The Work After the Algorithm: Dr. Rajan Gupta on AI in the Real World

When much of the world was still trying to understand where artificial intelligence might go next, researchers were already working on questions that would later become central to business. How should machines learn from complex data? How should intelligent systems explain their behaviour? Could information be selectively removed? Could machines recognise relationships and structures that conventional systems struggled to represent?

The same questions now sit inside enterprise decisions. A company deploying AI has to decide what its systems should remember, what they should forget, how they should explain an answer, how they should reason across relationships and how reliably they can operate once their outputs begin affecting customers, employees and transactions.

Dr. Rajan Gupta entered AI through that research-first world. His early work in the late 2000s involved image data, encryption and information outside the structured tables that dominated much of analytics. Images, audio, video, smartphones and digital interactions were expanding both the amount and variety of data available, and he became interested in what would happen once human attention could no longer interpret it at the required scale.

Research remained part of his development as the field matured. His work later extended into machine unlearning, explainable language models, graph neural networks, deep clustering and generative recommendation systems. The subjects vary technically, yet they remain connected by a business concern that has become harder to avoid: whether intelligence can stay reliable, interpretable and controllable as its consequences grow.

At Deutsche Telekom Digital Labs, that concern now meets production environments. Dr. Rajan’s current work includes agentic AI, global system architecture, production engineering, vendor assessment and the translation of business problems into scalable AI systems.

Earlier I was shooting for output of a model. Now I am shooting for business outcomes.

A model can perform well technically and still struggle to create value once customers, processes, cost, infrastructure and organisational realities enter the picture. Dr. Rajan’s work sits in that operating zone, where AI is judged by the decisions, systems and economics built around it.

The Range Behind the Decision

Dr. Rajan’s formal education moved through computer applications, management, business analytics and computer science, followed by doctoral and postdoctoral research. He continued formal study while his professional career was developing, gradually building a range that connected technology, analytics and business.

His route involved several changes in direction. He initially prepared for medicine before moving into computer science, and later chose research when commercial work was no longer giving him the technical depth he wanted. The decision was practical: AI was moving quickly, and he wanted a stronger grasp of the technology beneath the applications.

Enterprise AI now rewards that kind of range. The difficult choices sit between architecture, adoption, economics, governance, customer behaviour and organisational readiness. Technical leaders have to understand enough of the model to question its limits, enough of the product to judge usage, and enough of the business to recognise whether intelligence is creating value or simply creating activity.

When Accuracy Stopped Being Enough

One of Dr. Rajan’s most useful product lessons came while working on an analytics platform intended to make sophisticated data science accessible to people without specialist technical backgrounds. His instinct was to keep adding capability before exposing the product widely to customers.

His CEO questioned whether users actually needed the full scope before the product could begin solving useful problems. If roughly 20 algorithms were enough to create value, releasing earlier would allow customer behaviour to shape the next stage of development rather than waiting for a catalogue of 100.

Do they really need those hundred algorithms? Or are they content with twenty of them? If they are content with twenty of them, why are you wasting time building another eighty?

Dr. Rajan describes the experience as a shift from a scientist’s way of solving a problem toward a product-oriented way of working. Customer needs could differ materially from what an engineering team considered ideal. Delaying delivery in pursuit of completeness could cost time, attention and commercial momentum.

Generative AI has made the same lesson more urgent. Models improve quickly, products evolve while they are being built and customer expectations change alongside the technology. Strong technical judgment now requires a sense of proportion: how much sophistication a problem deserves, how much uncertainty the market can resolve, and when another layer of refinement stops changing the business outcome.

Context Is Where Intelligence Becomes Useful

Dr. Rajan often returns to context when discussing the quality of enterprise AI. He breaks it into three practical components: understanding the domain, understanding the user and retaining enough relevant history for one interaction to inform the next. Context, in his view, is the difference between a system that answers and a system that understands where the answer belongs.

His customer-service example is deliberately ordinary. A customer reports a failed transaction and receives confirmation that the issue is being handled. The next day, the customer asks for an update and is asked to explain the entire problem again. By the third interaction, the system may know the policy perfectly and still offer a poor experience because it has no usable continuity.

If I have already given the complaint yesterday, and today again you are asking me the same thing, then it is not intelligent for me.

Enterprise systems need more than access to documents or policies. Business rules, user permissions, past interactions and operating processes have to come together before an answer feels relevant in the situation where it is being used.

Context will become one of the more important sources of enterprise differentiation as foundation models become easier to access. Many companies will begin from similar model capability. Their outcomes will differ because one company understands its workflows, customers, history, risk boundaries and decision environment more deeply than another.

The Real Return From Automation

Dr. Rajan’s view of automation becomes more useful once the conversation moves beyond the task being automated. HR provides a practical example. Employees repeatedly ask about leave, insurance, benefits and company policies even when the information already exists somewhere inside the organisation.

