Three systems may use the same model and carry completely different levels of authority. The useful question is not whether AI is present. It is what the system is allowed to do.
There is a field experiment that makes the cost of getting this wrong concrete. In “Navigating the Jagged Technological Frontier”, researchers from Harvard, Wharton, MIT Sloan, Warwick and BCG gave 758 consultants 18 realistic tasks.
On tasks inside the model’s competence, consultants using AI completed 12.2 percent more tasks, 25.1 percent faster, and 40 percent produced higher quality work. On tasks outside it, the same people performed 19 percentage points worse than colleagues working without AI at all.
Same model. Same users. Opposite outcome, decided entirely by whether the task suited the tool and whether anyone had checked.
That is an authority problem wearing a productivity costume. Most teams describe every AI-enabled workflow as automation, and that word hides the part that matters.
An AI system can suggest a reply, send the reply after approval, or decide to refund the customer without asking. The interface may look almost identical in all three cases. The authority is not.
In our operating framework, assistance produces information for human use, automation executes an explicit rule, and delegation transfers bounded authority to choose and perform actions. These are working definitions, not universally standardised industry categories.
Confusing them creates two opposite failures. Teams treat harmless drafting tools as if they were autonomous machines. Or teams grant decision-making authority to agents while continuing to call them assistants.
A reliable implementation names the authority before it names the tool.
Assistance produces work for a human decision
An assisted system helps a person perform a task but does not own the decision or execute its consequence. It may retrieve information, compare documents, summarise a meeting, draft a response, identify missing fields, propose a classification or flag an unusual pattern.
The person can accept, change or reject the output. More importantly, the person is expected to examine the relevant evidence before acting.
AI drafts three subject lines, a marketer selects one. AI summarises a contract, a lawyer reviews the contract itself. AI identifies invoices with possible mismatches, finance verifies them. AI proposes a support response, an agent sends the final message.
Assistance is not risk-free. A plausible draft can anchor the reviewer’s judgment. A summary can omit the clause that matters. A recommendation becomes a rubber stamp if workloads are too high. The jagged-frontier study is a warning here specifically: the harm arrived through people who trusted a confident output on a task the model was bad at.
Human review only counts when the person has enough time, enough competence, access to the source, authority to reject the output, and a reason to notice an error.
A button labelled “approve” does not create meaningful oversight.
Automation executes an explicit rule
Automation performs a predefined action when defined conditions are met. The intelligence may sit in the design of the rule rather than in a model.
Move a file when its status changes to approved. Send a renewal reminder 30 days before expiry. Calculate a total from verified fields. Create a project folder from an accepted form. Stop campaign spend when a hard limit is reached.
A language model can participate in an automated workflow, but it should not be used where a simpler deterministic rule is sufficient. Deterministic systems are easier to test, explain, reproduce, audit, secure and price.
If the business rule is “send a reminder 30 days before the contract ends,” a generative model adds cost and variability without improving the decision.
Delegation grants bounded authority
Delegation occurs when a system is permitted to choose among actions and execute one without case-by-case human approval. An agent chooses which approved supplier to contact and sends the request. A system reschedules appointments inside a permitted window. An agent reallocates budget within a predefined limit.
Delegation is not defined by whether the system “reasons.” It is defined by whether the system can change the world.
The moment an AI can send, publish, spend, delete, approve, deny, order, schedule or modify a system of record, it has authority. That authority needs a permission boundary.
The Authority Ladder
Six levels describe what a system is allowed to do.
Level 0, observe. The system reads permitted inputs and produces no output beyond internal analysis. Detect recurring ticket categories for a process review.
Level 1, draft. The system prepares an output for human review. Draft a reply from an approved knowledge base.
Level 2, recommend. The system proposes a decision and presents supporting evidence. Recommend whether an invoice should be approved, with cited purchase-order lines.
Level 3, prepare action. The system completes every reversible preparation step but waits for explicit approval. Prepare an email, recipients and attachment, then require a person to press send.
