When intelligence enters the workflow

For most organizations, the first encounter with artificial intelligence is personal. An employee drafts an email faster, summarizes a long document, or asks a question that once required a search. The work improves a little. The risk is contained. The person remains fully responsible for what happens next.

The second encounter is different. The model classifies an incoming request and routes it. It proposes a price, flags a transaction, drafts the reply a customer receives, updates a record, or decides which case a team sees first. Intelligence is no longer beside the workflow. It is inside it, and its outputs begin to shape what the organization does.

That transition is where most of the value is found. It is also where most of the risk begins.

From assistance to consequence

When a model assists a person, its output is a draft. Someone reads it, judges it and decides whether to use it. Errors are caught, or at least owned, by the person who acted on them.

When a model participates in a workflow, its output becomes an event. A ticket moves to a different queue. A supplier invoice is approved for payment. A customer receives an answer. A priority changes. The next person in the process may never see the reasoning, only the result.

It is useful to distinguish four levels of involvement:

• The model suggests, and a person decides whether to act.
• The model drafts, and a person edits and sends.
• The model decides, and a person reviews before the decision takes effect.
• The model acts, and people review a sample afterward.

Each level may be appropriate for a different task. The mistake is increasing the model’s authority without redesigning the work around it.

What the evidence shows

Two of the most rigorous field studies to date point in the same direction, and their details matter.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied more than five thousand customer support agents given access to a generative AI assistant. On average, agents resolved about 14 percent more issues per hour. For novice and lower-skilled agents, the gain was about 34 percent. The most experienced agents saw little improvement. In effect, the system spread the practices of the best performers to those still learning.

A study by Fabrizio Dell’Acqua and colleagues at Harvard Business School, conducted with Boston Consulting Group, found a similar pattern. On tasks the model handled well, consultants completed more work, finished faster and produced higher-quality results. But on a task just outside the model’s capabilities, people using AI were less likely to reach the correct answer.

The researchers described this as a jagged technological frontier. Tasks of apparently similar difficulty can fall on opposite sides of it, and the boundary is not visible from the output alone.

The frontier runs through your processes

AI should be judged task by task, not department by department. A single workflow may contain steps the model performs reliably, steps it handles well with review and steps where it should not be involved at all.

A practical example:

Mapping the frontier requires testing on the organization’s own cases, including the difficult ones. Vendor demonstrations are built from tasks that sit comfortably inside it.

Oversight must be designed, not assumed

“A human in the loop” is often offered as the safeguard. In practice, review weakens when volume is high and the model is usually right. People approve what they are shown, especially when the queue is long and disagreement requires extra work. Review becomes a formality without anyone deciding that it should.

Effective oversight has to be designed. Reviewers need to see the evidence behind an output, not only the output. Disagreement should be easy to record. Less certain cases should receive stronger review, and completed work should be checked to see whether reviewers are actually catching errors.

The NIST AI Risk Management Framework, published by the United States National Institute of Standards and Technology, organizes this work around four functions: govern, map, measure and manage. Its emphasis on defined roles and responsibilities for human oversight applies to organizations of any size.

Ownership does not transfer to the model

When a model routes a request to the wrong team or sends an inaccurate answer, someone must own the outcome. That responsibility usually remains with the organization that owns the process.

This has practical consequences. The process owner must understand what the model does, approve where it is used, and decide how much autonomy it receives. The organization must be able to reconstruct what happened: which inputs were used, which version of the model or instructions was in place, and who reviewed the result.

Data becomes behaviour

Once intelligence enters the workflow, the quality of the organization’s information stops being a reporting concern and becomes an operational one. An outdated policy document becomes an incorrect answer to a customer. An inconsistent product definition becomes an incorrect recommendation. A missing field becomes a misrouted case.

A model is also only as useful as the context it can reach. Answering a customer question well may require order history from one system, entitlement from another and service notes from a third. Where those systems are poorly connected, the model either works with partial information or recreates the manual work between systems in a new form.

Preparing for AI often means getting the basics right first: clear ownership, reliable information, connected systems and fewer conflicting sources.

Measure the workflow, not the model

Model benchmarks describe capability in general. Organizations need to know whether a specific workflow has improved. The measures should describe the work, end to end.

• Cycle time from request to resolution.
• Errors that reach customers, suppliers or financial records.
• How often reviewers change or reject the model’s output.
• Share of cases escalated to a person, and why.
• Time for a new employee to reach full proficiency.
• Cost per completed case, including review and correction.

Two of these deserve particular attention. An override rate near zero may indicate an excellent model, or reviewers who have stopped looking. And the finding that AI helps novices most carries a second question: whether people are still learning the work, or only learning to accept the output.

The first question is often, “Where can we use AI?”