Skip to main content
Everything PR News
Digital PR & Communications

What AI Can't Do for a Communications Team (and How to Judge It)

EPR Editorial Team17 min read
Share
an overview of uncategorized content

A communications team cannot hand six jobs to AI: judgment calls, accountability for statements, source verification, confidential material, relationship work and live crisis decisions. About 82% of journalists now use at least one AI tool, according to Muck Rack's State of Journalism 2026, so the limits matter more each quarter. This page ties each limit to a documented case and ends with a quarterly checklist for judging what AI does well.

Originally published June 15, 2026. Updated October 13, 2026.

Key Takeaways

  • AI drafts, summarizes and searches quickly, but six jobs stay with people: judgment, accountability, verification, confidentiality, relationships and live crisis decisions.
  • The organization that publishes a statement is responsible for it. A British Columbia tribunal ruled in 2024 that an airline answered for what its website chatbot told a customer.
  • Fluent text is not verified text. A US federal judge sanctioned lawyers in 2023 for AI-invented case citations, and a Stanford study found retrieval-based legal research tools still erred in at least one of every six benchmark queries.
  • Confidential and embargoed material belongs only in tools whose retention, training and access terms the team has approved.
  • Judge AI on four outcomes: output quality, time saved, AI visibility and control adherence. Prompts run and seats licensed measure activity, not value.
  • Review the four outcomes every quarter with a written checklist, and end each review with a keep, restrict or retire decision for every tool.

What are the six limits of AI for a communications team?

The six limits are judgment, accountability, verification, confidentiality, relationships and live crisis decisions, and each one has a public record behind it. The table below pairs every limit with a documented case and the control that holds up in practice. The sections that follow explain each pairing.

LimitDocumented caseControl that holds
Judgment callsGoogle's May 2024 AI Overviews update, which addressed satire read as factA named person signs off audience, timing and trade-offs
Accountability for statementsMoffatt v. Air Canada, 2024 BCCRT 149One named approver for every external statement, including chatbot content
Source verificationMata v. Avianca sanctions, 2023, and the Stanford legal research studyEvery figure, quote and name carries a primary-source link and a named verifier
Confidential materialSamsung's May 2023 restriction on generative AI tools after a code uploadMaterial tiers matched to approved tools
Relationship workMuck Rack's State of Journalism 2026 survey data on pitch relevanceA person owns every journalist and client relationship
Live crisis decisionsApple's January 2025 pause of AI news notification summariesA pre-approved decision tree with named decision makers

Can AI make the judgment calls in a communications plan?

No. AI tools can propose options, but a person has to judge which option fits the audience, the timing and the context. That judgment is the part of the work that clients, executives and journalists hold a communications team responsible for.

Google's own account of its May 2024 AI Overviews rollout shows why context is hard for a model. Liz Reid, Google's head of Search, wrote in the post "AI Overviews: About last week" that the system had struggled with nonsensical queries and satirical content. Reid described a data void, meaning a topic with little high quality material, where a satirical page was the main source available. Google responded with better detection of nonsensical queries and limits on satire and humor content. Google published its explanation and described the changes it made.

The mechanism matters for a communications team. A language model predicts likely text from patterns, and a joke, a quote and a sincere recommendation can look alike once they are text. A person reads the room: whether a statement should go out today or after the hearing, whether a phrase invites the regulatory question the legal team is worried about, whether a competitor's announcement makes this the wrong week.

The practical rule is to name three decisions that a person signs off before anything leaves the building. The first is the audience and what that audience needs to hear. The second is the timing. The third is the trade-off the team accepts, such as speed against accuracy or reach against risk. AI can draft three versions of a message and pressure-test each one. A named person picks one.

a team working together on a strategy without ai assistance

Who is accountable when an AI-generated statement is wrong?

The organization that publishes the statement is accountable, whether a person or a tool wrote it. A 2024 tribunal ruling in British Columbia made that point about a chatbot, and the reasoning applies to any AI-generated text a team puts in front of an audience.

In Moffatt v. Air Canada, 2024 BCCRT 149, issued February 14, 2024, customer Jake Moffatt asked the airline's website chatbot about bereavement fares after a death in the family. The chatbot said a fare could be claimed retroactively, while the airline's policy page said it could not. The Civil Resolution Tribunal found that the airline did not take reasonable care to ensure its chatbot was accurate, and it rejected the argument that the chatbot was a separate entity responsible for its own actions. According to a legal summary on Mondaq, the airline was ordered to pay a total of $812.02.

