News

If you’ve done any shopping for marketing automation tools these days, you’ve probably noticed they all claim to be “powered by AI.” Apologies for splitting hairs, but that’s just not true.

A significant portion still rely primarily on rule-based automation, work identically to the platforms they replaced, triggering if/then workflows created by human engineers years ago. Give their systems an edge case to parse and you’ll soon see them send an inappropriate email, crash, or output some stale nonsense that wouldn’t matter to any living person.

Same label. Same old architecture. The problem? A rule bottleneck.

Traditional marketing automation relies on knowing rules ahead of time. A lead hits a score? Send an email. A prospect completed behaviors A, B, and C? Trigger sequence Y. Wait Z days, send the follow-up email.

A rules engine can execute these commands flawlessly – but it can only execute what it knows. When a situation arises that doesn’t fit the rules, what do you do? You update the rules.

Marketing is messy. Prospects take unpredictable journeys, trends come and go overnight, and audiences who loved your message last week don’t care about it this week. But rule-based systems can only improve when given new rules to fire. Engineers can’t possibly keep writing rules faster than the world changes.

The industry has been papering over this problem with AI buzzwords. Sprinkle some Neural Network magic on a rule engine, and suddenly you’ve got yourself an “AI platform.” Engineers who look past the updated sales brochures still find the same good old-fashioned if/then statements, patched up with trendy new nomenclature for the latest round of funding.

The difference between legacy automation and true agentic AI is that true agentic AI won’t just patch up the last generation of marketing automation tools – it will replace them. Agentic AI isn’t defined by capabilities so much as by the way decisions are made.

Rules engines ask, “what rule should fire next, given this input?” Agents ask, “what action should I take to get closer to my goal?” This is subtle but critical. Agent theory holds that the system knows its goal, its current context, and a list of available actions it can take.

Based on those three pieces of information, it can reason as to which action will bring it closer to accomplishing its overall objective. This extends far beyond executing canned responses - it’s deciding what to do.

Agentic systems maintain goals, reason over available actions, invoke tools, evaluate intermediate results, and adapt their plans as new information becomes available. The architecture is fundamentally iterative rather than purely reactive.

You know where this is going.

An agent can adapt if a campaign stops performing. It can coordinate with other agents who manage different subsets of that workflow. And it can do so without a human engineer going back into the system to rewrite the rules every time the world changes. The system manages its goals.

One important architectural decision that separates good agent implementations from the rest is specialization. Should you build one big generalist AI system to handle everything or many specialized agents, each performing their own task?

Specialization comes up often in discussions around AI, from medical doctors to Renaissance men. There is broad utility in generalization, but singular accuracy in specialization. The family doctor can handle any symptoms you throw at them. But when you need to be absolutely certain about your diagnosis, you see a specialist.

That’s because specialists aren’t smarter than the generalist – they’re just trained on narrower data. Likewise, generalist AI models aren’t going to produce great results for highly specific use cases. OpenAI’s models can write you a marketing strategy. They can craft creative assets. But they can’t produce marketing assets that:

• Fit the pixel ratio requirements of a given publisher
• Match your brand’s color palette
• Align with your target audience’s emotional affinity profile
• Incorporate mentions of trending topics from the previous day

They can’t do all of those at once, either. And you shouldn’t expect them to. For hard problems with specific solutions, you should build specialized agents (sometimes called “agent crews”) that own a narrow subset of your workflow.

One crew might specialize in strategy generation, while another focuses on creative writing. One might select publishers while another analyzes performance. Separately, these crews create atomic workflows that a generalist system would struggle to manage.

There’s another argument for specialized, privately hosted models that isn’t made enough: data privacy.

Whenever you use a public large language model (LLM) to write marketing copy, your data is being uploaded to someone else’s infrastructure. “We don’t use customer data for training” is easy to say but barely offers any assurance. Inputs are still being ingested, processed, stored, and handled according to what that provider’s internal policies dictate.

And those policies can change… most corporate lawyers have never looked at the data use section of public AI providers Terms of Service, let alone dissected it line-by-line.

