The unit of decision is the step, not the process
“Should we automate order handling?” is the usual question, and it is the wrong size. A process is a sequence of steps, each with its own inputs, failure modes and cost of failure. Parasuraman, Sheridan and Wickens (2000) made this point for automation in general: it is not one thing applied to a whole task but a level chosen separately for each class of function, from gathering information through analyzing it, selecting an action and carrying it out. A step that only gathers information can be delegated almost fully; a step that commits the organization to an external party is a different object, even inside the same process.
Agents sharpen the point. An autonomous agent is a player whose payoff is an objective function, the check, reward signal or budget rule it is scored against, rather than any desire of its own. Krakovna and colleagues (2020) define specification gaming as behavior that satisfies the literal specification of an objective without achieving the intended outcome. Whether a step is safe to delegate therefore depends on whether we can write a specification for it that the cheapest passing move actually satisfies, which is a property of the step, not of the process or the model.
So before we automate anything, we sort: every step is assigned to one of four lanes, and the lane carries the operating rules.
The four lanes
Human-led. Judgment, accountability and relationships that should stay with people, by design rather than by default. Steps land here because we decided so, not because nobody got around to automating them.
Existing software and assistance. Steps your current systems or assistants already handle well once the data and interfaces are fixed. Much that is pitched as an agent project is a data-quality project in disguise. If a rules engine, a workflow tool or an assisted person does the step correctly on clean inputs, the fix is the inputs.
Delegated autonomous workflow. Bounded work an agent can run under explicit acceptance criteria, budget and escalation rules. All three exist in writing before the first run. The acceptance criterion cannot be satisfied by asserting completion. The budget stops the run without a human noticing. The escalation rule names who receives the step when the agent stops, and with what evidence.
Candidate agentic unit. Teams or functions ready to operate as an agent-run unit sooner than the rest of the organization. This lane is about a cluster of steps, not a single step: steps dense in coordination with each other, with a measurable joint output and few irreversible touchpoints with the outside. Such a cluster can run as a unit with its own agents and gates and an accountable human owner.
The lanes are not a maturity ladder. Some steps are human-led because of what they are, whatever the technology does; others only until a check that can fail exists. The sort records which is which.
Three tests that sort a step
Three properties decide the lane, and each has an economic treatment that applies to an agent with an objective function as well as to a person with a contract.
Reversibility. Dixit and Pindyck (1994) built the theory of investment under uncertainty on two observations: most investment decisions are hard to reverse, and the uncertainty around them resolves over time, so waiting for better information has value. Option value is the worth of keeping a decision open until more is known. When an agent’s action can be undone at low cost, holding the decision for a human has little option value. When it cannot, as with a payment, a contractual commitment or a deletion, the option value is high, and that is where a human approval bound to a specific proposal earns its cost. Reversibility pushes a step toward the delegated lane; irreversibility pushes it toward human-led, or toward a delegated lane with a hard escalation at that action.
Measurability. Holmström and Milgrom (1991) showed that when an agent performs several tasks and only some are measured, rewarding the measured ones pulls effort away from the rest. For a software agent the result is literal: it optimizes what the check measures and nothing else. A step is a candidate for delegation only if the outcome we want is the one the check measures, and the check can fail without the agent’s cooperation. Where the part that matters is unmeasurable, such as whether a customer felt heard, the step is human-led or assisted, and no prompt changes that.
Coordination cost. Coase (1937) asked why firms exist if markets allocate resources well, and answered that using the price mechanism has costs: discovering prices, negotiating and enforcing contracts for each transaction. Williamson (1979) developed this into transaction cost economics, which asks how the attributes of a transaction decide the governance structure that handles it, inside an organization or across its boundary. For agents the question becomes how much coordination a step needs. A step that exchanges a clean artifact with its neighbors can be delegated on its own. A cluster of steps that exchange partial state with each other constantly, where splitting them between human and agent hands would create a handoff, and a coordination cost, at every turn, is either kept together as human work or is the raw material for a candidate agentic unit. The unit’s boundary sits where coordination cost drops.
Two hypothetical sorts
Both scenarios are invented; neither describes a client.
Hypothetical one: returns handling at an online retailer. A customer requests a return; the request is validated against order and policy; a label is issued; the item is received and inspected; a refund or replacement is decided; the refund is paid; the stock record is updated. Sorting by step: validation, label issuance and the stock update are existing software once the policy is encoded and the order data is clean, which it is not, so the first deliverable is fixing that. Inspection is human-led for now because the measurable part, whether the item matches the manifest, is small next to the judgment, whether damage is wear or abuse. The refund decision for in-policy, undamaged returns is a delegated autonomous workflow, with a check that recomputes the refund from the order record and a daily budget cap. The refund payment is the one irreversible action. Below a threshold its cost is small enough that the recomputation check suffices; above it, the payment escalates to a named person with the evidence attached. Nothing in the sort required an agent on day one, and two deliverables improve the human-run process immediately.