The amount is small and the principle is not. The tribunal treated the chatbot as part of the airline's website, which means every FAQ, chatbot answer, social reply and AI-drafted release carries the publisher's name. A communications team should assign one named approver for every external statement, record which tool produced the first draft, and keep a log of what any customer-facing bot has said. The log turns a dispute from memory against memory into a review of a record.

Can AI verify its own sources?

No. A generative tool can produce a fluent citation, quote or figure that does not exist, so a person has to check each factual claim against a primary source. Two documented cases show the pattern in law, and the same pattern applies to a press release.

On June 22, 2023, Judge P. Kevin Castel of the US District Court for the Southern District of New York fined two lawyers and their firm $5,000 in Mata v. Avianca. Their brief cited cases that ChatGPT had invented, and the lawyers continued to stand by the cases after the court questioned them. The judge wrote that the lawyers abandoned their responsibilities, and the court record shows the failure was one of checking, not of using a tool.

The second case is a measurement. Stanford RegLab and Stanford HAI researchers ran a preregistered test of AI legal research tools that use retrieval, a method that pulls in real documents before the model answers. The tools improved on a general-purpose chatbot, and the researchers still found hallucination rates of at least one in six across the benchmark queries. Retrieval lowers the error rate and does not remove the need for a human check.

In communications work, the equivalent failure is an invented statistic, a misattributed executive quote or a wrong date in a release. The control is a verification rule that is easy to audit: no draft is approved until every figure, quote, name and date has a primary-source link and a named verifier. Teams that need a worked example of the full workflow can see the six stages in the AI PR stack workflow.

A verification log needs five fields: the claim, the link to the primary source, the date the source was checked, the name of the verifier and the approval status. The log adds minutes to each draft and gives the editor, the client and any later reviewer a record of what was checked and by whom.

Is confidential client material safe in an AI tool?

Confidential material is safe only in a tool whose retention, training and access terms the team and the client have approved for that class of material. A consumer chat account the team does not control should be treated as outside the organization.

Samsung offers the clearest public example. Bloomberg reported on May 2, 2023 that the company restricted employee use of generative AI tools after staff uploaded sensitive code. Fortune's coverage of the memo says Samsung described the restriction as temporary and committed to building its own tools for translation and document summaries. The company's response was the one a communications team should copy: it set a rule, explained the security concern and worked toward an approved environment.

A tiered rule keeps the policy short enough to follow. Each material tier maps to the places it may go.

Material tierExamplesWhere it may go
PublicPublished releases, public filings, posted dataAny approved drafting or research tool
InternalDraft strategy, internal metrics, unpublished plansBusiness accounts with approved retention and no-training terms
RestrictedEmbargoed news, deal terms, unreleased financials, non-public crisis facts, personal contact dataOnly tools approved in writing by the client and legal, otherwise no AI

Vendor terms differ and change, so the team should read them before assigning a tier. The procurement questions on AI tools for communications teams cover training use, retention, security terms and exit, and the sector pages on healthcare communications show how a regulated field applies the tiers.

Can AI do the relationship work with journalists?

No. A relationship rests on trust built between people over time, and journalist survey data shows that relevance, not volume, decides whether a pitch is read. This limit is documented by survey data rather than by a single incident.

Muck Rack's State of Journalism 2026 report surveyed 1,044 journalists between January 30 and March 2, 2026, with 897 responses in the final analysis. It found that 86% of journalists say PR pitches inspire at least some of their stories, and that 88% immediately disregard pitches that miss their beat. Nearly half said they seldom receive pitches that match their coverage. The same report found that 82% of journalists use at least one AI tool, up from 77% the year before.

The reading for a communications team is that pitches matter and fit decides the outcome. AI is well suited to the research behind fit: sorting reporters by beat and recent coverage, summarizing what a reporter has written, drafting a first version. A person decides whom to call, what the reporter would want to know, and when to follow up. The reporter relationship stays with the named person whose email the reporter opens.

Can AI make live crisis decisions?

No. A live crisis requires decisions about facts that are still changing, legal exposure and who speaks, and a named person has to own each decision. AI can monitor, summarize and draft for review, and it cannot carry the consequences of a wrong call.

Apple's 2025 experience with AI notification summaries shows what happens when automated text meets a developing news event. In December 2024 the BBC complained that an Apple Intelligence summary falsely said a murder suspect had shot himself. As CNN reported on January 17, 2025, Apple paused the notification summaries for news and entertainment apps in beta software, said it would make improvements in a future update and added clearer labeling that the summaries come from AI. The BBC said it was pleased Apple had listened. The pause was a human decision to stop an automated process, which is the kind of call a crisis plan has to assign to a named person.

