Agents force process redesign

# Agents force process redesign The first wave of corporate excitement around agentic artificial intelligence treated it like a smarter chatbot. Many companies assumed they could bolt an autonomous system onto an old workflow, add a few prompts, and watch the process run itself. The early results are showing that this is not how the technology behaves in practice. Reporting published on April 7 and April 8, 2026, by *MIT Technology Review* and *TechRadar* describes a different lesson: when software can plan, act, and hand work to other systems, the workflow itself has to be rebuilt around that autonomy rather than lightly upgraded. (technologyreview.com) That shift starts with what “agentic” actually means inside a business process. A large language model answers a prompt. An agentic system is expected to do more: interpret a goal, choose steps, call tools, wait for results, recover from failure, and decide when to ask a person for help. In other words, it behaves less like a search box and more like a junior operator with access to software. Once companies ask a system to operate instead of merely suggest, the weak spots in the old process become obvious. (technologyreview.com) Traditional enterprise workflows were usually designed for either people or rigid automation. A person could notice ambiguity, chase down a missing approval, or stop when a case looked unusual. Conventional automation, by contrast, worked best when every step was predefined and every exception was rare. Agentic systems sit awkwardly between those models. They are flexible enough to take initiative, but not reliable enough to be trusted without boundaries. That is why simply inserting them into a legacy process often creates confusion instead of speed. (deloitte.com) The most common failure point is orchestration. In this context, orchestration means the control layer that decides what the agent can do next, what tools it may call, how long it should wait, what happens if a step fails, and when a human must review the outcome. If that control layer is vague or missing, an agent can move quickly through easy cases and then stall, loop, or take actions nobody intended when the real world gets messy. (deloitte.com) Approvals are another pressure point. In many companies, approvals were designed for human workers who understand policy and can summarize edge cases for a manager. An autonomous agent changes that rhythm. It may produce work at machine speed, but the organization still has legal, financial, and compliance checkpoints that move at human speed. If those checkpoints are added as an afterthought, the process jams at exactly the point where risk is highest. (technologyreview.com) Exception handling turns out to be even more important than the happy path. A procurement request with complete data is easy. A procurement request with a missing contract clause, a supplier mismatch, and a budget conflict is where the design either holds or breaks. Human teams have long relied on tacit knowledge to deal with those cases. Agentic systems need those judgment points made explicit: what counts as normal, what triggers escalation, and who owns the decision when the model is uncertain. (technologyreview.com) That is the core argument in the new reporting. *MIT Technology Review* says the operating model has to shift toward “humans as governors and agents as operators,” with people setting goals, policy constraints, and handling exceptions rather than manually driving every step. *TechRadar* similarly argues that the hard part is not the model itself but the organizational redesign needed to support it. The headline is not that agents are failing. It is that old processes are failing when asked to host agents. (technologyreview.com) This is why many pilot programs look impressive in a demo and disappointing in production. A pilot often succeeds on a narrow set of clean tasks with close supervision and a forgiving audience. Production work adds delays, conflicting data, external systems, audit requirements, and ambiguous cases. The more autonomy a company gives an agent, the more it needs durable controls for retries, handoffs, and oversight. Without those controls, the agent is not really autonomous; it is just pushing hidden cleanup work onto employees downstream. (deloitte.com) The redesign challenge also changes who inside the company owns success. This is no longer just a model team problem. Platform teams have to turn recurring patterns into product primitives: delegation rules, retry logic, escalation paths, approval gates, observability, and audit trails. Those are not cosmetic features around the model. They are the infrastructure that makes autonomous behavior usable inside a real business. (deloitte.com) That pushes companies toward a more explicit map of work. Instead of documenting only the ideal sequence of steps, they have to document uncertainty itself. Which decisions can be automated with policy constraints? Which tasks require two systems to agree? Which cases must stop for human review? Which failures should trigger a retry, and how many times? Process design used to focus on efficiency. Agent-first design has to focus on controllability as well. (technologyreview.com) Consultancies and enterprise architects have been moving toward the same conclusion. Deloitte’s late-2025 guidance on agentic orchestration emphasizes governance and operational controls, while Amazon Web Services has argued that autonomous capabilities require architectural changes closer to distributed systems design than prompt engineering. Capgemini has gone further by framing the shift as “zero-based” process redesign, meaning companies should rebuild some workflows from the outcome backward instead of preserving inherited steps that only made sense for human labor. (deloitte.com) The practical implication is blunt. If a workflow depends on a person quietly fixing bad inputs, interpreting policy contradictions, or noticing when a case “just looks wrong,” then that workflow is not ready for agents. Adding a large language model may hide the weakness for a while, but autonomy will expose it. Companies that want reliable agentic systems will have to redesign the process so the judgment points, stop conditions, and exception routes are visible to the software and to the humans supervising it. (technologyreview.com) So the real story is less about smarter models than about stricter operations. Businesses are learning that agentic artificial intelligence is not a drop-in layer for existing work. It is a forcing function that reveals where approvals are informal, where exceptions are undocumented, and where ownership is fuzzy. The firms that benefit first are likely to be the ones that treat orchestration, escalation, and human checkpoints as part of the product from day one, not as cleanup after the pilot goes live. (technologyreview.com)