Dr. Rajan refers to an internal AI system that has handled more than eight million transactions, allowing employees to find information conversationally and reducing repetitive demand on HR.

If routine queries are handled by AI, HR can focus on culture, collaboration and people development.

Productivity will increasingly be judged by redeployment, not simply reduction. Savings matter, but the larger opportunity may lie in shifting people toward work that improves judgment, relationships, speed and organisational capability. Automation becomes more valuable when leaders know what human attention should move toward next.

When Agents Reach the Edge of the Enterprise

Agentic AI brings a different set of constraints into view because the system begins moving from answering questions toward performing actions. Dr. Rajan describes work around a concept in which a single intelligent interface could operate across several applications, allowing a person to request a flight booking and complete the transaction without manually navigating different apps.

The idea was compelling at the interface. Implementation became harder once the agent reached the systems underneath. Applications and back-end environments had often been designed around conventional software interactions and human users, leaving them poorly prepared for agents that needed to exchange information, invoke functions and complete transactions programmatically.

Conceptually it looked fantastic... When we started to implement, we realised that the systems were not ready.

Integration becomes part of the intelligence problem. An agent can reason effectively and still fail because an API is unavailable, a permission is unclear, an identity cannot be verified or an exception has never been formalised. Greater autonomy places more pressure on the operating environment around the model.

Human employees have spent years learning how to work around organisational friction. They know whom to call, which shortcut is accepted and when a process has an unwritten exception. Agents force more of that tacit knowledge into systems, making weak integrations and unclear ownership visible quickly.

The Economics That Arrive With Scale

AI changes the financial responsibility of technology leadership once usage grows. Inference creates a recurring cost. Query volumes can multiply quickly. External model providers can change pricing. Architecture affects latency and infrastructure expenditure. Evaluation, monitoring, security and human oversight continue after launch.

Dr. Rajan argues that technical leaders have to understand these economics much earlier because architecture decisions now flow directly into the cost profile of the use case. A pilot that looks inexpensive at low volume can behave very differently once usage reaches production scale.

Powerful is something which gets your work done efficiently... if you are able to get the work done consistently over a period of time, that is powerful.

AI makes technology choices behave more like operating choices. Cost, latency, reliability and vendor dependency increasingly sit inside the same conversation as user experience and business value. Leaders who understand that connection are less likely to confuse a more capable system with a more useful one.

The Discipline of an Investable AI Business

Dr. Rajan’s view of AI has also been shaped by exposure to venture-side evaluation and technical due diligence. Looking at AI products through a funding lens changes the question. The issue is no longer only whether the product works, but whether it can become a business customers continue using, paying for and integrating into their workflow.

The venture lens sharpens five tests. The first is market pain: whether the problem is urgent enough to survive beyond curiosity. The second is workflow ownership: whether the product becomes part of how work actually happens. The third is data advantage: whether usage improves defensibility over time. The fourth is technology strategy: whether the product can survive platform shifts. The fifth is economics: whether the business still works when compute, inference and customer acquisition costs are fully counted.

The same logic appears in Dr. Rajan’s enterprise AI thinking. A demo can impress quickly. A durable AI product has to survive adoption, integration, cost and repeated use. Venture evaluation brings him back to the same operating question from another angle: whether intelligence creates value after it leaves the presentation layer and enters real work.

The Three Layers Enterprises Cannot Outsource

Dr. Rajan is comfortable using external technology when another organisation has solved a problem well and can bring speed or scale that would be expensive to reproduce internally. His caution concerns the knowledge an enterprise gradually gives away while assembling those external capabilities.

Three layers carry particular strategic weight: business process intelligence, proprietary data intelligence and integration intelligence. Business process intelligence is the organisation’s understanding of how work actually gets done. Proprietary data intelligence is the ability to interpret the enterprise’s own data in ways that reflect business reality. Integration intelligence is the accumulated understanding of how internal systems, workflows and dependencies connect.

A model provider can bring stronger technology. A systems integrator can accelerate implementation. Years of operating history, exceptions, customer behaviour and organisational dependencies still have to be understood by the company that will live with the consequences.

Mid-market companies face the sharpest version of this tension. Large enterprises can afford deeper internal capability across process knowledge, data intelligence and integration engineering. Smaller enterprises often feel stronger pressure to adopt AI quickly and cheaply, while relying more heavily on vendors whose advantage sits precisely in the layers they struggle to build internally.

A more strategic build-versus-buy conversation begins with comprehension. Leaders have to decide which capabilities can come from outside and which forms of knowledge must remain close enough to the business for the company to evaluate performance, challenge vendors, adapt systems and protect long-term operating control.

When Machine Authority Becomes Management Responsibility

More capable AI raises a basic question of ownership. Dr. Rajan is clear that deployed systems remain the responsibility of the people and organisations that design their operating boundaries. Humans decide what information the system can access, what tasks it can perform and how much authority it receives.