But what about controls your organization can enforce? Do your developers scrub data for PII before generating content with an LLM? That only works if everyone in your company memorizes your data policies and uses tools responsibly. One rogue employee attaching a spreadsheet full of internal pricing to a prompt breaks your compliance.

But if the model itself is hosted privately, that’s one major source of exposure that goes away. Your data never leaves your infrastructure. There’s no ingestion point to transmit it to a third-party, no training feedback loop that will process it, and no agreement to parse about how that company will handle your data “moving forward.”

Because AI in the enterprise has reached a maturity level where governance is a legitimate concern, many teams treat it as a bulk edit at the end of AI-generated content. Have humans review and approve. That’s fine, and many teams require this today. But governance should be built into the system at a fundamental level.

Well-built agents have guardrails at every stage of the decision-making process. That means models that make predictions within set bounds. That means observability that can trace every word generated back to its origin.

That means third-party benchmarking to prove your models perform well against industry standards, not just internal testing. Governance shouldn’t just be applied to outputs – it should be inherent in the architecture.

Buying criteria for agentic AI will vary by company, but as requests for proposal accelerate to keep pace with innovation in the industry, here are a few considerations every enterprise buyer should ask about:

• Goals vs. rules - Is this system actually agentic? Or is it just automating workflows with AI tools bolted on? The first step is asking vendors point blank what their system does when it encounters data it doesn’t know how to parse. Rules engines will point to specific fallback rules that execute. Agents will talk about reassessing their goal and weighing their available actions until they decide on the next best step.
• Models and hosting - Where are the models hosted? Are they specialized and trained on domain-specific data? This answers two questions at once – vendor capability, as well as data privacy concerns.
• Long-term memory and context - Enterprise agents become dramatically more useful when they retain organizational context over time. Rather than treating every interaction as a new conversation, they can accumulate institutional knowledge, remember previous decisions, and personalize future actions while remaining within governance boundaries. Persistent memory allows agentic systems to improve continuously without requiring engineers to encode new rules after every edge case.
• Hallucination - No current LLM is immune to hallucinations. The important architectural question is how the system detects, bounds, and mitigates them before they affect downstream business processes. Specialists hallucinate less in their domain of expertise. Prediction window guardrails limit how far an AI system can go “outside the data.” Human approval gates before sending anything live catch anything that slips through.
• Governance / auditability - Can the system provide traceability for every output it generates? Is the system’s accuracy benchmarked against a third-party, or just internally verified?

There's an additional argument that often gets overlooked in discussions focused on capability: cost structure.

Token-based pricing from large model providers creates a fundamentally unpredictable cost model for enterprise deployments. Every question, every generation, every iteration costs tokens — and iterating toward an acceptable output for a complex campaign task can consume a significant volume of them.

Enterprise subscriptions impose usage caps that create their own operational friction. The more AI-dependent your workflows become, the more acute this pressure grows.

Organizations that own and host their own specialized models are not subject to this dynamic. There is no token meter running. The economic relationship is closer to infrastructure than to a metered service - you bear the cost of building and maintaining the system, and in return you have predictable marginal cost. For organizations at scale, that math changes substantially.

AI marketing platforms will continue to flood the market with AI-sounding languages attached to rules engines. But for enterprises who truly want to deploy agentic AI, there’s a far better solution. Domain specific, privately hosted agents that don’t leave your organization exposing itself to risk.

As agentic systems mature, the organizations that differentiate between genuine autonomous architectures and AI-enhanced workflow engines will be better positioned to capture sustainable competitive advantage.

We've featured the best business intelligence platform.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The 2026 World Cup is already under way. Billions of eyes are on the pitch. The story that matters to anyone running a supply chain, though, is playing out in warehouses, customs terminals and distribution centers spread across three countries.

For the first time, the tournament spans three host nations: the United States, Canada and Mexico. That means millions of product lines, thousands of supplier handoffs and cross-border compliance requirements across three distinct regulatory environments, all compressed into a window with zero tolerance for error.

