For years now, vendors have been competing on benchmark scores, inference speed and model capabilities as businesses try to work out where AI can fit in with their daily workflows, but experimentation has evolved into actual deployment and what’s important is changing.

Accuracy and measurable business outcome is now more important than ever, with questions around return on investment, accountability and governance being raised across all sectors.

This is especially important in ecommerce, for example, where AI-generated imagery must be totally accurate. Marketers no longer have any qualms over using AI to produce ad variants and subtle market tweaks, but the quality of the output needs to be consistently high to ensure that the product is being accurately reflected.

Even the smallest changes in color, texture or dimensions can have major reputational risks, like a drop in customer trust or a rise in product returns. We already know that poorly executed AI strategies reduce customer trust, and that a lack of branding consistency also undermines the perception of a brand.

AI usage patterns are changing buying habits

But this challenge is also playing out as pricing models for AI subscriptions evolve. The boom started off with seat-based pricing and token consumption models, but we’re entering a new era where wasted AI is no longer being chargeable.

Zendesk, for example, recently announced that it would only be charging its customers when they realized verified outcomes, and Photoroom CEO Matt Rouif believes this pricing strategy could be a big hit for AI customers as customers shift from capability-led adoption to assurance-led adoption.

To explore how enterprise attitudes toward pricing and generative AI outputs are shifting, how organizations are measuring success, and why assurance and accountability are growing in importance, I spoke with Rouif.

• How are enterprise buyers’ expectations of generative AI changing as adoption matures?

As generative AI moves from experimentation into production, enterprise buyers are becoming far more focused on control, accountability and measurable business outcomes.

The first phase of adoption was largely about proving that AI could generate a credible output, whereas the current phase is about whether those outputs can be trusted inside commercial workflows, where accuracy and reliability directly influence customer experience and business performance.

Increasingly, enterprise leaders are asking more about governance than model capacity, assessing how outputs are evaluated and where accountability sits when something goes wrong.

In e-commerce, this is significant because a product visual is a central part of the buying decision. If an AI-generated image changes a colour, or alters or removes a detail, the issue moves beyond creative quality and becomes one of product fidelity, with direct implications for consumer trust, returns and commercial performance.

Our marketplace research reflects that changing expectation, with 55% of consumers saying poorly executed AI-generated or heavily edited product images make them trust an online marketplace less, while 77% expect marketplaces themselves to ensure product listings are accurate and trustworthy.

The same thinking is increasingly shaping enterprise procurement, where organisations are moving beyond asking whether AI can generate an output and toward whether those outputs can meet agreed commercial standards at scale.

That broader shift is turning assurance into a procurement consideration, with enterprise buyers increasingly expecting AI vendors to define quality standards upfront and stand behind outputs once they are operating inside live commercial workflows.

• What are organisations using to measure whether AI initiatives are delivering meaningful business value?

The way organisations measure AI is changing significantly, with early adoption often judged by visible output and perceived model capability, while today the discussion sits much closer to traditional business performance.

Leadership teams are now focused on whether AI can make production materially more efficient, less resource-intensive and more commercially useful. The question is no longer how much AI can produce, but how much of that output is genuinely usable inside a live enterprise workflow.

In e-commerce, that distinction becomes particularly important because production does not end when an image is generated. If teams still need extensive manual review, correction and quality assurance before an asset reaches a customer, the bottleneck has simply moved downstream rather than disappeared altogether.

Enterprise buyers are therefore placing much greater emphasis on output readiness than output volume, measuring whether AI-generated assets are accurate enough to be deployed with confidence rather than simply generated at scale.

Similarly, commercial standards are becoming a more meaningful measure of AI maturity than generation quality alone, reinforced by our March 2026 buyer analysis, which reflects 37% of enterprise buyers name inaccurate visuals as their top pain point in AI visual production.

The analysis reflects a clear organisational requirement for a distinct framework determining whether outputs are fit for purpose before they become customer-facing.

• What challenges remain when deploying AI-generated content in commercial environments at scale?

The biggest challenge is that commercial deployment exposes the difference between generating content and governing it. A single AI-generated image can appear convincing in isolation, but enterprise commerce depends on thousands, and often millions, of assets being accurate enough to support purchasing decisions.

Scale magnifies small inconsistencies, and those inconsistencies quickly become operational and commercial risks rather than creative ones. As a result, enterprise organisations are increasingly investing in validation as much as generation.

Small changes to a product’s colour, shape or packaging may pass a superficial creative review while still misrepresenting the product itself.