A communications team can prepare for the same pressure before a crisis starts. A pre-approved decision tree names who decides whether to speak, who approves the first statement, who checks facts and who has legal sign-off. The team can then use AI inside that tree for monitoring, a first holding-statement draft and a summary of coverage, as described on AI tools for crisis communications teams. Every sentence that goes out still has a named approver.

How should a communications team judge what AI does well?

A communications team should judge AI on four outcomes: output quality, time saved, AI visibility and control adherence. Activity measures such as prompts run, seats licensed or drafts produced show that a tool is used, not that it helps.

Each outcome needs a baseline taken before the tool arrives, a named owner and a review date. Without a baseline, a team can only say that work feels faster. With one, the team can show that a tool saved a measured number of hours or that it added revision rounds.

OutcomeQuestion it answersTypical ownerEvidence
Output qualityIs AI-assisted work accurate, on voice and approved without extra rounds?Editor or workflow ownerVerification log, revision count, error log
Time savedDo net hours fall once checking and editing are counted?Team leadTimed samples before and after
AI visibilityDo AI engines name the brand or client, and describe it correctly?Visibility ownerFixed prompt set across five engines
Control adherenceDid the team follow the tool, confidentiality and disclosure rules?Governance ownerSampled audit of recent work

Activity measures fail for a specific reason: they rise when a tool is used badly. A team that runs more prompts because drafts keep failing looks busier, and the number says nothing about whether the work improved. Outcome measures fall when a tool is used badly, which makes them the safer guide.

How do you score AI output quality?

Output quality is scored with four measures: the verified-claim rate, revision rounds, voice fit and escaped errors. Together they show whether AI-assisted work reaches approval accurately and without extra effort.

MeasureHow to calculate itReview rule
Verified-claim rateFactual claims with a primary-source link and a named verifier, divided by all factual claims in the draftAnything below 100% blocks approval
Revision roundsEdit rounds from first draft to approval, compared with the same task done without AIIf AI drafts need more rounds, revise the prompt or drop the use
Voice fitAn editor scores samples from 1 to 5 against a fixed brand voice rubricScores below 4 trigger a prompt or style guide change
Escaped errorsErrors found after publication, marked as traced to AI output or notAny escaped error triggers a review within a week and a change to the checking step

The scoring scale and thresholds are starting points for a team to adjust, not industry standards. What matters is that the team fixes them before the quarter begins, so a result cannot be argued after the fact.

How do you measure time saved by AI?

Time saved is the baseline hours for a task minus the hours the AI-assisted version takes, with checking and editing counted as part of the AI-assisted version. A tool that drafts in two minutes and needs an hour of correction has not saved time.

Pick three to five repeatable tasks, such as building a media list, drafting a release, summarizing a week of coverage or preparing a briefing. Time each task without AI across a small sample, then time it with AI across the same number of samples. Record draft time, verification time and revision time separately.

Illustrative scenario: a release draft takes 90 minutes without AI. With AI, drafting takes 20 minutes, verification 25 minutes and revision 15 minutes, which totals 60 minutes. The net saving is 30 minutes, or one third. The figures are invented to show the arithmetic, and a team should replace them with its own timed samples. The full hours-saved calculation, including a sample, is on EPR's guide to measuring ROI on AI tools.

How do you track AI visibility?

AI visibility is tracked by running a fixed set of buyer questions through the main AI engines on a schedule and recording which brands each answer names. The result shows whether a client appears when buyers research the category, and how the answer describes the client.

EPR Research's Citation Share Index defines Citation Share as the share of AI-generated answers in which a brand appears. Its method runs buyer-intent prompts across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, and it reports figures as modeled estimates because AI engine logs are not public. A team tracking its own brand can use a fixed set of 30 to 60 buyer questions, run it monthly and divide the answers that name the brand by all answers.

The qualitative read matters as much as the count. Check whether the description is accurate, whether a competitor is named first and whether the framing matches what the team would have chosen. A rise in Citation Share with an inaccurate description is a problem to fix, not a win to report. The layers that produce this result are laid out in the AI Communications Stack.

How do you check that AI controls are followed?

Control adherence is checked by sampling recent AI-assisted work and testing it against the written rules for tools, confidentiality, verification and disclosure. A policy that is never audited is a document, not a control.

Each quarter, pull a sample of finished work, for example ten pieces across different team members. For each piece, confirm the draft was made in an approved tool and account, that the material tier matched the tool, that a verifier is named and the primary-source links are present, and that any required disclosure to the client appears. Record misses by cause, such as an unclear rule, a missing tool or a skipped step, because the cause decides the fix.