Hypothetical two: dependency upgrades in an engineering organization. A platform team spends much of its time upgrading libraries, applying security patches, running tests, fixing breakages and opening pull requests. The steps coordinate constantly, the joint output is measurable (dependencies within policy, build green, executed test count unchanged, no new vulnerabilities above a set severity), and the single irreversible touchpoint is merging to the main branch. This cluster is a candidate agentic unit. Its agents run upgrades, repair breakages and produce evidence; a reviewer agent scored on agreement with independent checks gates the pull request; merge requires a human approval bound to the digest of that change. The team’s human lead owns the unit’s output and budget. The rest of the organization may be years from this; the sort found one unit that is ready.
The sorting decision, drawn
Read the diagram top to bottom, one step at a time. The first question comes first so that the human-led lane is a choice rather than a residue. The second diverts the many steps that need better data, not an agent. The third is the delegation gate, with three parts because a check without a budget runs forever and a budget without an escalation stops in silence. The fourth applies only after several steps are sorted, since a unit is a property of a cluster. A step that fails the third question returns to the first two lanes with a note on what would have to be true for it to move.
What the human is left with
Bainbridge (1983) described the ironies of automation in process plants, and two apply directly. The designer who tries to eliminate the operator still leaves the operator the tasks the designer could not think how to automate, so the human is left, in her words, with an arbitrary collection of tasks. And the human asked to monitor a system that rarely needs intervention has been given an impossible task, because attention to a source where little happens cannot be sustained for long.
Sorting by lane answers both. The human-led lane is filled first and on purpose, so people keep a coherent set of judgments, not the residue. And no human monitors the delegated lane continuously: the check monitors, the budget stops, and the human appears at a named escalation with evidence in hand. A delegated workflow that needs a person watching a dashboard is in the wrong lane.
Why the sort has value when deployment is deferred
An organization that will not deploy agents this year still gains from sorting now. NIST’s AI Risk Management Framework (2023) describes a Map function that establishes the context in which a system’s risks are framed, with governance running throughout. The framework says that after mapping, users should have enough contextual knowledge “to inform an initial go/no-go decision about whether to design, develop, or deploy an AI system” (NIST, 2023, p. 25). A lane assignment for every step is that context, written in terms an operations leader can act on. The second-lane fixes, clean data and working interfaces, pay off in the human process immediately. The third-lane acceptance criteria measure the human process in the meantime. And when a unit becomes ready, the decision is made against a map that exists, which is option value in practice.
What to do on Monday
Pick one process and list its steps. Fill the human-led lane on purpose. Send to the second lane everything your systems would handle with clean inputs, and write down the cleaning needed. For what remains, try to write the check that can fail, the budget and the escalation. Where you can, you have a delegated workflow; where you cannot, you know what must become true first. Then look across the delegated steps for a cluster that could run as a unit.
Onomos and agent.ceo belong to the same company group, and the lane rules here are the ones that group applies to its own agents. If you want a process sorted into lanes with its acceptance criteria and first increment written, Onomos’s autonomous-systems preparation project begins with a 20-minute fit conversation.
Sources
- Onomos (2026). Live site, read for the lane definitions and service description. https://onomos.ai
- Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). “A model for types and levels of human interaction with automation.” IEEE Transactions on Systems, Man, and Cybernetics, Part A 30(3), pp. 286-297. DOI record opened at https://api.crossref.org/works/10.1109/3468.844354 (publisher page https://ieeexplore.ieee.org/document/844354 refused automated access).
- Krakovna, V. et al. (2020). “Specification gaming: the flip side of AI ingenuity.” DeepMind blog. https://deepmind.google/discover/blog/specification-gaming-the-flip-side-of-ai-ingenuity/
- Dixit, A. K. and Pindyck, R. S. (1994). Investment under Uncertainty. Princeton University Press. https://press.princeton.edu/books/hardcover/9780691034102/investment-under-uncertainty
- Holmström, B. and Milgrom, P. (1991). “Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design.” Journal of Law, Economics, and Organization 7 (special issue), pp. 24-52. https://academic.oup.com/jleo/article/7/special_issue/24/1073412
- Coase, R. H. (1937). “The Nature of the Firm.” Economica 4(16), pp. 386-405. DOI record opened at https://api.crossref.org/works/10.1111/j.1468-0335.1937.tb00002.x (publisher page refused automated access). Prize citation read at https://www.nobelprize.org/prizes/economic-sciences/1991/summary/
- Williamson, O. E. (1979). “Transaction-Cost Economics: The Governance of Contractual Relations.” Journal of Law and Economics 22(2), pp. 233-261. DOI record opened at https://api.crossref.org/works/10.1086/466942 (JSTOR page refused automated access). Prize citation read at https://www.nobelprize.org/prizes/economic-sciences/2009/summary/
- Bainbridge, L. (1983). “Ironies of Automation.” Automatica 19(6), pp. 775-779. Author’s reproduction at https://www.complexcognition.co.uk/2021/06/ironies-of-automation.html; scanned original read at https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf
- National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf