Kostas Karolemeas
Writing archive

· agentic AI

Agentic AI Will Be Won in the Workflow

The next phase of enterprise AI will reward teams that combine agent technology with process expertise, reliable execution and measurable business outcomes.

agentic AIenterprise AIworkflow automationbusiness transformation
Teal, copper and coral painted planes overlap across a cream ground, with translucent aqua layers and a small cobalt accent.

I expect agentic workflows to be one of the defining themes of 2027. The test will be whether organizations can turn increasingly capable AI into better operations.

That was the argument behind my comment on Tom Read's LinkedIn post, which discussed Apollo's assessment of AI's financial payoff. Read raised a useful distinction between systems that generate responses and systems trusted to take action inside a business.

My view is that the next phase will reward teams that combine AI engineering with a deep understanding of business processes. Model capability matters. So does knowing where work waits, why exceptions occur, who can authorize a change, and what a completed case is worth.

The vendors and internal teams that bring those capabilities together will be best placed to turn experimentation into transformation.

What the payoff debate actually tells us

In an October 24, 2026 note, Apollo chief economist Torsten Slok argues that the financial benefits remain concentrated. His charts contrast rising revenue per worker and profit margins among the Magnificent Seven with flat measures for the remaining S&P 500 companies, alongside a broader productivity series.

That is a reason to demand better evidence from AI programs. It does not, by itself, establish that every deployment outside those seven companies has produced no benefit. The S&P 493 is also a group defined by excluding seven companies, rather than a clean classification of all non-technology businesses. Aggregate financial results cannot isolate AI's contribution from everything else affecting a firm.

Evidence at the task level answers a narrower question. In Generative AI at Work, published in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents. Access to a conversational AI assistant increased issues resolved per hour by 15% on average, with substantial differences across workers. The authors explicitly caution against generalizing from one tool, occupation and firm.

Both observations can coexist: a useful improvement within a particular workflow and no clearly established economy-wide payoff in the indicators under discussion. Neither establishes when agentic systems will deliver broad financial returns.

My 2027 expectation is a judgment about where implementation effort should go. It is not a date that these studies prove.

The interface is only the beginning

Conversational interfaces can be valuable. The customer-support research is evidence of that. A chat window can also be the entry point to a system that performs substantial work behind it.

The limitation appears when a vendor's delivery stops at producing an answer, while the business still depends on someone to interpret it, locate the right record, transfer the information, obtain approval and check that the change succeeded.

An answer may save time at one step. Whether that improves the overall process depends on what happens next.

Consider an invoice discrepancy. A chatbot might explain why an invoice does not match a purchase order. That explanation helps the person investigating it. A more complete system could assemble the supporting records, identify the type of discrepancy, request missing evidence, route the case to the right approver and record the authorized resolution.

This is an illustrative design, not a claim about a deployed system or a measured saving. Its purpose is to show how much work sits between an explanation and a business outcome.

Process expertise determines where autonomy belongs

The invoice example raises questions that a model demonstration cannot settle.

Which record is authoritative? Is the discrepancy a genuine exception or a timing difference? Who can approve a price change? What happens when the goods receipt is missing? If a system times out after submitting an update, how do we check whether it succeeded before trying again?

Those questions shape the architecture. They also require the people who understand how the operation actually works.

Anthropic's guidance on building effective agents, originally published in December 2024, distinguishes predefined workflows from agents that dynamically choose their steps and tools. It recommends starting with the simplest approach and adding complexity when it improves outcomes. The tooling has evolved since publication; that architectural distinction remains useful here.

I would use deterministic rules for a known approval threshold, conventional automation for a reliable record update, and a model where interpreting messy evidence or selecting an investigation path adds value. An agent can investigate a discrepancy without receiving authority to release a payment.

The practical skill is designing that combination: flexible interpretation inside explicit permissions, durable process state and clear responsibility for exceptions.

Measure the completed case

For a pilot, I would begin with one recurring process and a baseline. Record how long a case takes from arrival to resolution, how much human effort it consumes, how often it needs rework and whether the result meets the required quality standard.

Then compare the redesigned process with the existing one on a comparable mix of cases. A staged rollout or a suitable comparison group is more informative than a demonstration using only easy examples. If the new system gets the simplest cases while people retain every difficult exception, an average handling-time comparison can mislead.

For the invoice workflow, the scorecard could be compact:

  • Completion: the share of eligible cases correctly resolved, including cases escalated to a person.
  • Time: end-to-end resolution time and human minutes per case, including review and correction.
  • Quality: incorrect updates, reopened cases and unresolved exceptions.
  • Cost: total operating cost per correctly completed case, including model usage, integrations, supervision and rework, with implementation costs reported separately.

The business also needs a plan for any capacity released. Saving ten minutes does not automatically reduce spending or increase revenue. It may reduce a backlog, improve service, absorb additional demand or avoid future hiring. Each is a different outcome and should be measured as such.

This is the evidence I would want before expanding a pilot: a better process under realistic conditions, with the costs and remaining human work visible.

Transformation requires complementary work

There is a useful economic explanation for why buying technology and realizing its benefits can be separated by considerable effort. In The Productivity J-Curve, Brynjolfsson, Daniel Rock and Chad Syverson describe how general-purpose technologies require complementary investments in processes, products, business models and human capital. Their work also examines how intangible investments complicate productivity measurement.

That framework does not guarantee a return on today's AI spending. It identifies work that a technology purchase alone does not accomplish.

In an enterprise deployment, I would expect that work to include cleaning up responsibilities, connecting systems, training people, removing unnecessary steps and agreeing on what success means. Some improvements may come from ordinary process redesign. They still count as business value, but we should be honest about what produced them.

What I expect to distinguish the winners

The strongest AI teams in 2027 will need to explain more than their choice of model or agent framework. They will need to show how their systems improve a real operation and how those improvements survive exceptions, operating costs and everyday use.

For buyers, that means asking a vendor to walk through a difficult case from beginning to end: the evidence it reads, the actions it is permitted to take, the human decisions it needs and the result it can verify.

For builders, it means working closely with process owners and frontline teams. Their knowledge belongs in the design from the beginning.

That is where I see the opportunity in agentic workflows. The teams that know how to turn AI capability into dependable operational value will have the strongest claim to having delivered transformation.

Sources