An AI agent won't fix a broken business process

Agents already handle well-scoped, multi-step tasks remarkably well. Yet redesigning processes and tasks rarely fails on technology. It fails because no one has decided how work flows, who gets to decide and which data can be trusted. That is why the first sizeable agent initiative is not primarily a technical project. It is a way to make the missing decisions visible.

McKinsey's State of AI 2026 highlights a small group of organizations that capture enterprise-level value from AI. Three things set them apart: they pursue growth or innovation, not just efficiency; they redesign workflows instead of adding AI to the old model; and they tie leadership and operational discipline to execution. Nearly three in four top performers say they have fundamentally redesigned their workflows. Among the rest, the share is about one in four.

In Deloitte's research, only 16 percent of leaders consider their processes ready for agents. More than 70 percent have deployed AI without redesigning work, workflows or decision rights. Individual efficiency gains disappear before they reach the income statement.

“

An agent scales the way of working it is given. If that way of working is unclear, the ambiguity scales.

Meanwhile, at a manufacturing company

Picture an ordinary situation at a manufacturing company. A supplier reports that delivery of a critical component will be a week late.

The buyer reads the email, checks the order in the ERP and asks production what the delay affects. The warehouse is asked for the actual stock level. Sales assesses customer commitments. A substitute part is hunted down in spreadsheets and emails, and by asking an experienced colleague.

This would be a natural task for an agent: detect the exception, combine the data, work out the options and prepare the next steps.

But the agent cannot resolve the four questions the execution depends on:

Can the component be substituted?
Who approves the extra cost?
Whose customer gets priority?
Is the stock level in the system correct?

These are organizational decisions. If they have not been made, the agent produces proposals that look good and are still wrong.

Technology is enough for a scoped workflow, not for running the chain

The building blocks exist: retrieval, tool use, multi-step workflows and human approval points. Reliability still has to be assessed task by task. In long chains, errors compound.

Anthropic's Building Effective Agents remains a useful design principle: keep the solution simple, make the steps transparent and define the tools precisely. Add complexity only when it improves the outcome. Humans must stay in control.

Many say "the technology is ready", but that needs a caveat. A scoped workflow can be built now. Autonomously running an entire supply chain is a different level of requirement, and it should not be the first target.

Industry data shows the same. In an Economist Impact study, only one company in ten has integrated agents extensively into its supply chain. In Gartner's CSCO survey, 56 percent cite integration with legacy systems and processes as a key roadblock. The differentiator is not the model. It is whether the operating model is redesigned or the agent is bolted onto the old structure.

Organizational debt is the interest rate on scaling

Technical debt builds up when a quick fix is never cleaned up. Organizational debt builds up the same way in how work is divided, how decisions are made and how things get done. People learn to patch it: they call the right person, interpret incomplete data and remember the exceptions.

Recent research describes the same phenomenon as process debt. When workflows are undocumented and inconsistent, an agent does not behave predictably in production. The debt is not a side issue. It is the reason pilots stay pilots.

“

A quick test: ask three people from different departments who decides on expedited shipping and on what grounds. If the answers differ, there is debt.

In the supply chain, the debt shows up in ways like these.

Ownership breaks at department boundaries

Procurement buys, production manufactures, sales owns the customer. Resolving an exception end to end relies on everyone going the extra mile.

Decision rules are tacit knowledge

An experienced buyer knows when expedited shipping is worth it. The rationale is not written down and the authority is not defined.

Metrics pull in different directions

Procurement minimizes purchase price, the warehouse minimizes working capital and sales minimizes lead time.

Data has no shared meaning

In one system, delivery date means the dispatch date; in another, the date the customer receives the goods.

If an agent is built on top of this, it optimizes whatever metric it is given. With procurement's metric, it picks the cheapest substitute part. It cannot see that the lead time breaks a customer commitment if that commitment lives in a salesperson's email. The buyer approves the proposal, and the mistake repeats faster and more consistently than before.

Some of the tacit rules can be inferred from historical data. The difference between inferring and approving matters. When the agent shows that expedited shipping was chosen in 40 percent of cases with no recorded rationale, someone has to decide whether that is a rule going forward or a mistake. This is exactly the value of the first initiative: vague debt turns into a decision queue.

