In conversation with Brij Pahwa, DEPT® technology leaders Jonathan Whiteside and Yash Mody discuss the shift from AI assistance to agent-led workflows, the limits of autonomy, emerging enterprise use cases and why companies must rethink processes before deploying agents at scale.
Artificial intelligence is rapidly moving beyond tools that assist individual employees. Enterprises are now experimenting with systems that can perform tasks, interact with software and execute parts of complex workflows with varying degrees of autonomy.
But the transition from generative AI to agentic AI is not simply another technology upgrade.
Jonathan Whiteside, Global EVP of Technology at DEPT®, and Yash Mody, CTO APAC & DEPT® Adobe Practice, argue that agentic transformation requires organizations to redesign how work moves across people, processes and technology. In an interview with Brij Pahwa, the two leaders discussed what agentic engineering means, where enterprises are seeing measurable results and why human oversight will remain central to the next generation of AI systems.
Agentic engineering is a delivery model, not simply a technology
Whiteside described enterprise adoption of AI in engineering as progressing through three broad waves.
The first was “vibe coding”, where a user could prompt an AI model to generate an application or a piece of code. The second involved context engineering, which introduced more structured instructions and frameworks around how large language models were used.
Both waves, he explained, were primarily assistive. They enabled individual employees to perform their jobs faster.
Agentic engineering represents a wider organizational shift.
“I see agentic engineering more as a delivery model than I do a technology,” Whiteside said.
Instead of accelerating only one developer or employee, agentic engineering examines how AI can operate across an entire workflow. This could begin with developing a brief and defining requirements, continue through software creation and testing, and extend to deployment.
The objective is not simply to make an individual faster, but to increase the speed and effectiveness of the entire team and organization.
Mody agreed that enterprises are approaching an inflection point. Emerging technologies, he observed, typically begin by attracting widespread attention before organizations introduce standards, controls and repeatable methods for using them.
Agentic engineering is now entering that more structured phase.
Enterprises, Mody said, are beginning to consider how existing workflows can be replaced or redesigned as agentic workflows. The goal isn’t simply to replace workflows with agents. It’s to redesign workflows where humans and agents collaborate. It’s a transformation of the operating model itself. Look at Agentic OS, which companies, including ours, are developing to enable this partnership in the most efficient way.
Agents are not reliable or autonomous by default
Despite the industry enthusiasm surrounding AI agents, both leaders cautioned enterprises against treating autonomy as unsupervised operation.
Mody identified three major misconceptions.
The first is the assumption that agents are reliable by default. AI agents can produce incorrect content, but they can also take incorrect actions when executing instructions.
The second misconception is that an agent can be placed on top of a dysfunctional process and somehow repair it.
“Agents amplify dysfunction,” Mody said.
If an organization introduces an agent into an unclear, fragmented or poorly governed workflow, the technology could reproduce and scale the problems already present within that process.
The third misconception concerns autonomy itself. According to Mody, allowing an agent to operate autonomously should not mean removing human supervision altogether.
“Autonomy is earned incrementally,” he said. Organizations should determine the level of autonomy an agent receives based on its demonstrated reliability and the business risk involved. Different workflows require different levels of human oversight. Human checkpoints will continue to be required, particularly when an agent’s actions could affect customers, employees, finances or regulated business operations.
Whiteside similarly argued that agents must operate within tightly defined guardrails. Organizations need to establish the context available to the agent, the tools it can access, the scope of its activities and the checkpoints through which the quality of its work will be evaluated. “You can’t just let them loose,” he said.
For many organizations, Whiteside added, the AI model itself is not the primary obstacle.
“The context and the knowledge is the bottleneck we see in most organizations,” he said.
Software engineering is leading enterprise adoption
Among enterprise functions, software engineering is currently seeing some of the strongest demand for agentic systems, according to Whiteside.
Engineering work is often highly structured, information-rich and relatively easy to verify. Code either runs or it does not, giving organizations a clearer mechanism for determining whether an agentic process has succeeded.
Adoption becomes slower when the quality of an output depends on subjective judgment, taste or situational interpretation.
Content development is another area attracting enterprise interest, particularly when organizations need to adapt an existing piece of content into multiple formats for different channels.
