Autonomous Workflows and AI-Native Organizations

Picture a large financing business a generation ago. A field salesperson phones in a request to finance a customer’s purchase. Someone logs it on paper. The paper goes upstairs to credit, then to a department that adjusts the loan terms, then to a pricer, then to a clerical group that writes the quote letter. The whole trip takes six days on average and sometimes two weeks, long enough for the customer to find another lender. The company tries a fix: a control desk that tracks every request, so the salesperson can at least learn where the deal is. Turnaround grows to more than seven days.

This was IBM Credit, and the case became famous because of what two senior managers did next. They walked one request through all five steps and asked each office to process it at once instead of leaving it in a pile. The actual work took ninety minutes. Everything else was waiting between desks. The managers drew the conclusion that matters for this chapter: if every person in the chain became twice as productive, turnaround would fall by only forty-five minutes¹.

Replace “twice as productive” with “an AI assistant on every desk” and the arithmetic has not changed in more than thirty years. IBM Credit replaced its chain of specialists with one generalist, a “deal structurer”, who handled each request from start to finish with a new system to support the work and a small pool of specialists for the hard cases. Turnaround fell from seven days to four hours, and the number of deals handled rose a hundredfold, with slightly fewer people¹.

AI-enabled is not AI-native

Many large organizations in 2026 are AI-enabled. Their people have assistants, their teams have pilots, and their software vendors have added AI features to the tools they already used. Individual steps are faster. The structure of the work, meaning who hands what to whom and who waits for whose approval, is the one that existed before.

An AI-native organization starts from a different question: if AI had been available when this work was first designed, how would we design it? The answer usually removes steps rather than speeding them up, gives one owner the whole outcome, lets AI carry the routine path from start to finish, and sends the unusual cases to people who are equipped to judge them. AI Is Changing Everything called this the difference between automating a task and redesigning the work, and AI Use Case Discovery and Design traced the instinct back to 1990s reengineering. This chapter does not re-teach reengineering. It asks what changes when AI, rather than a person, can carry the routine path, and what that means for workflows and for the organization around them.

The evidence says the organization, not the tool, is now the constraint. In Microsoft’s 2026 survey of 20,000 knowledge workers, all of whom use AI at work, organizational factors such as culture, manager support and talent practices explained more than twice as much of the impact people reported from AI as individual factors did, 67 percent against 32². Those shares describe what drives reported impact, not how many workers agree with anything. The survey is a vendor’s and the impact is self-reported, but it points the same way as McKinsey’s finding, discussed in AI Use Case Discovery and Design, that workflow redesign is the practice most strongly linked to earnings impact from generative AI³.

AI-native is not a synonym for maximum autonomy, and it is not a declaration a company makes once. The goal is the best combination of AI execution, human judgment and authoritative systems for each workflow. Some workflows will rightly stay mostly human.

Four shapes of a workflow

It helps to describe a workflow by who coordinates it. In an AI-assisted workflow, a person still runs every step and AI drafts, summarizes or recommends inside some of them. In an AI-automated workflow, an event triggers AI to perform a bounded task, such as reading a document or classifying a request, and a system acts on the result. In an AI-orchestrated workflow, AI coordinates several steps across several systems: it gathers the context, checks the policy, prepares the action and routes the work. In an AI-native workflow, the process itself has been redesigned around those capabilities, so steps that existed only to move information between people have disappeared.

This ladder describes the shape of the work, not the authority of an agent. Agentic AI and Autonomous Actions set out how to grant authority, with least privilege and with approval where impact is high and reversal is hard, and those controls apply on every rung. The two questions are separate. A workflow can be fully redesigned and still require a person to approve every payment.

Nor is the top rung the goal everywhere. IBM Credit ended up with three versions of its process: straightforward cases handled entirely by computer, medium cases by a deal structurer, and difficult cases by a deal structurer working with specialists¹. Analysts expect the routine share to grow. Gartner, in a June 2025 release reported by the trade press, forecast that by 2028 at least 15 percent of day-to-day work decisions will be made autonomously by agentic AI, up from none in 2024⁴. That is a forecast, and it still leaves most decisions with people. Fully orchestrated workflows are, for now, more often designed than documented: published results at scale remain few.

