Agentic AI is the hot topic of the moment. However, that term covers a broad range of applications – from simple AI automations, through to more complex autonomous agents. A truly autonomous AI agent is very different from a simple automation, in that you give it a goal (rather than a script), and it reasons its way to an outcome. It can break a goal into individual steps and use tools or other systems to take actions in the real world and handle situations that may have never explicitly been programmed for.
The Martech 2026 report reveals that most deployments by marketers are currently automations rather than fully autonomous agents. With 80% of respondents in the Martech survey reporting that they operate agents in ‘assist only’ mode (AI suggests, human decides), and a further 38% require human approval before execution. This means a very large proportion of the small number of agents currently in use still operate with tight human oversight. Placing them firmly at the less-autonomous end of the spectrum of what ‘agentic’ can mean.
Though most organisations are only in the very early days of agentic usage, there are emerging visions of what the future of more autonomous agents might look like. Focusing on consumer usage, in the company’s annual letter the payments platform Stripe set out their vision of how agentic commerce may progress. Which is a useful model for understanding where AI shopping agents may be headed.
Right now, we’re at the start of this journey but Stripe’s view is that where we go from here depends entirely on the industry agreeing shared protocols. It is worth noting that the infrastructure race for this is already underway.
AI models have a finite context window (a limit to the amount of information they can reliably hold and reason with at any one time). If the context window is overloaded the performance degrades and you risk getting hallucinations, missed steps or decisions made on incomplete information. This is what Scott Brinker calls the Goldilocks rule: too much noise is as damaging as too little signal. The goal is context that is ‘just right’, something done via context engineering.
Context engineering is not just a technical challenge; it will require organisations to think carefully about how to make organisational and other knowledge accessible to AI. Traditional knowledge management has always been about storage and retrieval – creating resources to help people find information. With the emergence of AI agents, organisations now will need to think bigger and consider their ‘knowledge architecture’.
In practice knowledge architecture may begin with a foundation of readily available knowledge resources such as customer personas, brand assets, style guides. The next step is to broaden the range of knowledge that a marketing AI could be using to reason with. Such as – customer data (behavioural patterns, purchase history, zero and first-party insights), brand knowledge (strategy, tone of voice, creative principles, what’s worked and what hasn’t), campaign performance history (performance data, test results, the institutional memory of what this audience actually responds to), competitive intelligence (what’s happening in the market, how competitors are positioning themselves), alongside broader cultural and contextual signals (the world your customers are living in). Organisations may also include inputs from other functions such as sales, supply chain and finance to ensure a broader, more joined up perspective.
Most organisations are beginning with creating AI enterprise search, but the real opportunity may well come from the ability to reason across diverse knowledge sources in a more deliberate way than a general LLM engine.
The issue of ‘cognitive outsourcing’ (the increasing delegation of mental tasks to AI tools) is a risk in human/AI collaboration. In January this year, Anthropic (the company behind Claude) published a randomised controlled trial with software developers. Participants who used AI assistance scored 17% lower on tests of the skills they’d just used. The study used a grading scale where the AI group averaged 50% versus 67% for those who coded by hand.
Interestingly, the biggest gap was in debugging, or the ability to spot when something is wrong and understand why. This is significant because the capacity to oversee, verify and correct what AI produces is becoming increasingly important as AI is further integrated into processes. Put simply, the risk of simply outsourcing to AI to perform tasks and functions is that marketers do not develop the real expertise and skills needed to assess when the output is not good enough. Here, speed without judgment can become a liability.
The Anthropic study contained an important caveat. Using AI did not guarantee lower scores but instead what mattered was how people used it. Participants who asked follow-up questions, requested explanations, and engaged conceptually while coding retained significantly more than those who simply let AI generate the answers. The tool was the same but how the participants used it and their intention was different.
This connects directly to research from the Universities of Michigan and Chicago, which identified the importance of what they call ‘machine fluency’. This is the ability to instruct an AI agent to effectively align with your own objectives. Their findings showed a 73% of variance in AI performance derives from the user’s ability to instruct the AI. Machine fluency isn’t about technical know-how, it’s about clarity of thinking, quality of judgment, and the ability to direct AI with precision and purpose. The good news for marketers is that this is a learnable skill. The question is whether your organisation is investing in developing it.
On the human determinants of AI success, as AI capabilities improve rapidly, marketing leaders will increasingly need to consider how they can establish and lead hybrid teams of humans and AI agents to maximise effectiveness and impact, while not destroying trust.
When we talk about leading hybrid human and AI teams, the first shift is to frame this as orchestration rather than replacement. AI handles speed, repetition and pattern recognition and humans focus on judgement, context, relationships and ambiguity, but the aim is a system stronger than either alone.
This is fundamentally a cultural shift. Roles, expertise and status are affected, and the technical setup is often easier than the human transition. As part of this – clear boundaries matter. Teams should know when AI must defer to human judgement. Defined limits create safety and confidence.
Finally, hybrid teams should be built to learn. Feedback and adaptation must be designed in from the start. Rigid systems break but learning systems endure.
Download Econsultancy’s Future of Marketing report here