Every enterprise board now expects an agentic AI plan. Far fewer operating leaders can say which workflow should carry it first, and why.
That debate ran through the closed-door roundtable jointly hosted by iOPEX and Soroco in London, with senior operations and transformation leaders from telecom, banking, retail, financial data, and life sciences. They described pressure from the board and a fear of falling behind, with no reliable method for identifying the workflow in which AI will move a number that the CFO tracks.
The market data shows why the question matters. In McKinsey’s 2026 State of AI survey, 40% of large organizations report scaling AI agents, up from 27% a year earlier. Yet the share attributing any EBIT impact to AI has held flat at 37%.
The room’s conclusion was practical. Agentic AI pays off when enterprises decide what to fix before what to build, and when they run agents as an operation after go-live.
What Did Operations Leaders at the Roundtable Agree On?
- The first decision is the constraint: process, system, data, or automation.
- Agents need operational context, so workflow and data readiness come before autonomy.
- Vendor capability claims have converged. Evidence of P&L movement now separates providers.
- At scale, observability and cost per outcome become operating disciplines.
The First AI Decision is a Work Decision
“Which agent should we deploy?” is usually the first question. Leaders at the roundtable argued it should be the second. The first is what actually constrains the work. There are four candidates:
- The workflow is poorly designed.
- The underlying system creates unnecessary effort.
- The data is incomplete or unreliable.
- Automation is genuinely the missing layer.
Only the last one calls for an agent. Automating a badly designed workflow scales the problem.
Redesign is no small ask. A financial data leader noted that changing a workflow in a large enterprise touches controls, systems, and teams that rarely move together. So the choice of workflow needs evidence before it needs budget. McKinsey finds that nearly three-quarters of AI high performers have fundamentally redesigned workflows, against one-quarter of other organizations.
Why Does Observed Work Change the AI Business Case?
Most programs start with workshops, SOPs, system logs, and interviews. These describe how work was designed or how people remember it. Actual execution is messier.
Soroco’s Scout builds a work graph from how people and agents execute work across applications. It exposes effort, friction,, and variation, and simulates the effect of a workflow change before it reaches production. In Soroco’s agentification research, a process documented as one workflow typically runs in four to ten execution variants.
That evidence also sharpens the business case. One UK telecom operator at the roundtable had identified around £6 million in savings opportunity through work with Soroco and iOPEX. The value sat in specific operational changes the team could locate, size and assign an owner to.
Data Readiness Comes Before Autonomy
Once the workflow is chosen, the next constraint is context. An agent may know the documented process yet lack reliable customer data, exception history, system state, policy rules, or the informal judgment experienced staff apply every day.
Soroco tested this gap directly. Agents given only system credentials completed 12% of tasks. Agents given standard operating procedures reached 57%. Agents given the full work graph of how the work actually runs reached 90% (Soroco).
The readiness test therefore covers data quality and operational context together. Autonomy becomes safer when agents know where exceptions occur and when human judgment still applies.
Why Has the Vendor Pitch Stopped Being Useful?
Several leaders made the same point: every AI pitch now sounds alike, and few show meaningful P&L impact. Agents, orchestration, and governance are now the price of entry to the category.
A more useful test asks for operating evidence. Ask for the observed workflow friction and the sized value opportunity, the economic case against alternative fixes, the readiness work done beforehand, and the measured change in cycle time, cost, quality, and customer outcomes after deployment.
Customer experience belongs in that calculation, a point a telecom leader raised directly. An agent that resolves issues faster or prevents service failures can create value even where labor savings are modest.
What Changes When One Agent Becomes Hundreds?
Scale raised two connected concerns. A banking leader asked how to manage agents spread across systems without token costs climbing. A retail leader named observability as the top priority as AI scales. The cost concern is widespread: one in five organizations says AI operating costs, including tokens, constrain their AI use.
Token spend is easy to measure and says little about value. A cheaper agent that fails, escalates, or creates rework can cost more per successful outcome. The measures that matter are cost per completed workflow, exception rate, human intervention, latency, quality,y and movement in the business outcome.
This is the discipline iOPEX calls AgentOps. HFS Research recognized iOPEX in the HFS Horizons: Agentic Services, 2026 report for embedding governance, privacy, and explainability into agent operations through its AgentOps framework. iOPEX runs more than 1,000 agents in production, and its Command Agents execute across enterprise systems, with built-in auditability and human oversight.
What Operating Loop Should Leaders Put in Place?
The model that emerged connects discovery and production into a single loop.
Soroco strengthens the front of the loop by exposing how work really runs and modeling the future state. iOPEX carries the prioritized opportunity into governed production and runs it through AgentOps. Production telemetry then feeds back into discovery, so the workflow continues to improve after go-live.
Their Final AI Question was Economic
Leaders in the room kept returning to economics. A strong AI program can say where it deployed agents first and why that workflow earned priority, what changed in the operation, what it saved or earned, and who owns that result today. McKinsey finds that high-performing AI teams are twice as likely as others to have defined processes for measuring the impact of their AI initiatives.
Programs that cannot answer with evidence are running pilots at scale. Soroco gives leaders the empirical view to choose the right work. iOPEX keeps that work governed and measured once agents go live. Together, the two disciplines give the board an AI result it can audit.
Talk to iOPEX about identifying the workflows in which agents will move your P&L first.
