Organisations are Not Inherently Built for Agents, Why Must Redesign Workflows

A recent IDC study, commissioned by AWS, caught my eye. It said that 97% of organisations expected to fully deploy agents by 2027, and yet just 3% have managed to redesign workflow to scale agentic AI across departments. This expectation-reality gap is a matter of concern for many in boardrooms. A common question is often asked: What are we missing?

HFS offers a likely answer after surveying the Global 2000 organisations in 2025: leaders mistakenly view their initial successes in AI adoption as evidence of meaningful enterprise AI transformation, but in reality they're relatively simple AI copilots and task-level automation, not more complex, enterprise-wide AI initiatives such as AI governance, and human-agent collaboration

I believe there is more to it than meets the eye. The key to this disconnect lies in the core structure of an enterprise.

Overlaying AI, eating away at efficiencies

In the race to upgrade operations, meet customer expectations, and adhere to regulatory requirements, organisations are trying to find the quickest way to jump into the AI bandwagon. Most enterprises are bolting AI onto structures that were inherently designed for human-only work.

Of course, the pilots were impressive. But on implementation, the economics did not hold up. McKinsey confirms that while 78–80% of companies now use generative AI, around 80% report no material earnings impact from it. The consulting firm calls this the “gen AI paradox” and links it to the dominance of general-purpose copilots and chatbots over more targeted, function-specific applications. Dig deeper, and you realise that without redesigning the workflows for Agentic AI organisations are deploying agents into workflows, roles, and accountability structures that were never designed for machine autonomy.

So, should organisations that are not built for agents stop upgrading? Short answer: No. It is too soon to assume that these organisations should close or remain stuck with outdated systems.

Before you begin assessing AI systems and models, consider the core issue of running agent-native work on human-only operating systems.

Converting human-centric workflows to agent-native processes

Until now, enterprises engineered operations around the assumption that humans would perform every step – gather data, make judgements, update systems, and communicate the results. With agentic AI, this assumption will create duplications and redundancies. An end-to-end process redesign is an ideal way forward for AI governance, and human-agent collaboration

BCG reported that redesigning end-to-end processes increased productivity three times, lowered cycle time by 80%, and minimized costs by over 60%. A European bank that redesigned retail lending end-to-end around an agentic workflow achieved over 90% end-to-end automation for consumer loans and 50%+ productivity gains across retail lending.

Such an elevated level of autonomy necessitates more rigorous initial preparation:

• Map the process from the desired outcome backwards.
• Identify every decision point, dependency, and exception.
• Define which decisions an agent may make independently.
• Establish when human review is mandatory.
• Specify which systems an agent may access or update.
• Design escalation paths before deployment.

The objective is to decide which activities should disappear, which should be performed by agents, which require human judgement and how the entire process should be coordinated.

Clearing blurred lines of AI governance and human-agent collaboration

Defining decision rights to humans was straightforward. Humans would review dashboards, apply their judgement, and sign off on exceptions. With agents, especially multiple specialised ones, that model becomes more complicated. Leaders would either stifle autonomy by over-controlling or create risks otherwise.

To overcome this gridlock, organisations must stop relying on static authority structures and adopt dynamic governance that continuously coordinates who decides, when, and on what basis across both human and artificial deciders.

Deloitte’s RACI-O matrix provides an organisational mechanism for this. AI agents can share responsibility for execution or be consulted for analysis, while accountability and override rights remain with designated human owners.

In practice, this means every agent needs a human owner, every class of decision needs defined autonomy boundaries, and every failure mode needs a clear escalation path.

The challenge, therefore, is not simply where to deploy an agent or how much autonomy to give it. Organisations must rethink how work is designed and managed when humans and agents operate together. Processes, roles, decision rights, governance, technology, and organisational culture are interconnected; changing one without addressing the others can simply shift inefficiencies rather than eliminate them. For enterprises already managing complex legacy environments, doing this effectively requires experience in redesigning the operating model itself.

Choosing the right partner is a strategic decision

Redesigning the operating system of work is a complex undertaking. It requires organisations to consider processes, data, decision rights, roles, controls, governance, and culture together. Trying to do all this at such a massive scale with internal teams already stretched across BAU, compliance, and legacy transformation is a recipe for slow, patchwork progress.

This is why choosing the right partner is a strategic decision, and the right partner for agentic AI should cover the following four imperatives:

• Co-design end-to-end workflows that are agent-native from the start.
• Redefine roles and outcomes, helping leaders move from managing headcount to managing hybrid human-agent teams with clear accountability and override authority.
• Build governance and architecture for agents, including agent registries, observability, evaluation frameworks, and compliance controls that scale beyond pilots.
• Bring cross-industry patterns for what works (and what does not) in human-agent team design, so you do not have to learn every lesson the hard way.

The value of agentic AI will depend on how effectively it redesigns work around it. The right partner will turn agentic AI from a collection of isolated pilots into a fundamental redesign of how work gets done.