AI Agents and Intelligent Workflows

Take an illustrative case. Suppose your operations team pilots two AI agents on the same stream of routine requests for a month. Agent A completes nine requests in ten correctly, on its own. Agent B completes only seven in ten on its own and hands the other three to a person, with a note saying what it checked and why it stopped. Asked which one to scale, many leaders pick Agent A. The higher number looks like the better agent.

It may be the worse one. The comparison turns on a detail the headline number hides: what Agent A does on its tenth request. If it acts anyway, and the mistake surfaces weeks later in a complaint, a write-off or a rework queue, then its tenth case is a silent error. Agent B’s three cases are handoffs. A person sees each one on the day, with the context attached. Which agent is better depends on what each of those costs, and in many enterprise workflows a silent error costs far more than a handoff.

The questions that matter are about agents at work inside the enterprise: where they fit, how to delegate to them, and how to judge them. How an agent works as a technology, its loop of planning, tool use and checking, belongs to From AI Assistants to AI Agents in Module 02, and the shift from answering to acting in your systems of record belongs to From Copilots to AI Agents in Module 01. Here the question is managerial. You are about to give a piece of software a job.

Three outcomes, not one success rate

Every task an agent attempts ends in one of three ways. It is done and verified: the agent acted, and a check confirmed the result. It is handed off: the agent stopped and passed the case to a person. Or it is done wrong and unnoticed: the agent acted, the result was wrong, and nobody knows yet.

A single success rate merges the first and the third, because a wrong action and a right one both count as “completed” until someone notices. That is why the first question to ask about any agent is not how often it finishes, but how often it is wrong without saying so, and how you would know. The third number cannot come from the agent’s own log. By definition, nobody has seen those errors. It has to come from sampling: a person regularly checking a random set of completed tasks against what should have happened.

Illustrative arithmetic makes the point concrete. Take the two agents from the opening and give them 10,000 requests a month. Agent A acts on all of them: 9,000 are right and 1,000 are wrong without anyone noticing. Agent B acts on 7,000 and hands off 3,000, and it is not perfect either: suppose 100 of the 7,000 it handles are wrong. Now suppose a silent error costs 300 to find and fix later, in staff time, refunds and goodwill, and a handoff costs 20 of staff time on the day. Agent A’s 1,000 silent errors cost 300,000 a month. Agent B’s 100 silent errors cost 30,000 and its 3,000 handoffs 60,000, or 90,000 a month in all.

The claim needs this framing to be honest. A lower completion rate is not a virtue in itself. Agent B pays 60,000 in handoffs to avoid 900 silent errors, so it comes out ahead only when a silent error costs more than 60,000 divided by 900, about 67. Below that price, Agent A would win as it stands. And if Agent A could hand off its own tenth case instead of acting on it, its 1,000 handoffs would cost 20,000 and it would beat Agent B easily. Agent B wins here because it knows where its competence ends and stops there. Accuracy, Hallucination and Reliability, in Module 06, explains why a system that can decline to answer is more trustworthy. The managerial version is simpler: put a price on a silent error before you compare agents, and the right choice often becomes clear.

Give agents the workflows that suit them

Agents are being tried widely, and mostly not yet at scale. McKinsey’s 2025 global survey, as reported in coverage of it, found 23 percent of respondents saying their organizations were scaling an agentic AI system somewhere, and another 39 percent experimenting. In any single business function, no more than 10 percent reportedly said they were scaling agents¹. Gartner forecasts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value or inadequate risk controls; that is a forecast, not a measurement². One common pattern behind both numbers, and the one a leader can most directly control, is an agent applied to a workflow it does not suit.

Five questions sort the candidates.

The first question often removes the most candidates. If the path is fixed, an agent is the expensive way to do it. Anthropic’s engineering guidance on building agents draws the same line between workflows, where models and tools follow predefined code paths, and agents, which direct their own process. It advises “finding the simplest solution possible, and only increasing complexity when needed”, and warns that agentic systems “often trade latency and cost for better task performance”³. A request that always takes the same five steps should be automated with rules. The agent belongs where the request arrives in messy language and the next step depends on what the last one found.

The third question decides how much the agent can be trusted. If a system can confirm the result, such as a booking that shows as accepted or a record that matches its source, the agent can check its own work before it reports success. If only a person can judge the result, a person has to approve it.

The fifth question is easy to skip. A Stanford study of 1,500 workers across 104 occupations found them positive about automating 46 percent of the 844 tasks studied, most often because it would free time for higher-value work⁴. An agent dropped on a team that did not ask for it is likely to be worked around.

Consistency: the same task a thousand times

A demonstration runs a task once. A workflow runs it thousands of times a month, and an agent can solve a task on Monday and fail the identical task on Tuesday. When a leading model of 2024 was tested on simulated retail service tasks, it succeeded at the first attempt on 61 percent, but on fewer than a quarter of tasks in all eight attempts⁵. Models have improved a great deal since, so treat the figures as a snapshot. The method is what lasts. From AI Assistants to AI Agents showed how small error rates compound across the steps of one task. Repetition is the second multiplier, across the many runs of the same task. When a team or vendor reports a success rate, ask how often the agent succeeds when the same case is run again, on your own cases, not theirs.

