The next wave of enterprise transformation will not be led by humanoid robots but by smart content AI that understands, interprets and acts on the documents organisations handle every day. As agentic AI moves from experiment to infrastructure, the real competitive advantage will come from how intelligently businesses manage the information they already own.
For years, the public imagination has been distracted by videos of humanoid robots climbing stairs, opening doors and dancing in factory showrooms. The narrative has been seductive and straightforward. Physical work would be the first to fall, machines would replace manual labour, and knowledge work would remain safe for longer. That story is now colliding with a more uncomfortable reality. The most immediate disruption is not happening in warehouses or on loading docks, but in offices, where routine white-collar tasks are quietly being absorbed by software.
Dr Bates, a Cambridge Computer Science academic and now CEO of SER Group, believes the industry misunderstood where AI would bite hardest. “The misperception was that AI would be good at replacing menial jobs,” he says. “What we have learned is that those jobs are much harder than you think because robots are complex and fine motor skills are challenging to replicate. The surprise is that AI has turned out to be good at replacing simple white collar tasks, not so much blue-collar work, and I do not think anyone was expecting that.”
The labour market data is starting to confirm that picture. Entry-level roles in administration, clerical support, and junior analysis are becoming scarcer, while automation spreads through back-office processes that once relied on human pattern recognition. Bates points to recent figures on graduate and apprenticeship vacancies as an early warning signal, not an anomaly. “For the first time, vacancies for graduates, apprenticeships, internships, and junior white collar workers have fallen by over thirty per cent,” he says. “That is roughly a quarter of the jobs market, and it is happening because of AI. This has crept up on us, and I think we are now going to see month-to-month drops in entry-level clerical work, which is a massive part of the market.”
At the same time, many executives still treat AI as an abstract horizon rather than a present operational concern. The reality is more prosaic and more immediate. The real battleground is not a futuristic robot but the document, the email and the workflow.
From dark data to living knowledge
Most enterprises remain heavily dependent on documents, even if the paper has long disappeared. Contracts, invoices, orders, claims, resumes, design packs, meeting notes, and compliance records now circulate as PDFs, images, messages, and attachments across every corner of the organisation. Much of this content is unstructured or semi structured, difficult to search and poorly connected to the systems that depend on it. Bates sees this as the core opportunity for smart content AI.
“Every large organisation you can think of has hundreds of millions, if not billions, of documents,” Bates continues. “A lot of that is dark. The question is what hidden nuggets exist in that dark data that you do not know about. If you think about something as simple as your daily Teams or Zoom calls, you’ll see ideas, presentations, documents, and minutes being created all the time. Where is that, how do you link to it, and when is it presented to you now you need it?”
He argues that the same problem exists across long-established systems. “You have information in ERP, in CRM, in email systems, and in repositories that have been around for twenty or thirty years. There are still insights in that material, especially in the last year or two, but most organisations have no effective way to unlock it and compare it against what is happening now.”
The goal of smart content AI is to turn these archives into an active intelligence layer. That shift goes beyond basic extraction. Systems must be able to read documents, understand their structure and meaning, and relate them to the broader business context. They need to spot fraudulent invoices by recognising deviations from historic patterns, not just by matching fields. They should be able to connect a customer complaint to a service ticket, an order, a support email and a contract clause without human intervention.
Bates sees early progress in this direction, but believes the potential is far from exhausted. “We already see businesses that understand what is in emails, automatically launch orders, start workflows and look for fraud. There is so much more opportunity out there. We do not even know what all the use cases are going to be yet.”
Building the wisdom backbone
Delivering that capability requires more than incremental upgrades. Intelligent document processing, enterprise content management, and process automation have traditionally evolved as separate categories, each with its own vendors, architectures, and teams. Bates argues that fragmentation is now the primary constraint on progress.
“You need to bring together three worlds that have historically been separate,” Bates explains. “One is intelligent document processing, which uses AI to understand documents. Another is enterprise content management, which handles storage, management, search, archiving and collaboration. The third is process automation, particularly robotic process automation around documents. Those must come together because you need to understand, automate, collaborate, store, and search in an AI-native way. Unless that happens, you are not going to be able to do the things we are talking about.”
Where that convergence is happening, the results are tangible. In manufacturing, incoming orders arrive as emails with diverse attachments and inconsistent formats. Smart content systems can now interpret the attachments, extract relevant fields, and launch chains of actions. CAD tools can be opened with the correct design template, ERP systems can be updated with order details, and fulfilment workflows can be triggered against defined service levels, all without manual intervention. Similar patterns are emerging in insurance claims processing, contract review, HR onboarding and legal approvals.
The more profound change comes when organisations stop treating these as isolated automations and start viewing them as a connected fabric. Bates describes the direction of travel as a “wisdom backbone” that runs across the enterprise. “It is not just one system linking to another,” he says. “You are removing silos of documents and processes and bringing them all together. You can see that this customer in the SAP system is doing something in the CRM system: they have logged a service ticket and are not paying their invoice. There may be an opportunity, or there may be a risk, and that insight can be surfaced to the sales team or the legal team at the right time.”
When content, context and process are unified in this way, knowledge stops being a static record and becomes an operational asset. That is where agentic behaviour begins to emerge.