What belongs on a quarterly AI review checklist?

A quarterly AI review checklist has one section for each of the four outcomes and a closing decision section. The checklist below can be copied into a team document and completed with evidence, not impressions.

AreaQuestion for the quarterEvidence to pullAction if the answer is no
Output qualityDid every approved draft reach 100% verified claims?Verification logTighten the approval gate and retrain the team
Output qualityDid AI drafts need fewer or equal revision rounds than human drafts?Revision counts by taskRevise prompts or limit the use case
Output qualityDid voice-fit scores stay at 4 or above?Editor score sheetUpdate the style guide and prompt library
Output qualityWere there escaped errors, and were they traced?Error logAdd a checking step at the point of failure
Time savedDid net hours fall on each tracked task?Timed samplesRetire the use or redesign the workflow
Time savedIs verification time shrinking as prompts improve?Verification time by quarterImprove source supply in the prompt
AI visibilityDid the monthly prompt set run on schedule across all five engines?Run logAssign an owner and fix the schedule
AI visibilityIs the brand named, and is the description accurate?Answer archiveCorrect source pages and published facts
AI visibilityIs any competitor named first on key questions?Prompt results by enginePlan content and placements for those questions
Control adherenceWere all sampled pieces made in approved tools?Account and tool recordsClose the gap with a seat or a rule
Control adherenceDid material tiers match the tool used?Sampled auditRetrain and restrict the tool
Control adherenceWas required disclosure to clients made?Client filesAdd disclosure to the approval step
Vendor termsHave training, retention and exit terms changed since last review?Current vendor termsReassign the tier or switch the tool
DecisionDoes each tool get a keep, restrict or retire decision?Review notesDo not close the review until decided

What do you do with the quarterly results?

Each quarterly review ends with one decision per tool: keep, restrict or retire. The decision follows the evidence in the checklist, and it is recorded with a date and an owner so the next review can test it.

DecisionTriggerAction
KeepNet hours saved, quality at target, no control missesContinue, and raise the target for next quarter
RestrictQuality or control misses in one use case but gains in othersLimit to the use cases that passed and retrain on the rest
RetireNet hours do not fall, or errors escape despite fixesRemove the seat and move the budget to a tool that passed

Share the result with the people who use the tools. Teams follow rules they can see working, and a short summary of what was kept, restricted or retired makes the next review easier.

Conclusion

Every limit on this page comes back to one fact: a person's name sits on the statement. Judge AI by whether it makes that person's work faster, more accurate and easier to audit, and review the evidence every quarter.

Frequently Asked Questions

What can't AI do for a communications team?

AI cannot take responsibility for a statement, verify its own sources, hold a confidence, build a reporter relationship, read timing and context or make a live crisis decision. It can draft, summarize, search and organize, and a named person decides what goes out.

Can AI replace a PR agency?

No. AI replaces parts of the production work inside an agency, such as first drafts and research summaries. The relationships, judgment, strategy and accountability that clients hire an agency for are not tasks a tool performs.

What is the biggest risk of using AI in communications?

Unverified specifics. A model can produce a confident statistic, source or quote that does not exist, and the publisher answers for it. Every factual claim should carry a primary-source link and a named verifier before approval.

Should communications teams use AI tools at all?

Yes, for production work inside written rules. Muck Rack's State of Journalism 2026 reports that 82% of journalists use AI, so a team that avoids the tools still operates in a field where the other side uses them. The question is where the tools sit in the workflow, not whether to use them.

How do you measure whether AI is working?

Measure outcomes instead of activity: output quality, net time saved, AI visibility and control adherence. Take a baseline before the tool arrives, review each outcome every quarter, and decide to keep, restrict or retire each tool.

How often should a team review its AI tools?

Quarterly is a workable cycle. Run the AI visibility prompt set monthly, because answers from AI engines change more often than a quarterly review catches, and use the quarterly review for quality, time, controls and vendor terms.

EPR Editorial Team
Written by
EPR Editorial Team

The Everything-PR Editorial Team is the staff byline for news, analysis and features on public relations, crisis communications, reputation, AI visibility and digital discovery. Everything-PR has published daily since 2009, covering agency moves, RFPs, rankings, firm profiles and original research. Every article is reviewed by a human editor before publication. Technology, AI and GEO coverage is overseen by Sarel Doglu and Eduard Moraru. Coverage follows the Editorial Policy, and substantive corrections are noted on the article under the Corrections Policy.

Related reading

Other news

See all

Most brands are invisible inside AI search. Is yours?

EPR publishes the data every week.

Free. Weekly. Unsubscribe anytime.