The time sits between the steps

Why do assistants on every desk change so little? Because in processes like IBM Credit’s, most of the elapsed time is not work. It is waiting: in a queue, in an inbox, for an approval, for the next team to pick the case up. Suppose a process takes 20 days end to end and contains two days of actual work. An assistant that cuts 30 percent off the working time of every step removes 0.6 of a day, and the customer still waits 19.4 days. Redesigning the flow so that one owner carries the case, with AI gathering and drafting, removes most of the queues.

The queues exist for a reason, and the reason is the organization chart. In 1968 the computer scientist Melvin Conway observed that organizations which design systems produce designs that copy their own communication structures⁵. Business processes follow the same law. A process that crosses five departments has five steps and four handoffs because there are five departments. Put an AI assistant inside each department and the AI inherits those boundaries. The handoffs survive, now between faster desks.

Redesign therefore starts by asking four questions of each step. Which steps should disappear, because they exist only to move or re-enter information? Which should be combined under one owner? Which can AI carry on the normal path? Which genuinely require human judgment? If a process is broken, adding AI produces a faster broken process. And because the queues follow the structure, removing them usually means changing the structure too.

Design the exception path on purpose

An AI-native workflow is built around a simple operating pattern. AI handles the normal cases, verified by rules and checks, and the cases that do not fit go to people. That pattern only works if four things are defined before autonomy is raised: what counts as normal, what counts as an exception, how an exception is escalated and to whom, and when the workflow must stop altogether, for example when inputs look unfamiliar or a check fails repeatedly.

The exception path is not a leftover. When AI takes the easy cases, the human queue is made of the hard ones, so the people who staff it need more skill, more context and more authority than the people who used to process everything. They also need practice: Operational and Workforce Risk showed how skills decay when people only supervise, and Data, Integration and Operational Costs showed how exception rates drive the real cost of a workflow. A redesign that counts the cases AI handles and ignores the queue it leaves behind has designed half a process.

AI orchestrates; systems of record keep the truth

A common misreading of AI-native is that AI replaces the ERP, the CRM, the finance ledger or the safety database. It does not. Those systems of record hold the authoritative state of the business, carry the audit trail and enforce many of the controls that regulators and auditors rely on. In an AI-native design, AI becomes a coordinating layer above them. It reads from them, prepares actions, routes work between them and writes results back, through the same permissions and records as any other user.

Two consequences follow. First, an orchestrated workflow is only as good as the data and connections beneath it; Multimodal and Agentic AI explained the choice between connecting through interfaces and acting on screens, and From Copilots to AI Agents gave the test of who actually acts in the system of record. Second, a failure in a layer that coordinates many systems can spread across all of them, so monitoring, boundaries and rollback matter more as orchestration grows, not less.

The workflow should then be measured as a workflow. The useful numbers are end to end: cycle time, the share of cases completed correctly the first time, the exception rate, the cost per completed case. Counting assistants deployed or agent runs says nothing about whether the customer, the patient or the supplier is better served.

The organization follows the workflow

If the queues follow the structure, a redesigned workflow needs a redesigned structure. Roles move from doing each step toward owning outcomes, verifying AI output and handling exceptions. In Microsoft’s 2026 survey, 86 percent of those AI users said they treat AI results as starting points rather than finished products². Managers start to plan capacity that includes both people and AI. Microsoft’s 2025 report proposed a “human-agent ratio” as a management metric, and 45 percent of the leaders it surveyed named expanding team capacity with digital labor as a priority for the next 12 to 18 months⁶.

Some companies have changed the top of the structure as well. In 2025 Moderna merged its human resources and digital technology functions under one Chief People and Digital Technology Officer, describing the aim as architecting the flow of work and replacing separate workforce and technology planning with a single work plan⁷. It is one company’s choice, not a template, but it shows the logic: when work is done by people and AI together, planning one without the other leaves gaps.