Our marketplace research reflects the commercial importance of that challenge, with 63% of consumers saying variation in product imagery, branding or presentation makes a seller or marketplace appear unreliable, while 51% believe marketplace listings often look acceptable but still fail to give them complete confidence in what they are buying.

At catalogue scale, those numbers represent a significant volume of assets that look passable on review but carry real commercial risk once they reach customers.

The unresolved enterprise challenge is therefore not simply producing better AI outputs but building systems capable of catching and correcting failures before they become customer-facing.

• How should businesses balance creative flexibility with consistency, accuracy and reliability in AI-generated outputs?

Businesses should think of product truth as the fixed foundation, with creative flexibility built around it rather than replacing it. AI is exceptionally good at adapting content for different channels, audiences and formats, but those creative decisions should never compromise the factual attributes of the product itself.

The practical way to operationalise that distinction is to define upfront which product attributes are locked - colour, dimensions, materials, key details, and which elements sit within the creative range.

That gives teams a clear framework for evaluating outputs rather than relying on subjective review at the point of approval.

Our research suggests consumers are already making that distinction themselves, with only 33% of UK consumers saying they are comfortable with AI-enhanced product images if clearly labelled, while 41% disagree.

That result matters because it shows transparency alone is not sufficient - the more important question is whether customers believe the image accurately represents what they will receive. For enterprise organisations, that means the governance framework has to be built around factual accuracy as the primary standard, with creative flexibility operating within those boundaries rather than alongside them.

• Where do specialist AI tools add value compared with more general-purpose AI models for enterprise workflows?

General-purpose models have dramatically expanded what AI can create, but enterprise deployment depends on much more than generation capability alone.

Organisations increasingly need systems that understand the commercial context in which those outputs will be used and can support quality evaluation, workflow integration and consistency at scale.

The value therefore shifts from the model itself to the operating system built around it. A general-purpose model may produce an attractive product image, but enterprise teams also need confidence that the product remains accurate, consistent at catalogue scale and that failures can be identified before assets become customer-facing.

The long-term value of specialist AI therefore comes less from producing visually impressive outputs and more from solving repeatable commercial problems. In visual production, that means building systems capable of evaluating product fidelity, reducing manual review and creating structured validation processes that organisations can rely upon consistently at scale, rather than depending solely on subjective human approval.

• What forms of accountability or assurance are enterprise customers increasingly looking for from AI vendors?

Enterprise buyers are increasingly looking for AI vendors that can operate within clearly defined commercial standards rather than simply offering more capable models. That means agreeing success criteria before deployment, evaluating outputs transparently and creating clear processes for handling exceptions when outputs fall short.

As AI becomes embedded within operational workflows, assurance is becoming every bit as important as capability. A failed AI output during experimentation is largely an inconvenience, whereas a failed output inside a live commercial environment can affect customer trust, listing performance and revenue.

Our marketplace research reinforces why this is becoming a board-level discussion, with 51% of consumers saying they would switch to a different marketplace entirely if another platform offered clearer, more accurate product images, while 62% believe marketplaces should actively help sellers improve listing quality.

That expectation increasingly extends beyond marketplaces to the technology providers supporting them, with buyers beginning to look for accountability mechanisms that resemble those expected from other enterprise software providers.

• Looking ahead, what do you see as the next major shift in enterprise AI adoption over the next 12–24 months?

Over the next 12 to 24 months, I expect enterprise AI adoption to move decisively from capability-led adoption to assurance-led adoption.

Businesses will continue to care about model quality, speed and efficiency, but those attributes will become increasingly expected rather than differentiating.

The organisations creating the greatest value will be those capable of embedding AI into revenue-critical workflows with confidence, governance and measurable accountability.

In practical terms, that means much greater emphasis on evaluation, validation and operational trust. Enterprise buyers will increasingly ask how outputs are verified, how failures are handled, how responsibility is shared and how AI systems integrate into existing governance frameworks.

Commerce is likely to be one of the first industries where that transition becomes visible because consumers remain cautious about AI within the buying journey. Our research found that only 24% of consumers already use, or are happy to use, AI tools to help them shop online, while 59% remain uncomfortable doing so.

The next chapter of enterprise AI will therefore be defined less by what models can generate and more by whether organisations can deploy those outputs repeatedly, responsibly and with confidence.

AI’s timeline is very much still being written, but one thing is clear – companies are now in the midst of shifting from experimentation to widespread implementation after having determined strong use cases, with security and trust now becoming higher priorities.

