How to Choose a PR agency for Artificial intelligence (AI)

By Aidan Sowa, CEO and Founder, Frontier Industries | October 2026

In 1950, Alan Turing proposed a parlour game: could a machine hold a conversation well enough to fool a judge? Seventy five years later the machines pass, and the judges have changed. Your buyers no longer ask whether the model works in a demo. They ask who audited it, what it was trained on, and what happens when it is wrong. Seventy eight percent of organizations used AI in 2024, up from 55 percent a year earlier (Stanford AI Index, 2025). Adoption is not your problem any more. Scrutiny is.

Key numbers for this decision:

I. Discrimen

A launch announcement is not a credibility asset

Standfirst: Coverage tells the market you exist. It does not tell the market you are safe to buy.

Most AI companies brief agencies as if the contest were about attention. It is not. The contest is about permission: permission from a chief information officer to run your model on customer data, permission from a regulator to operate in a scored or advised decision, permission from a board to put your name in an annual report. Each of those permissions is granted by people who read differently from consumers. They read for failure modes.

That is why the first distinction matters. A generalist technology agency sells reach into publications. An AI literate agency sells fluency in the three languages your buyers actually speak: capability, governance, and liability. If a pitch deck cannot tell you how the agency would handle a story about your model producing a biased output, the deck is decoration. In AI, the crisis plan is the portfolio.

Aphorism for the wall: In this market, trust is the product and the model is the feature.

II. Machina

Your story is re-scored every time the benchmarks move

Standfirst: The press cycle in AI is driven by evaluation results, and evaluations decay in weeks.

Understand the machine you are hiring into. AI journalism runs on a benchmark calendar. When Stanford HAI published the 2025 AI Index, it recorded scores on demanding tests such as SWE-bench rising by more than 60 percentage points in a single year. A claim that made you look frontier in March can make you look mid-pack by September, and journalists keep the receipts.

Your agency therefore works inside a moving evidentiary frame. It needs people who can read a model card, interrogate a training data summary, and translate an evaluation table into a sentence a business editor will print. If the team cannot explain the difference between a benchmark result and a deployment result, they will write releases your technical buyers will quietly discount. Read more about how these guides frame evidence in the resources library.

There is a staffing consequence hiding inside that cycle. Your agency needs at least one person who follows the research conversation itself, the preprint servers, the major conference proceedings, the open source release notes, because that is where tomorrow trade headlines are drafted today. A team that learns about a capability shift from the same press release your competitors issued is a team that will always be writing the second day story. Second day stories are where narratives go to be summarised by other people.

The machine has a second gear: policy. The United States introduced 59 AI related federal regulations in 2024, more than double the prior year (Stanford AI Index, 2025), and the EU AI Act phases in obligations by risk class. News in this sector is written at the seam between capability and compliance, and your agency must be able to stand on that seam without wobbling.

III. Exemplum

The companies that win coverage bring receipts, not adjectives

Standfirst: Editors reward the vendor who publishes the evaluation, not the vendor who praises the model.

Watch how the credible end of this market behaves. The firms that hold share of voice across a full year are the ones that release reproducible evaluations, publish incident postmortems, and put named customers on the record with measurable outcomes. Their communications teams treat a limitations section as an asset, because procurement teams treat it as one too.

Contrast the pattern that fails. A vendor announces a breakthrough with no methodology, a friendly journalist repeats the claim, and a technical publication then spends a week dismantling it. The correction travels further than the launch. Your agency should be the party in the room who asks, before anyone else does, what the strongest hostile reading of your claim looks like. If they will not play that role in the pitch meeting, they will not play it in the newsroom either.

The lesson is structural, not moral: in AI, substantiation is a communications discipline before it is a legal one.

IV. Damnum

The cost of a careless claim is regulatory, not just reputational

Standfirst: In most sectors a bad headline fades. In AI it can be filed, cited, and priced.

Price the downside correctly. The Federal Trade Commission has warned companies that AI claims are subject to the same truth in advertising rules as any other claim, and that there is no AI exemption from those rules. The agency has already brought actions over deceptive AI marketing. A sentence your agency drafts to sound bold in a release can become exhibit material in an enforcement context.