The operational scale of this tournament is without precedent.

Now that the group stages are live, the pressure on supply chains is real and immediate. The lessons surfacing are worth paying attention to, because they apply well beyond sport.

Official merchandise for an event of this scale moves across multiple countries, customs jurisdictions and retail channels simultaneously. Product identification has to work at every stage of that journey, from the manufacturer's floor to the stadium vendor's shelf. The label that cleared customs in Los Angeles may face entirely different requirements in Toronto.

The deeper problem is structural. Today's supply chains still largely operate as disconnected islands. Each site, supplier, co-packer and carrier maintains its own systems and repeatedly re-enters the same product and compliance data. That fragmentation creates built-in waste at every handoff: redundant setup, duplicate records and inconsistent label versions. Under normal conditions, these inefficiencies are costly but can be masked by day-to-day operations. Under the pressure of a live global tournament, they become critical.

Demand shifts are happening in real time. A host city that reaches the knockout stages sees fan merchandise demand surge overnight. Supply chains built on static, batch-processed labelling data are finding they cannot respond at that pace.

The speed of these shifts can be surprisingly tangible. In Atlanta, shortly after the Spain-Cape Verde match, I walked through the airport and was struck by how many people were wearing Cape Verde jerseys. In the space of a few hours, merchandise that had been relatively low-profile had become highly visible, underscoring how quickly demand signals can emerge and spread during a global event.

What the World Cup is making visible in real time is a shift that has been under way for several years. Product identification is no longer a print-and-forget exercise. It is a live data problem.

The industry is moving from fragmented, internal systems to connected, multi-partner ecosystems. The organizations managing the tournament's supply chain most effectively are those that have made this shift: rather than each stakeholder operating in isolation and recreating the same product and compliance data from scratch, they are working within a shared, real-time environment where information flows seamlessly across systems, suppliers, customers and geographies.

The benefits of this approach are measurable. Organizations that can operate with real-time visibility and trusted data across their extended value chain can reduce delays, prevent errors at source and respond faster when disruption hits. In sectors where production downtime can exceed $1-2 million per hour, that responsiveness is not a nice-to-have but operationally critical.

The consequences of siloed product data are well understood in theory. A tournament of this scale is making them visible in practice.

When product data does not flow seamlessly across sites and trading partners, the failure surfaces in predictable ways. Rejected shipments at customs. Compliance failures that stall distribution. Production downtime while teams manually reconcile data across systems. Against the backdrop of a global event with fixed deadlines, those failures are not recoverable.

The organizations absorbing those costs right now are those still operating inside the organization perimeter - managing product identification as an internal function rather than a network-level capability. The distinction matters. Supply chain resilience increasingly depends on the ability to coordinate accurate product data across every site, trading partner, and customer - creating a connected ecosystem in which product identity can be shared, trusted, and acted upon seamlessly.

Connected, network-driven approaches to product identification are no longer a future aspiration. The World Cup is demonstrating their value in real time, at a scale most supply chains will never encounter but from which every supply chain can learn.

The direction of travel is clear. Organizations that can rapidly adapt labelling requirements across plants and partners, share trusted product data in real time and eliminate the manual rework that comes with disconnected systems will outperform those that cannot. That is as true in retail, pharma and automotive as it is in a stadium in Los Angeles.

The World Cup will be over in a matter of weeks. The infrastructure challenges it is exposing will still be there when it ends. The organizations that use this moment to address those fundamentals will be better placed for whatever high-pressure deadline comes next.

We've listed the best product information management software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The conversation about AI in space keeps arriving at the same image: a floating supercomputer processing the world's information from 400 miles up. It’s a compelling narrative, but there’s a gap between what companies are hoping to build and what is being built today.

Today, satellites run on fixed power budgets measured in watts with strict constraints. Bandwidth is scarce enough that every byte reaching the ground has to earn its place.

Those limits push the field toward architecture that looks more like a nervous system, than a data center – lightweight models running onboard that interpret sensor data in near real time and convert observations into structured events. The event reaches the ground.