As systems move closer to execution, governance has to become operational. Which actions can proceed automatically? When should a person intervene? Can the organisation reconstruct what happened later? Who can suspend or change the system?

Dr. Rajan’s research background gives these questions a technical lineage. Explainability, memory and unlearning begin inside computer science. Once AI enters customer and employee environments, the same subjects connect directly to privacy, regulatory responsibility and the organisation’s ability to inspect machine behaviour.

Governance needs to move from policy language into operating design. Permissions, escalation paths, decision logs and intervention rights will matter more as machine authority expands.

The Skill Shift Beneath the AI Talent Market

Foundation models are changing what enterprises need from technical talent. During the earlier data-science cycle, many organisations built teams around people who could design, train and tune models. Much of frontier model development now sits inside companies with research capabilities, data and compute at a scale individual enterprises rarely need to reproduce.

Dr. Rajan sees more enterprise value moving toward engineers who can take existing intelligence and make it useful inside products. Integration, evaluation, production reliability, workflow understanding, cost and system behaviour become part of the job alongside knowledge of models.

The role of the conventional data scientist has reduced. The need now is for AI engineers who can use, integrate and orchestrate intelligence.

Dr. Rajan’s view of future work is direct.

We will ultimately be annotators and enablers for the AI. We will not be generating anything on our own.

For enterprises, the implication is practical. Human value moves toward framing problems, shaping systems, reviewing outputs, setting direction and connecting machine capability to operating reality. The most valuable people will understand how AI changes the work around it.

The Operating Burden of Leadership

Dr. Rajan’s move into leadership required a similar broadening of perspective. Earlier roles gave him considerable autonomy as a researcher, consultant and individual contributor. Larger teams changed the unit of performance. Decisions increasingly had to work for an organisation and a group of people whose expertise, incentives and constraints could differ from his own.

He describes the order simply: organisation first, team next, then self.

Organisation first, team next, then self.

Communication, patience, conflict management and change management became more important as his responsibilities expanded. Technical judgment alone could no longer carry the quality of the final outcome.

Perfectionism also became more selective. Dr. Rajan describes himself as someone who naturally wants work to be in order, yet enterprise settings rarely offer perfect information or unlimited time. Customers may need a usable solution earlier, markets may move and teams need decisions before every uncertainty can be resolved.

Leadership Lessons From Dr. Rajan Gupta’s Work

• Bring customer evidence into the product while it can still change the product. Technical expertise provides direction; real behaviour reveals priorities that internal teams often miss.

• Treat context as an enterprise asset. Domain knowledge, user understanding and memory allow broadly available models to produce company-specific value.

• Follow the human capacity released by automation. Efficiency has limited value when freed capacity has no destination.

• Prepare the operating environment for autonomy. Agents depend on processes, interfaces, permissions and integrations explicit enough to support reliable action.

• Make AI economics part of engineering judgment. Technical ambition has to survive unit economics once usage expands.

• Keep proprietary understanding close to the organisation. Outsourcing execution should never mean outsourcing comprehension.

• Define responsibility before expanding machine authority. Governance is strongest when it is designed before the failure.

• Develop people who can connect models to operating reality. The scarce skill is system judgment.

• Preserve the ability to simplify. Restraint becomes an advantage when technology is moving faster than adoption capacity.

When the Institution Becomes Part of the Technology

Artificial intelligence has moved from research environments into the operating core of companies with remarkable speed. Questions that once belonged primarily to computer science are now embedded in customer experience, workforce design, product economics, governance and organisational decision-making.

Enterprise competence in AI looks different from conventional technology competence. Model capability matters, but so do the systems around it, the context available to it, the economics supporting it and the people who remain responsible for its decisions.

Dr. Rajan Gupta’s career spans both sides of that movement. His early work developed close to the research questions beneath machine intelligence. His current responsibilities bring him closer to what happens when intelligence meets real processes, large-scale infrastructure and business expectations.

Companies will have to make sharper choices as intelligence becomes embedded deeper into work. Which processes deserve automation? Which decisions require human judgment? Which knowledge must stay inside the organisation? Which systems need to be rebuilt before agents can act safely? Which AI costs will remain manageable once pilots become daily infrastructure?

The real test will sit less in the announcement of AI capability and more in the organisation built around it. A strong model inside a weak operating environment will still produce weak outcomes. A capable agent inside unclear workflows will still expose confusion. A promising automation programme without disciplined redeployment will still leave value on the table.

Enterprise AI will mature through operating discipline: knowing what to automate, what to retain, what to simplify, what to govern and what to keep close enough for leaders to understand. The same discipline applies whether Dr. Rajan is looking at an enterprise deployment or an AI venture. Capability is only the starting point; value depends on workflow, economics, defensibility and the operating system around the technology.