Then there is the trust ledger. Only 32 percent of Americans say they trust AI, against 72 percent in China (Edelman Trust Barometer, 2025), and trust in technology companies among Americans has fallen from 73 percent a decade ago to 63 percent. Your buyers know their own customers are sceptical. When a chief information officer chooses you, they are spending their internal credibility, and they will choose the vendor whose public record looks boring under audit.

An agency that understands this prices caution into its drafting. An agency that does not will eventually make you the example in somebody else lecture.

V. Genus

The controlling variable is whether the agency can survive a technical cross-examination

Standfirst: Logos and media lists are table stakes. Technical defensibility is the differentiator.

When you score candidates, weight one variable above the rest: can this team defend your claims in front of a sceptical specialist audience? Everything else, relationships, creative range, price, matters after that gate. Test it directly. Give each finalist agency a real claim from your roadmap and ask them to write the version they would defend, the version they would soften, and the version they would refuse to publish. Their answers will tell you more than any case study.

None of these types is wrong in the abstract. Each is wrong for a specific mismatch. Your job in selection is to find the mismatch before the market does.

VI. Segmentum

A model lab, an AI feature, and an applied AI vendor need three different agencies

Standfirst: The category label on your website matters less than the scrutiny class of your buyer.

Segment your shortlist by the buyer who can end your quarter. A foundation model company sells to researchers, platforms, and governments. Its press lives in technical publications and policy media, and its proof burden is evaluation transparency. An applied AI vendor selling into hospitals, banks, or insurers sells to risk committees. Its proof burden is validation inside a regulated workflow, and its agency needs fluency in that vertical regulator, not just in AI.

A consumer AI product faces a third regime: app store economics, creator media, and sentiment that can turn on a single viral failure. The agency that excels in one regime will often underperform in another, because the journalists, the evidence, and the vocabulary differ. Ask every candidate which regime most of their current AI revenue sits in, and make them prove it with named coverage. The public relations services page outlines how a programme is scoped by regime rather than by tactic list.

VII. Occasio

The opening is in governance, because almost nobody communicates it well

Standfirst: Capability claims are crowded. Credible governance communication is still rare air.

Here is the opportunity your selection should exploit. The Stanford AI Index (2025) noted that AI incidents are rising while standardised responsible AI evaluations remain uncommon among major developers. In plain terms, the market is long on capability theatre and short on documented restraint. A company that communicates its evaluation practice, its data provenance, and its incident process clearly can occupy ground that bigger rivals have left empty.

There is a commercial reason to take the opening seriously, beyond virtue. Procurement questionnaires in enterprise AI now routinely ask for evaluation methodology, red teaming practice, and data handling documentation. Every public artefact your agency helps you publish, a model card summary, an evaluation explainer, an incident transparency note, becomes a document your sales team can attach to those questionnaires. Communications stops being a megaphone and becomes part of the evidence pack that shortens a security and risk review. Agencies that grasp this write differently: they draft for the auditor second and the headline first, and they keep a register of which public claims map to which internal proof.

Your agency should arrive with a point of view on this. Ask candidates how they would build you a governance narrative that a regulator could quote back at you approvingly. The strong answers will mention the NIST AI Risk Management Framework as a structuring device, because buyers and auditors already use its Govern, Map, Measure, and Manage functions as a shared vocabulary. The weak answers will offer you a thought leadership calendar. Choose accordingly.

VIII. Structura

A two person specialist team with senior time beats a famous logo with junior staffing

Standfirst: In AI communications, the byline on the pitch matters more than the brand on the building.

Structure beats size in this category for a practical reason. The journalists and analysts who decide AI narratives are few, technical, and allergic to being pitched by someone who learned the product that morning. A small agency where the person in your selection meeting is the person who calls the editor will outperform a large agency where your account is a training ground.

Interrogate structure in the selection process. Ask who exactly will draft technical material, who has relationships at the three publications your buyers actually read, and how many accounts those named people carry. Ask about the bench for an incident weekend, because AI incidents do not respect office hours. The answers are auditable. The logo is not. Background on how an industries focused team is organised is on the team page.

IX. Consilium

Run the selection like a procurement, with scored evidence and a paid pilot

Standfirst: Decide on artefacts, not chemistry. Chemistry is what the artefacts produce over time.

Use a scorecard and make every finalist produce the same artefacts: a claims audit of your current site, a mock response to a model failure scenario, and a 90 day narrative plan tied to your sales calendar. Score the artefacts blind if you can. The process feels slower for one week and saves you a year.

