For the past three years, corporate conversations about AI have mostly revolved around adoption. Which tools should we use? Which teams should use them? Are competitors moving faster? How much should we invest?
Stanford’s 2026 AI Index suggests that, for many companies, adoption is already well underway. In 2025, 88% of surveyed organisations said they were using AI in at least one business function, compared with 78% a year earlier. More than half were already using it across three or more functions. In consumer goods and retail, marketing and sales are among the areas where usage is particularly high.
The report is full of numbers showing just how quickly the technology is improving and spreading. But reading through its 400-plus pages, another issue keeps appearing: the infrastructure surrounding AI is developing much more slowly.
Stanford itself describes a growing gap between what AI systems can do and our ability to manage them. Governance frameworks, evaluation methods and organisational systems are struggling to keep pace with the technology.
For companies, and particularly for brands, this may prove more important than another improvement in model performance.
AI is giving companies a lot more capacity
The productivity gains are becoming difficult to dismiss.
Research collected by Stanford shows customer-support teams resolving roughly 14% to 15% more issues per hour with AI assistance. Developers using GitHub Copilot completed 26% more pull requests in one study. Marketing teams using multimodal AI to produce advertising increased output per employee by 50%.
Companies are also beginning to associate AI with business results. Among respondents surveyed by McKinsey and included in the Stanford report, 67% linked AI use in marketing and sales with revenue gains, while 64% said AI had improved innovation.
In marketing, the attraction is obvious. Producing ten versions of a campaign is easier. Local teams can create material without waiting several weeks for central production. Agencies can explore far more creative routes. CRM, social content, product descriptions, presentations and advertising can all be produced in greater quantities and much faster.
None of this is particularly controversial anymore.
What companies have spent less time thinking about is what happens when the amount of material being produced increases much faster than their ability to review it.
A brand producing five times as much content has also created five times as many opportunities for something to be slightly wrong.
And “wrong” is rarely limited to spelling mistakes or obviously false information. A piece of communication can be perfectly written and still feel completely unlike the brand. A campaign can respect every visual guideline and undermine the positioning. An influencer may deliver excellent reach and still be the wrong association. A local adaptation may be grammatically flawless but culturally tone-deaf.
These are judgment problems, not production problems.
AI is very good at helping with one of them.
The strange unevenness of AI matters more than headline benchmarks
One of the more useful concepts in the Stanford report is what researchers call AI’s “jagged frontier”.
The performance of current models is surprisingly uneven. They can solve problems that most people would consider extremely difficult and then fail at tasks that appear trivial.
Stanford gives a memorable example. Gemini Deep Think achieved gold-medal-level performance at the International Mathematical Olympiad. Yet the strongest model tested on an analogue-clock benchmark managed only 50.1% accuracy. AI agents have made huge progress on computer tasks too, but still fail roughly one third of attempts on a structured benchmark designed to test real computer use.
There are similar problems with factual reliability. In one benchmark covered by Stanford, hallucination rates among 26 leading models ranged from 22% to 94%, depending on the task and the way information was presented.
This is important because companies understandably tend to judge AI by its best demonstrations. A system writes an excellent strategic summary, analyses a complicated spreadsheet or produces an impressive campaign concept, and we naturally start giving more weight to its next output.
But competence in one task does not guarantee competence in the next.
The problem becomes more significant as AI moves further into actual company workflows. A strange answer in a personal ChatGPT conversation has very little consequence. The same behaviour becomes more serious when AI is drafting customer communications, producing advertising, making recommendations, interpreting internal rules or taking actions through an agent.
Marketing has an additional problem: most important decisions are contextual
Brand work is particularly difficult to automate because there is rarely one objectively correct answer.
Consider a partnership proposal.
An AI system can analyse the partner’s audience, engagement, historical controversies, pricing, geographical reach and perhaps even whether there is an obvious overlap with the brand’s target consumer.
That still does not answer the most important question: should this brand be associated with this person?
The answer depends on positioning, history, ambition, competitors, cultural context and sometimes on an instinctive understanding of what the brand should never become.
The same is true of creative work.
Producing twenty campaign routes can save an enormous amount of time. But somebody still has to recognise which one contains an idea worth pursuing. Producing 200 headlines is useful only if the organisation is capable of identifying which five actually sound like the brand.
Stanford’s research on productivity reinforces this distinction. The strongest improvements tend to appear in structured tasks where outputs can be measured relatively easily. Results are more mixed when work requires deeper reasoning and judgment. One study included in the report found that experienced open-source developers became 19% slower while using AI, even though the developers themselves believed AI had made them faster.
Marketing organisations should pay attention to that difference. Much of what makes a brand valuable exists precisely in the areas where there is no obvious scoring function.
Companies are beginning to build governance, but they are behind
Stanford’s Responsible AI chapter gives a fairly sober picture of what is happening inside organisations.
