is who can see when the model inside the cells has begun editing the institution back.

Source Discipline

The sources for this essay should be read by type. Microsoft and Google product documentation establish what the tools say they can do, what limits they disclose, and what data-handling claims they make. Those vendor documents are not proof that a given deployment is safe, complete, or properly configured.

The workbook itself needs the same discipline. A generated formula explanation, chart title, dashboard summary, or model-written note should not be cited as the source of a claim. For consequential work, the source is the workbook version, source tab, source range, external file, connector result, query, formula, model interaction record, and human review that produced the accepted cell state.

SpreadsheetBench is useful as a task benchmark, not as a governance certificate. Its leaderboard is also a moving public artifact: scores, entrants, verified status, benchmark versions, and product-specific agent scaffolds can change faster than institutional policy cycles. The original SpreadsheetBench manipulation benchmark and SpreadsheetBench 2's end-to-end workflow benchmark should not be blended into one score. A benchmark score does not answer whether a workbook is auditable, whether source data was appropriate, whether a generated formula matches a local business rule, or whether a regulated decision can be defended after review. The Panko paper and Senate report show the older spreadsheet-error problem; they do not by themselves measure modern AI features.

SR 26-2 and the NIST AI Risk Management Framework supply governance vocabulary, not a one-size rule for every office workbook. The practical standard is therefore source discipline: feature claims from vendors, benchmark claims from benchmark maintainers, error claims from spreadsheet research, and governance claims from regulators or standards bodies should not be collapsed into one general story of "AI is good at spreadsheets." Internal links below provide site vocabulary and continuity; they are not independent evidence for external product, benchmark, legal, or regulatory claims.

Sources

Return to Blog