Close the process with a paid 60 day pilot scoped to one narrative. An agency confident in its AI practice will accept a pilot with clear success measures. An agency selling you a retainer fog will resist one.

X. Quaesita

Questions buyers ask before they sign

How much AI specialisation does an agency really need?

Enough that its senior staff can read your model card and your regulator latest guidance in the same afternoon. Below that threshold, you are paying tuition for their education with your reputation as the classroom.

Should an AI company hire a separate crisis firm?

Only if your incident risk is extreme, for example public facing systems with safety exposure. For most vendors, one agency with a rehearsed runbook and a named senior bench is faster than a handoff between two firms mid-incident.

What does credible AI coverage cost?

Retainers vary widely by market and scope, and any firm quoting before a claims audit is guessing. Expect a specialist programme to be priced like senior counsel, not like content production, because the scarce input is judgement.

How fast should we see results?

Narrative placement in trade and research press can land inside a quarter. Governance credibility compounds more slowly, across evaluation cycles and procurement references. Anyone promising category ownership in 90 days is selling you a slogan.

Can an agency help with analyst relations as well as media?

Yes, and in AI the two are converging, because analysts now interrogate evaluation practice the way journalists do. Ask whether the same senior staff would handle both, since split ownership produces split stories.

What should we prepare before briefing agencies?

A claims register, your evaluation summaries, your data provenance notes, and your incident history. Handing these over in selection is the fastest way to see which agency knows what to do with them.

Is thought leadership worth anything in AI now?

Only when it carries evidence. Opinion without data is the most abundant substance in this market, and editors price it accordingly, at zero.

How do we judge an agency work in the first quarter?

Agree the measures before the kickoff, in writing. Look for placement quality in the publications your buyers cite, the accuracy of how your claims survive editing, and whether analyst and procurement conversations start referencing your evidence unprompted. Clip volume alone will flatter a weak programme and indict a strong one, because governance stories are fewer and heavier than launch stories.

XI. Methodus

How this guide was built, and its limits

This guide rests on published third party research, chiefly the Stanford AI Index Report 2025, the McKinsey State of AI survey programme, the Edelman Trust Barometer 2025 technology sector findings, and the G2 Buyer Behavior Report 2024. Figures are quoted from those publications and from regulator material published by NIST, the European Commission, and the Federal Trade Commission.

The limits deserve stating. Survey based adoption figures depend on self reporting and vary by sample. Trust percentages shift by market and by question wording. Agency pricing is not standardised, so no figure here should be read as a quote. Where this guide generalises about agency types, individual firms will break the pattern in both directions, which is precisely why the selection artefacts in section IX exist.

XII. Incipe

Where to start

Frontier Industries builds public relations programmes for technology companies whose buyers audit before they buy, and its AI practice is staffed by people who can defend a claim in technical company. If you want a claims audit of your current narrative before you brief anyone else, start with a conversation: begin here or call 401-219-4207. Bring your hardest claim. It is the right first test.

Sources

1. Stanford Institute for Human Centered AI, AI Index Report 2025. https://aiindex.stanford.edu/report/ 2. Stanford HAI, 2025 AI Index release summary, BusinessWire. https://www.businesswire.com/news/home/20250407539812/en/Stanford-HAIs-2025-AI-Index-Reveals-Record-Growth-in-AI-Capabilities-Investment-and-Regulation 3. McKinsey and Company, The State of AI in Early 2024. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2024 4. Edelman, 2025 Trust Barometer, Technology Sector Insights. https://www.edelman.com/trust/2025/trust-barometer/report-tech-sector 5. National Institute of Standards and Technology, AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework 6. European Commission, Regulatory Framework for Artificial Intelligence (EU AI Act). https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai 7. World Economic Forum, Future of Jobs Report 2025. https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf 8. Gartner, Worldwide IT Spending Forecast, January 2025. https://www.gartner.com/en/newsroom/press-releases/2025-01-21-gartner-forecasts-worldwide-it-spending-to-grow-9-point-8-percent-in-2025 9. G2, Buyer Behavior Report 2024. https://research.g2.com/hubfs/2024-buyer-behavior-report.pdf 10. OECD, AI Policy Observatory and Principles. https://oecd.ai/en/ai-principles