What agentic AI really looks like
Agentic AI is one of the more overused terms in the current cycle, yet in the content domain, it has a clear, practical meaning. Instead of waiting for a user to ask a question, the system operates continuously in the background, monitoring streams of information, identifying relevant signals and taking bounded actions within a defined arena. Bates is succinct in his description. “A truly agentic system is operating asynchronously to the main process,” Bates adds. “It behaves like a curator or a knowledge worker in its own environment. It processes information, raises escalations and takes autonomous actions, but always within a predefined arena. When it hits the boundary, it hands off to a human.”
The key difference from earlier generations of automation lies in reasoning and timing. An agent can watch for patterns across time rather than reacting to a single trigger. It might see that an employee has applied for external roles, logged negative survey responses and shown declining engagement in internal communications, and flag retention risk before a resignation letter arrives. It might recognise that a contract clause conflicts with existing agreements and route it for special review.
Yet even the most advanced systems remain bounded by experience. Bates is clear that AI will never encounter every possible scenario. “There will always be things that fall outside the range of what has been seen before. Those will be flagged for a human. The principle is that knowledge workers should become far more productive because they focus on high-level tasks and are interrupted only when there is a decision to make.”
Learning on the job deepens this capability but introduces new complexities. Feedback loops can tune models as they encounter fresh patterns, yet they can also amplify noise and drift. Bates warns that careless retraining can erode performance rather than improve it. “Large language models have been trained on internet-scale corpora, but as they generate new content, that output starts feeding the next generation. You risk losing fidelity if low-quality or esoteric outputs get treated as ground truth. It is like a village that never broadens its gene pool. Without discipline, you can create abnormalities in the next generation of models.”
For enterprises, this means that governance and data curation must sit alongside experimentation. Agentic capability is not a licence to leave systems unsupervised.
Circuit breakers for autonomous systems
As more decision-making shifts towards AI, the question of oversight becomes central. Bates returns repeatedly to the need for robust control mechanisms, drawing a parallel with algorithmic trading. “In financial markets, we have had agentic AI for a long time,” he continues. “Trading algorithms have caused flash crashes. The lesson from that world is that you need circuit breakers. If certain parameters are hit, you shut things down. In practice, it was too complex to put circuit breakers around every algorithm, so regulators imposed them on the market itself as a blunt instrument.”
He believes enterprises must design something more precise. “You need what I think of as a risk firewall around AI. You monitor its output, and if it goes outside predetermined parameters, you cut it off. That is essential for mission-critical systems. People have not been here before, and they get excited about what AI can do, but you must be careful what it has direct access to.”
The consequences of failure depend on the domain. In healthcare, misclassification or misrouting can have life-critical implications. In finance, errors can cascade into significant losses. In customer operations, misjudged decisions can erode trust and damage relationships. Bates argues that governance audits must become standard before deployment. “You need to ask what the worst that could happen is, not just what the best case looks like. Explore aggressively but understand the downside and put the right protections in place.”
Regulation will add further pressure. As governments confront the social impact of automation on white collar employment, there is already discussion of how to tax or constrain AI in ways that reflect its economic footprint. Bates is openly curious about where this might lead. “If white collar workers are being replaced by algorithms, will governments start to tax AI itself. Maybe that becomes a new form of tax that nobody had seriously considered.”
Designing work around the human edge
One of the more nuanced challenges in this shift is the risk of de-skilling. Suppose AI systems handle routine tasks and humans intervene only in exceptional cases. In that case, there is a danger that the next generation of workers will never develop the baseline skills that used to come from repetitive exposure. Bates sees both benefit and risk. “If you are the owner of a small accounting business, AI can check tax returns and flag issues, which frees you to spend high-quality time with your clients,” he continues. “That is a clear win. The other side is that entry-level jobs people used to rely on to get into the market may go away, particularly for semi-skilled roles. That is the part that worries me.”
The answer is unlikely to lie in slowing the adoption of intelligent systems. Instead, it will require a deliberate rethink of how organisations train, mentor and expose junior staff to meaningful work. In many ways, this mirrors past industrial transitions. When mechanical automation entered factories, apprenticeships and vocational pathways had to change. In the age of agentic content systems, knowledge work needs an equivalent redesign.
Bates remains optimistic about the broader trajectory. He has been waiting for this moment since his days as a young computer scientist. “It has been my dream to see systems with some level of true intelligence assist people,” he explains. “We will adapt the job market because humans always have. When coal mines closed, it was painful, but the world had to adapt to new forms of work. Now people will have more time to focus on high-level tasks, and software will finally be able to help in a meaningful way.”
For executives, the immediate question is not whether Rosie the robot will arrive this decade. It is how quickly they can start building an intelligent fabric around the content, processes and decisions that already define their organisations. Bates sees no reason to wait. “We are still in the digital realm, not the physical one,” he concludes. “We can embrace AI-powered document intelligence today and roll it out. The opportunity is there right now.”
The enterprises that act on that opportunity will not look futuristic at first glance. Their offices will not be full of walking machines. Their advantage will come from quieter changes. Fewer manual searches, fewer lost documents, fewer missed signals, and far more decisions made with the full weight of organisational knowledge behind them.