AI-native redesign repeats a promise made once before, and it can repeat the failure. Reengineering in the 1990s promised redesign around a new technology, and one of its originators, Thomas Davenport, explained why so many programs disappointed in an article called “The Fad That Forgot People”. In a 1994 survey he cited, 73 percent of companies had used it to eliminate an average of 21 percent of jobs, and 67 percent of 99 completed programs were judged mediocre, marginal or failed⁸. Redesign done to people, as a headcount program, failed. Redesign done with people, who know where the exceptions hide, stands a better chance, and in much of Europe the law requires their representatives to be involved.

Story: the safety cases that kept arriving

What follows is a composite, drawn from patterns common in pharmaceutical safety departments; it describes no single company, and its details are illustrative.

A mid-sized pharmaceutical company’s drug safety department processes reports of suspected side effects from doctors, patients, partners and the literature. Valid cases must reach the EU regulator’s database on time: every suspected serious reaction within 15 calendar days, and non-serious cases from the EU within 90⁹. Volumes have grown for three years. Each case passes through five teams: intake, triage, data entry, medical review and submission. Most cases are routine. The serious ones arrive mixed in with them, and the teams spend their days chasing deadlines.

The head of safety has three proposals on the table. The first puts an AI assistant into each of the five teams; a vendor can deploy it within a quarter, and every team expects to be faster. The second redesigns the flow: AI reads each incoming report, extracts the data, checks for duplicates and drafts a complete case in the safety database, and one case owner carries the case to submission, with physicians assessing every serious case. It would take longer, change roles and need the works council’s involvement. The third, favored by finance, lets AI process and submit non-serious cases with no person involved, because they are the bulk of the volume.

Decide before you read on.

The department chose the second option and refused the third. The first would have made five teams faster and left four handoffs and four queues between them, which was where the days went. The third failed a regulatory test as well as a quality one. The European Medicines Agency expects AI to support adverse event report management, but it remains the company’s responsibility to validate, monitor and document model performance and include AI operations in its pharmacovigilance system¹⁰. A stream with no human in it would also have removed the place where unfamiliar reports, the ones most likely to matter, get noticed.

They started with one source of non-serious reports, ran the new flow alongside the old one, and measured on-time submission, quality-review findings and cost per completed case. The serious stream changed only after the routine one had a record. The former data-entry staff became case owners, and the training plan was built from the exceptions the pilot produced. The lesson is not that pharmacovigilance is special. It is that the option with the most autonomy and the option with the least disruption both kept the wrong thing: one kept no person accountable, and the other kept the old structure.

What this means for leaders

Stop counting assistants and start mapping workflows. Pick a few end-to-end processes that matter to customers or regulators, walk one case through them, and measure how much of the elapsed time is work. Redesign from the clean-sheet question, decide the normal, exception, escalation and stop rules before raising autonomy, and keep systems of record authoritative. Then change the structure to match the workflow, with the people who do the work and, where the law requires it, their representatives. An enterprise will run assisted, automated and redesigned workflows side by side for years. The aim is not to declare the company AI-native. It is to stop treating another copilot as the operating model.

Check yourself

1. Giving every team an AI assistant makes an organization AI-native.MythFact
2. At IBM Credit, most of a seven-day turnaround was spent waiting between departments.MythFact
3. The goal of an AI-native workflow is zero human involvement.MythFact
4. In an AI-native design, AI replaces the systems of record.MythFact
5. Under EMA guidance, a drug company stays responsible for validating and monitoring AI used in pharmacovigilance.MythFact
6. Reengineering in the 1990s failed mainly because the technology was not ready.MythFact

Reflection: one workflow, end to end

What comes next

This chapter looked inward, at how work runs when it is redesigned around AI. The same logic reaches the market. The next chapter, AI-Native Products and Business Models, asks how AI changes not only how a company operates, but what it sells, how customers use it and how the business creates and captures value.