The question is no longer about whether employees are willing to embrace AI, because that much is clear. It’s now about whether their employers know how AI tools are actually being used, whether they’re providing the right type of solutions, and whether their governance supports real-world use cases.

Off the back of that, companies are now struggling to tame shadow AI as workers go off to explore their preferred tools, rather than being confined to workplace-provided alternatives. But while organizations have years of experience handling shadow IT, shadow AI is presenting new challenges.

Shadow AI is harder to tame than Shadow IT – gaining visibility is the first step

Rather than being blocked from downloading certain software, workers can almost painlessly head to their chosen AI tool directly from the browser or via a personal account without approval or restrictions. As much as two-thirds (67%) of enterprise AI use now takes place through unmanaged personal accounts, even when an organization already provides enterprise-grade licenses.

But those sanctioned AI tools are clearly working for employees, who are seeing higher productivity. At the end of the day, this is a major win for companies who are under pressure to prove ROI, but shadow AI presents security risks that enterprise-grade software generally negates.

Teramind has revealed that 86% of organizations lack visibility into how data moves to and from AI tools, and it’s not just knowledge workers who are to blame. Nearly seven in 10 C-suite execs also admitted to prioritizing speed over security.

I spoke with Teramind VP of Strategy Leeron Walter to understand why shadow AI has become more of an issue than we might’ve thought, and what organizations can realistically do to regain visibility and control while continuing to meet workers where they feel most comfortable and productive.

• How do you define shadow AI, and why does it happen inside approved tools?

Shadow AI is any AI usage that operates outside organizational visibility and governance - whether through banned apps, personal accounts, or AI features embedded in tools you already pay for.

The reason it's hiding inside approved platforms is simple: vendors are racing to embed AI into everything. Your licensed Microsoft 365, your PDF reader, your CRM - they all have AI features now.

Our research shows 67% of enterprise AI usage runs through unmanaged personal accounts on corporate-licensed platforms. The perimeter didn't move. It dissolved.

• Do executives actually follow the AI policies they sign off on?

Not always. Our data is unambiguous: 69% of C-suite leaders prioritize speed over security when using AI tools, versus just 37% of frontline employees.

Executives feel competitive pressure more acutely, so they rationalize bypassing policies.

• What goes through an employee's head when they choose productivity over compliance - and can companies change that?

They're doing a fast cost-benefit calculation: "Missing this deadline hurts me now. A data breach is someone else's problem later." 60% of employees in our research said productivity benefits outweigh security risks when deadlines are involved.

You don't fix that with more restrictions - 48% said they'd use AI even if it were explicitly banned. You fix it by making the secure option just as fast and frictionless as the risky one. Remove the tradeoff entirely.

• Is Gen Z really more likely to work around AI rules?

Yes, but not because they're reckless - because they're impatient with policies that feel arbitrary. For them, AI is a basic utility, like a search engine.

Blocking it doesn't register as a security measure; it registers as the company being behind. Meet them with speed and enablement, not bureaucracy.

• Why do traditional DLP tools miss AI traffic?

Because they were built to catch files moving, not ideas being processed. Shadow IT was about unauthorized storage - a file uploaded to Dropbox.

Shadow AI is about unauthorized processing - sensitive data pasted into a chat prompt. There's no file transfer to intercept. The data moves through an encrypted browser session, and legacy DLP tools are pattern-matching against file types and network transfers, not semantic content in a chat box.

The threat model changed; the tools didn't.

• What does the first 90 days of gaining AI visibility actually look like?

Days 1–30: Observe, don't block. Deploy behavioral telemetry to build a full Shadow AI inventory - browser extensions, clipboard activity, personal account usage inside approved platforms. Understand what's actually happening before you touch anything.

Days 31–60: Categorize risk. Which tools train on user data? Which departments depend on them? This is when you find out Engineering lives in an unvetted coding assistant.

Days 61–90: Enable and enforce. Roll out approved alternatives for high-risk tools. Implement real-time coaching - block the risky action, surface the safe alternative immediately. Goal: not zero AI usage, but 100% visible AI usage.

• What does an enablement-first AI approach actually look like - and how do you stop it becoming shadow AI with extra paperwork?

You build paved roads. Give employees a fast, secure, approved AI path so they don't need to go off-road. That means enterprise AI tools with zero-retention data policies, integrated into existing workflows - not buried in a separate portal.

To avoid it becoming theater, your AI tool approval process needs to be agile. If the review takes six months, employees use the consumer version today and say nothing. Govern the data, not the application - allow the tool, but monitor and control what data flows through it in real time.