The number of businesses with no responsible AI policies fell significantly, from 24% to 11%, while AI-specific governance roles increased by 17% in 2025. At the same time, companies continue to report knowledge gaps, limited budgets and regulatory uncertainty as major obstacles to putting those policies into practice.
There are other reasons for caution.
The number of documented AI incidents recorded by the AI Incident Database rose from 233 in 2024 to 362 in 2025.
Transparency among model providers has also deteriorated in some areas. Stanford reports that the average Foundation Model Transparency Index score fell from 58 in 2024 to 40 in 2025, with large gaps remaining around training data, computing resources and what happens after models are deployed.
This creates an awkward situation for companies. They are putting more AI into their operations while the systems being used are becoming more powerful, more complex and, in some respects, harder to inspect.
The sensible response is not to stop using AI. There is too much value in the technology for that argument to make much sense.
It does mean companies need to become much more precise about what happens between an AI output and a real-world decision.
Some outputs can probably be used with almost no supervision. Others should be checked. Some decisions should require explicit approval. Certain uses may simply not be worth automating.
Those distinctions will vary dramatically from one company to another.
Brand guidelines were built for a much slower organisation
There is another practical issue for marketing teams.
Most brand governance was designed around documents.
Companies create a brand platform, visual guidelines, tone-of-voice rules and sometimes extensive local-market playbooks. They distribute them to agencies and employees, run training sessions and expect everyone to apply them.
Even before generative AI, this system was imperfect.
Anyone who has worked across several markets will have seen how quickly the same strategy starts producing different interpretations. Headquarters understands the positioning one way. The local marketing team adds its own interpretation. The agency adds another. A creator sees something else entirely.
Generative AI adds yet another layer, except it can produce hundreds or thousands of interpretations.
The answer cannot simply be a longer brand book.
Companies will increasingly need ways of turning their brand principles into practical decision criteria that can be applied while work is being created.
Not only whether the logo is correct, the typography is approved or the tone is friendly enough.
More difficult questions.
Does this idea actually reinforce our positioning?
Is the claim credible?
Does this partnership belong in our world?
Would we make the same decision in Tokyo and Paris?
Does this make sense commercially but damage the brand over time?
There is nothing new about those questions. What changes with AI is how frequently they will have to be answered.
Local culture makes the problem harder
One relatively small section of Stanford’s report should also interest anyone managing an international brand.
AI performance remains uneven across languages and dialects.
On HELM Arabic, a model developed specifically for Arabic outperformed GPT-5.1 and Gemini 2.5 Flash. On a Slovenian commonsense test, several leading models lost close to half their accuracy when tested in a regional dialect rather than standard Slovenian.
This is a useful reminder of the limits of automated localisation.
A translated sentence can be linguistically correct without making cultural sense. Humour, status, social codes, references and even the perceived aggressiveness of a phrase can change considerably between markets.
For global brands, particularly in categories such as luxury, beauty, fashion, hospitality and entertainment, those differences are not details. They are often where the brand relationship is built.
It would be ironic if global companies used AI to increase the amount of locally produced content while making that content less locally intelligent.
What happens when agents become normal?
There is one reason this governance issue deserves attention now rather than in a few years.
Corporate use of AI agents is still relatively limited.
Stanford reports that scaled agent deployment remains in the single digits across almost all business functions. Even in IT and knowledge management, where adoption is more advanced, roughly two thirds or more of respondents still report no agent use.
In other words, most companies are dealing today with systems that largely produce things for people.
The next generation will increasingly do things for companies.
The difference is significant.
Once an AI can publish, contact, purchase, recommend, negotiate, modify or approve rather than simply generate a draft, governance can no longer sit at the end of the process.
It has to become part of the process itself.
That applies to AI governance generally, but there is a specific implication for brands. Companies will need to make their brand strategy usable as a decision system, not simply as reference material.
Stanford’s report does not say this directly. It does, however, provide a remarkably strong picture of the conditions that make it necessary: adoption is accelerating, output is increasing, performance remains uneven, governance is lagging and autonomous systems are only beginning to enter companies.
For several years, being early with AI provided an advantage.
That advantage will become harder to sustain as the technology becomes available to everyone.
What companies choose to automate, what they choose to review, and what they decide should remain under human judgment may eventually matter much more than which model they use.
And for brands, that makes governance much more than a compliance subject.
It becomes part of brand strategy.
---
Rethinking how AI should work inside your brand
Moon Above Advisory works with leadership and marketing teams to define where AI can create genuine value, what should remain under human judgment, and how AI can be introduced without weakening the brand in the process.
Explore our AI Brand Transformation offer →
https://moonaboveadvisory.com/
From brand guidelines to brand decisions
We also built Signal, a Brand Decision System designed for the problem described in this article.
Signal helps teams, agencies and partners review campaigns, content, partnerships, claims and other brand decisions against the brand strategy before they go live.
Not another set of guidelines to read. A way to apply them when decisions are actually being made.
Discover Signal →
https://signal.moonaboveadvisory.com/