Design the handoff, not only the action

If handoffs are what make Agent B worth scaling, their quality matters as much as their number. A handoff that says only “unable to complete” sends the person back to the start. A good one hands over the work done so far.

Two design choices follow. First, the person receiving handoffs needs the skill to resolve them. As agents take the easy cases, the human queue fills with the hard ones, and the people doing the work see fewer routine cases to stay in practice on. Lisanne Bainbridge called this an irony of automation four decades ago: the more a system automates, the more its remaining human tasks depend on skills that automation stops exercising⁶. Second, the agent’s own rules for stopping must be written, not left to the model’s sense of doubt. The rules name the topics, amounts and kinds of customer that always go to a person, whatever the agent’s confidence. Autonomous Workflows and AI-Native Organizations, in Module 10, takes the exception path up to the scale of the whole organization.

Write the agent a mandate

Delegating to an agent is delegation, and good delegation is written down. A mandate has five fields, and each one answers a question an auditor, a regulator or an angry customer will eventually ask.

The outcome must be checkable: “resolve supplier invoice queries” is a hope, while “every query answered with the matching purchase order and a confirmed payment date” is a goal an agent can be held to. Systems follows least privilege: the agent reaches only what the workflow needs. Acts alone and hands off together set its authority. How far up the scale that authority goes, from recommending to operating on its own, is the subject of Agentic AI and Autonomous Actions in Module 06, which also covers how authority is widened on evidence. The owner is the field most often left blank. It should be the business leader whose results the agent affects, not the team that built it.

Story: the fraud system that never handed off

One of the clearest public records of software given a job and the authority to finish it comes from before generative AI. In 2013, Michigan’s Unemployment Insurance Agency switched on a new system, MiDAS, built to detect benefit fraud automatically. In tens of thousands of cases it decided fraud on its own, with no person involved. Its fraud notices, IEEE Spectrum reported, were designed in a way that almost ensured some claimants would admit to fraud by mistake, and some findings rested on missing or corrupt data⁷. A fraud finding meant repaying the benefits plus a penalty of four times the amount, collected if necessary by seizing tax refunds and garnishing wages⁸. MiDAS was rule-based software, not an AI agent. That makes it a cleaner lesson, because nobody can blame a model: what failed was the mandate.

Judged by what it produced, the system was a success. Fraud findings rose about fivefold compared with the old system, and the fund that received the penalties reportedly grew from about 3 million to more than 69 million in little more than a year⁷. Every case it touched counted as finished. None was handed off.

The errors were silent until the people affected forced them into view through appeals and lawsuits. In September 2015, after a federal lawsuit was filed, the agency stopped letting the system decide fraud on its own. A later review, as IEEE Spectrum reported, found that of 40,195 cases the system had decided alone, 85 percent, or about 34,000, were wrong. Of a separate set of 22,589 cases with some human involvement, 44 percent were wrong⁷. In January 2024, a Michigan court gave final approval to a 20 million settlement for more than 3,000 people falsely accused between 2013 and 2015⁹. Australia’s Robodebt scheme, told in AI vs Automation, followed the same pattern.

Read the case as a missing mandate. The outcome, “find fraud”, was counted but never checked. The system could act alone on everything, including penalties, and had to hand off nothing. Nobody sampled its finished cases until the courts did. And the 44 percent is a warning of its own: a person somewhere in the process is not the same as a designed handoff, with the evidence, to someone who has the time and skill to decide. Price the silent errors, and the fivefold rise in findings stops looking like success. An AI agent can fail the same way, faster. The defenses are the ones in this chapter: a written mandate, stopping rules, handoffs with context and a weekly sample of completed work.

What this means for leaders

Treat an agent as a new role in a workflow, not a feature in an application. That means choosing the workflow with care, writing down its authority, designing how it hands work back, and judging it on all three outcomes. The leadership decisions are the same ones you make when you delegate to a person, made explicitly because software will not ask when it is unsure unless you tell it to.

Agents are the strongest form of the lesson that runs through this module: value comes from AI built into real work and measured by outcomes. They do not only inform the work; they do it. That is why the controls named in this chapter, from handoff rules to audits, matter more here than anywhere else in the module.

Check yourself

1. The agent with the higher completion rate is the better one to scale.MythFact
2. An agent’s own logs can tell you its silent-error rate.MythFact
3. If a request always follows the same steps, rules or ordinary automation are usually the better choice.MythFact
4. An agent that succeeds on a task once will succeed on it every time.MythFact
5. A good handoff includes what the agent checked and why it stopped.MythFact
6. The team that built the agent should own its results.MythFact

Reflection: price the tenth case

What comes next

This module looked at where AI creates value; every chapter also touched on what happens when it is wrong. The next module makes that the subject. The AI Risk Landscape opens Module 06 by mapping where AI risk comes from and why it depends on how a system is used, not only on how accurate it is.