Whiteside distinguished this from original creative development. Enterprises are moving more confidently in areas where content can be checked for accuracy or compliance. Adoption is more cautious when the output requires empathy, creative judgment or a nuanced understanding of brand identity.
The pattern suggests that organizations are currently most comfortable deploying agents where verification is fast, measurable and repeatable.
From demographic segments to reasoning-based personalization
Within customer experience, Mody expects agentic systems to move personalization beyond fixed demographic segments and predefined conditions.
Traditional personalization has generally relied on information such as geography, income, customer category or previous transactions. Customers are placed into segments and shown messages selected for that broader group.
Agentic systems can potentially introduce a much richer level of context.
Instead of relying only on demographic attributes, an organization could interpret a customer’s interests, behaviours, channel preferences and interactions to determine how a product or message should be presented.
Mody pointed to platforms such as Adobe Journey Optimizer and Adobe Real-Time CDP as part of an ecosystem that is becoming increasingly contextual and agent-driven. Adobe describes Journey Optimizer as a customer journey management solution for orchestrating personalized engagement across channels, while Real-Time CDP is designed to harmonize and activate customer data.
In this model, customer experience platforms would no longer operate only as orchestration tools. They could become an underlying environment through which agents gather context, interpret behaviour and coordinate interactions.
Mody expects this to make digital communication feel closer to a human conversation, rather than an automated system delivering messages based on static rules.
Human intervention becomes more important as AI scales
The continued need for human supervision does not make agentic AI redundant, Whiteside argued. It changes the nature of the work performed by people.
An individual may eventually supervise a large number of agents, directing their work, handling exceptions and verifying that the output meets the intended objective.
In software development, for example, agents could perform much of the code creation between the initial specification and the final quality-assurance process. Humans would still determine what should be built and verify whether the resulting product meets the requirement.
The role of the employee consequently shifts from executing every task to providing direction, judgment and verification.
Mody added that definitions of correctness are often contextual. A system affecting a limited internal process does not carry the same consequences as one affecting thousands or millions of customers.
Organizations must therefore account for broader questions of fairness, responsibility and social impact. These judgments cannot always be reduced to binary technical rules.
For Mody, the future of enterprise AI will continue to require a human-first approach, particularly when decisions involve ethical, cultural or societal considerations.
Data, APIs and verification remain major barriers
Fragmented data is one of the most visible barriers to enterprise-scale agentic AI.
Information may be distributed across numerous systems, each with different access rules, structures and levels of quality. Organizations must determine what data an agent can use, which actions it can take and how those activities comply with privacy and governance requirements.
Mody also highlighted the problem of deterministic verification: how does an organization confirm that an agent has completed the correct task in the correct manner?
Once the number and speed of agent-led activities move beyond what a human can inspect individually, enterprises will require automated validation systems and clearly defined quality gates.
The condition of enterprise APIs presents another challenge.
Agents require stable and consistently defined interfaces to interact with organizational systems. In many enterprises, however, different teams have developed APIs using different naming conventions, schemas and definitions. Agents can struggle when the underlying systems do not share a consistent understanding of the data being exchanged.
Whiteside framed the wider issue as one of trust. Organizations need the ability to audit what an agent has done, identify the information it used and trace the actions it performed.
He suggested that the autonomy granted to an agent should correspond with how easily its actions can be reversed.
An agent can be given greater freedom when an action is inexpensive, low-risk and easy to undo. Actions with significant consequences require tighter boundaries, permissions and human checkpoints.
“Match the autonomy of an AI agent with the reversibility of its action,” Whiteside said.
Enterprises are trusting AI with verification before creativity
One of the clearest enterprise use cases discussed during the interview involved quality assurance for marketing assets.
Whiteside shared the example of an unnamed supermarket brand in the United Kingdom that receives advertising assets from multiple brands. These assets must be checked for brand consistency, legal requirements, advertising claims and regulatory compliance.
DEPT® developed an automated quality-control process involving 25 AI checks, according to Whiteside.
He said the system achieved approximately 95 per cent accuracy during the process, compared with approximately 80 per cent for a manual team reviewing the same assets.