Level 4, act and notify. The system acts inside a defined boundary and immediately records and reports what it did. Reschedule an internal meeting inside agreed working hours where all participants have authorised that behaviour.
Level 5, act and escalate exceptions. The system operates independently inside a narrow domain and stops when a defined boundary is crossed. Process low-risk stock transfers under a fixed value while escalating shortages, conflicts and unusual requests.
Level 5 is not the target for every workflow. Many valuable systems should remain at Level 1 or 2.
Maturity is not measured by the amount of authority transferred to a machine. It is measured by whether the assigned authority is appropriate and controlled.
Classify the task using four variables
Consequence. What is the maximum credible harm if the action is wrong? Financial loss, customer harm, contractual effect, privacy, discrimination, reputation, safety, loss of records.
Reversibility. Can the action be undone completely, quickly and cheaply? Drafting is more reversible than sending. Scheduling is more reversible than paying. Archiving may be more reversible than deletion.
Detectability. Will the team know that an error occurred? A visible formatting error is easy to detect. A missing contractual qualification inside a polished summary is not.
Evidence quality. Does the system act from complete, current and authorised information? A system connected to ten years of ungoverned files does not have institutional knowledge. It has institutional noise.
The default authority rule
Start at the lowest level that produces useful value. Increase authority only after the team can demonstrate stable inputs, measurable output quality, known failure modes, working exception handling, complete action logs, successful recovery tests, and a named owner.
This is the opposite of designing the most autonomous demo and then adding safety around it.
A worked example: customer refund requests
Assistance. The system retrieves the order, return policy and previous messages. It drafts a response. A support agent decides.
Automation. If a cancelled order has not shipped and the payment is still pending, a deterministic rule reverses the authorisation.
Delegation. The system evaluates evidence, chooses whether the case fits an approved exception, and issues a refund up to a fixed amount.
Those are three different operating systems. For the third, the team must define maximum refund value, eligible products and markets, required evidence, customer-history access, abuse indicators, excluded cases, daily action limit, escalation conditions, audit log and reversal process.
Calling all three “AI customer service” erases the controls.
Human oversight needs design
Article 14 of the EU AI Act establishes human-oversight requirements for high-risk systems. Under Regulation (EU) 2026/1744, those Chapter III requirements apply from 2 December 2027 to Annex III high-risk systems and from 2 August 2028 to Article 6(1) and Annex I high-risk systems.
Systems outside those classifications may still face transparency, literacy, privacy, consumer, employment or sector-specific obligations. The underlying design lesson holds regardless: people need to understand the system’s capacity and limits, notice anomalies, interpret outputs correctly, and be able to disregard or reverse them.
The company should be able to answer: who is watching, what are they expected to notice, what evidence can they inspect, what action can they stop, how quickly can they intervene, and what happens outside working hours.
Without those answers, “human in the loop” is decorative language.
How we describe a system
We describe every proposed AI workflow using two labels: the function (retrieve, classify, draft, recommend or act) and the authority level, 0 through 5.
“An AI assistant for campaign operations” is too vague. “A Level 2 system that identifies pacing anomalies and recommends a correction using approved campaign data” is a design that can be evaluated.
The first description sells possibility. The second creates accountability.
Which decisions deserve which level is the question answered by a recurring decision map, which is where this work starts. In practice the systems we build sit inside Scale, where an action costs money the moment it executes, or inside Origin, where the risk is publishing something nobody approved.
AI does not become useful when it can do more. It becomes useful when everyone knows exactly what it is allowed to do.
If you are about to give a system permission to act, tell us what it will be allowed to change and we will tell you which level it should start at.
Sources
Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon and Lakhani, “Navigating the Jagged Technological Frontier”, field experiment with 758 BCG consultants, published in Organization Science.
NIST AI Risk Management Framework Core.
NIST Generative AI Profile.
OWASP Top 10 for Large Language Model Applications.
European Commission, AI Act implementation and human oversight requirements.
Regulation (EU) 2026/1744 amending the AI Act.
Current to 30 July 2026. This article provides an operating framework, not legal advice.