The first target is a workflow that forces decisions

A good first target is a recurring, business-critical workflow whose outcome can be verified and whose errors can be contained. It needs an owner with the authority to change how the work is done.

McKinsey describes the transition as three horizons: adoption, automating existing work, and reinvention, where roles, workflows and the operating model are redesigned. Enterprise-level value clearly shows up only in the third. Leaders who report redesigning workflows are more than five times as likely to report enterprise-level value as those who leave workflows unchanged.

In the supply chain, practical first targets include processing order confirmations, resolving delivery exceptions and enriching item master data. They all involve combining scattered information and preparing a decision.

Assess candidates with these questions:

How much does the problem cost?
Is the data available?
Can success be verified?
What would an error lead to?

Scope the first implementation tightly: one plant, one component group, an impact assessment of delivery delays and a proposed resolution for the buyer to approve. Automatic supplier switches and production plan changes are later extensions, assessed separately.

At the same time, ask whether all the current steps are needed. If three people assemble the same status picture for different meetings, a shared view doesn't make that work more efficient. It removes the work.

Divide work at the level of tasks and decisions

Don't divide work by role. Divide it by what is routine, what calls for judgment and what is accountability.

Step	Agent's task	Human accountability
Detecting the exception	Interpret the supplier's message and compare it with the order.	Define which exceptions require handling.
Assessing the impact	Combine order, stock and production data and flag gaps.	Resolve conflicts and assess impacts the data does not capture.
Preparing options	Retrieve approved alternatives and compile a cost and lead-time comparison.	Approve technical deviations and customer prioritization.
Execution and follow-up	Make authorized entries and track the agreed resolution.	Decide on actions beyond delegated authority and own the outcome.

A language model is not the right place for inventory calculations, cost comparisons or planning constraints. Those are handled with rules, calculation, machine learning and optimization. The agent ties them into the workflow.

Approval is a step in the work, not a ceremonial click. Define who reviews, what they see and how quickly they must respond. If every case ends up in the same queue, the bottleneck moves rather than disappears.

Governance has to be designed from the start. Gartner estimates that by 2028 a large enterprise may have tens or even hundreds of thousands of agents. Only 13 percent of organizations consider their current governance adequate. The first initiative must not spawn an unmanaged agent sprawl.

Legacy isn't the first obstacle, data often is

An old ERP may prevent real-time reads or safe updates. That still doesn't mean the whole landscape must be replaced before the first implementation. Start with read access and entries made by people. Open write access only once access rights, validations and the audit log are in place.

Content quality is often a bigger obstacle than the interface. Identifying a substitute component requires reliable data on specifications and approved compatibility. A similar-looking item is not an interchangeable part.

Accountability does not transfer to the agent

The process owner is accountable for service levels, costs and outcomes. IT is accountable for technical operability and access management. Data owners are accountable for the quality of critical data. An agent's proposal does not shift accountability.

The expert's work shifts from gathering information to assessing exceptions, developing decision rules and working with suppliers. The skills needed are checking sources, spotting missing information and challenging a plausible-sounding proposal.

Tool training won't build this. People need time to describe their current work, design the new workflow and rehearse failure scenarios too. The change in roles, and how freed-up time will be used, must be communicated openly.

Action plan

Pick one workflow and name an owner
Define where it starts, where it ends and the business benefit you're after.
Map the real work
Include the emails, spreadsheets, waiting times and informal decisions.
Design the new division of work
Remove unnecessary steps and define the agent's authority, approvals and exception handling.
Fix the critical prerequisites for this workflow
Make sure the data, connections, access rights and behavior during disruptions are in order.
Advance based on evidence
Test with past cases, compare proposals with real decisions and move to limited use. Increase autonomy only based on results.

Processing time alone is not a sufficient metric. Also track decision quality, delivery reliability, extra costs and the review and correction work people do. A faster proposal is a benefit only when the whole workflow improves.

Supply chain AI readiness doesn't come from a model. It comes from bringing goals, data and decision rights together in the same workflow. These decisions can be made today. The first agent will show which of them are missing.