The client was not identified during the interview, so the performance figures remain attributed to DEPT® and Whiteside’s account of the project.
The example reflects a broader pattern in enterprise adoption. Brands may not yet trust AI to independently produce their most important creative work, but they are increasingly willing to use it for checking, validation and quality assurance.
These functions are time-consuming, repetitive and susceptible to human error. Automating them can reduce time to market without removing human judgment from the creative process.
Whiteside also referred to an e-commerce project in which, according to DEPT®’s internal comparison, the team delivered the project 3.8 times faster using agentic methods than the original estimate for a manual approach.
The faster delivery allowed the client to use part of the remaining budget for activities intended to promote the e-commerce platform.
Technical roles will become broader and more interconnected
Agentic systems will also change the skills enterprises require from their employees.
Mody expects the boundaries between roles to become less rigid. Teams will increasingly include people operating as decision-makers, orchestrators and implementers, with responsibilities often extending across those categories.
Technical specialists will need a stronger understanding of the complete delivery process.
A developer, for example, may also need to understand testing, deployment, product management and agile project management. Knowing only one programming language or a narrow part of the workflow may no longer be sufficient.
At the same time, broader knowledge should not replace technical depth.
People supervising AI-generated work must still understand what good work looks like. A developer cannot reliably ask an agent to improve the complexity or performance of code without understanding the underlying engineering concepts.
The ability to evaluate AI output will therefore become as important as the ability to generate it.
The AI-native enterprise will be built around adaptable systems
Predicting a fixed enterprise architecture is difficult because AI technology is evolving rapidly, Whiteside said.
Organizations should instead focus on flexibility. Decoupling systems through APIs can make enterprise infrastructure more adaptable and allow different technologies to interact without becoming permanently tied to one platform or model.
Modernizing data is another “no-regret” investment.
Making organizational data cleaner, more accessible and better governed creates value regardless of which AI system ultimately becomes dominant.
However, Whiteside believes the largest change will involve process reimagination rather than infrastructure alone.
Enterprises need to examine how work moves from an initial input to a final outcome. They must identify which tasks should involve AI, where human judgment remains necessary and what safeguards should govern the workflow.
DEPT® approaches this through process modelling, examining how a client might perform a particular type of work in the future rather than simply adding AI to its current operating structure.
Technology is then selected to support the redesigned workflow.
Mody similarly expects AI to become part of the underlying operating system of a business, much like cloud, network and server infrastructure are today.
Data cycles will become shorter, software will be delivered more incrementally and enterprise applications will evolve continuously rather than only through large monthly or quarterly releases.
He also expects Model Context Protocol-style tool registries and similar integration models to reduce the need for organizations to create a separate custom connection for every platform or application.
What CEOs should prioritise
For CEOs considering agentic AI, Whiteside’s first recommendation is direct: “Find a problem to fix.”
Organizations should begin with a significant business challenge that has a measurable impact. They should not start with a decision to implement AI and then search for a problem that justifies the investment.
His second recommendation is to “get your plumbing in order”.
This includes improving data access, identity management and governance. The underlying systems do not need to be perfect, but they must be sufficiently reliable for agents to access information and perform actions safely.
The third priority is investment in people.
Human involvement will continue to be substantial, but the required skills will change. Employees must be trained to design, supervise, evaluate and improve agentic systems.
Mody added that organizations also require a structured investment plan. Enterprises should define what they are spending, the outcomes they expect and the milestones through which progress will be evaluated.
The roadmap should help leaders decide whether a particular problem requires people, conventional software tools, large language models, smaller specialized models or AI agents.
It should also specify the level of access granted to an agent. An agent permitted to read information carries a different risk from one allowed to modify selected records or make unrestricted changes.
Workflow redesign is the real work
The long-term impact of agentic AI will not be determined only by the sophistication of the models enterprises adopt.
Its success will depend on whether organizations can improve their data foundations, define meaningful guardrails, redesign processes and prepare employees for roles centred on orchestration and judgment.
Agentic engineering, in this sense, is less about removing people from the enterprise and more about changing where human intelligence is applied.
“This isn’t a technology thing. It’s an operating model process change,” Whiteside concluded.
“Redesigning how the workflows work is the work.”
Related Articles