The raw pixel doesn’t.

In a conventional Earth observation pipeline, a satellite captures an image, downlinks it to a ground station, ground systems process the raw data, and the result reaches whoever needs it. Best case that happens in hours, but often it’s closer to a day.

For a disaster response team managing a flood in a low-lying river delta, or a conservation authority trying to locate the origin point of a wildfire in a remote national park, that day comes at a cost.

A satellite running inference onboard changes the math. Detection happens in seconds. What gets downlinked is a position, timestamp, or risk score. The bottleneck shifts from the space segment to the ground distribution, which is a comparatively manageable problem.

The use cases where this matters most are not the visible disasters people are watching. They’re the methane leak on a pipeline with no weekly inspection schedule, an oil spill beyond the reach of coastal patrols, a wildfire that began in a remote area before anyone had reported smoke. Onboard inference turns a passive imaging asset into an early warning system.

The tradeoffs being resolved in orbit are an extreme version of the same constraints facing any organization deploying AI outside a well-provisioned data center.

Cloud-native AI development often has a back up plan: when the model is too large add compute; when bandwidth is constrained, increase it; when latency is a problem, move the processing closer. In orbit, none of these options exist. You build within the envelope, or the system doesn’t function.

The result is a forcing function that enterprise architects rarely encounter at the same level. Industrial IoT deployments face intermittent connectivity. Autonomous systems can’t afford round-trip latency to a central server at decision time.

The shift from 'send everything, process centrally' to 'process locally, transmit what matters' is happening across multiple industries. Space is where that shift ran without a safety net.

The data reduction numbers transmitted in real time is not just 80-90 percent. Once processing happens on the spacecraft, the reduction for the real-time layer exceeds 99 percent. This is semantic compression. The satellite sends the meaning of what it saw, not the measurement it produced.

A conventional operator downlinking hundreds of terabytes of raw imagery daily is paying bandwidth cost for data that largely contains nothing of interest. With onboard inference, what's transmitted in real time is a structured detection event: a position, a timestamp, a risk score, and perhaps a small compressed image. That is hundreds of kilobytes, not terabytes.

An operator downlinks only what warrants examination, rather than blindly dumping the full data stream.

A model making decisions before a human is in the loop carries different requirements than one generating recommendations for human review. Ambiguity tolerance is lower. Inference behavior needs tighter scoping. This is the same conversation that medicine and finance have been having for years.

AI-assisted diagnostics and accountability distributed across platform, model, training data, and end customer rather than concentrated in a single layer. The space industry is joining a conversation other sectors have already been having for years.

In a cloud environment, model size and computational efficiency are optimization targets. Meaning they are important, but secondary to capability. In a constrained orbital environment, they are the primary design constraint from which everything else follows. A model that cannot run within the available compute envelope is not a model that gets deployed. There is no option to add a larger instance.

A maritime patrol aircraft that previously ran random vessel inspections now works from a ranked list of targets with risk scores attached. Some alerts will be false positives which is a physical reality of any probabilistic system. But the aircraft's operational effectiveness improves substantially compared to random patrolling or no monitoring at all. The AI narrows the search.

One satellite running an onboard model is a proof of concept. A constellation of hundreds running distributed inference is a different infrastructure problem as orbital AI scales.

Centralized orchestration becomes the bottleneck when constellations grow. Every decision can’t route through a ground station. Distributed inference is a requirement. Enterprise architects hit the same wall when a pilot deployment expands to thousands of edge nodes. The centralized model that worked in development becomes the thing that breaks in production.

The cloud infrastructure analogy supports this. Nobody builds a data center before launching an application. The pattern is shared infrastructure, with control at the model, mission logic, and decision layer. Those can be sovereign regardless of who owns the underlying compute.

It’s hard to develop constraint-based thinking in environments where adding compute is always on the table. The organizations that have built it tend to have faced conditions where it wasn’t.

The strategic advantage in edge AI over the next decade will not just be measured in the amount of compute available. It will also be measured in code deployed to the right place in the stack. Satellites are running that experiment first.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit