When to Give a New Hire an AI Agent to Run a Task

Coworkers, managers and hiring plans are all organized around people and titles. Delegation research makes the distinction explicit: delegation involves transferring authority, responsibility and accountability, plus clear specifications of roles and boundaries, not just chopping work into smaller pieces. The intelligent AI delegation paper argues that task allocation without accountability, capability matching and continuous performance monitoring produces brittle systems that fail in high-stakes settings. That is exactly the failure mode of "let the new hire figure out where the agent helps." Nobody owns the boundary, so the boundary moves by accident.

The practical consequence for onboarding is that you should write the boundary down before the hire starts. The 30-60-90 day plan format already exists for this: it sets expectations, aligns work to goals and gives both manager and hire a way to track progress and adjust. AIHR's guide to the 30-60-90 plan describes it as a document that defines milestones at 30, 60 and 90 days and is used by managers to catch problems early. Qooper's onboarding template breaks the same structure into phases: understanding the organization in the first 30 days, applying knowledge independently by day 60, and demonstrating measurable impact by day 90. Use that skeleton, but replace "learn the process" with "review the agent, then own exceptions, then own the rubric."

There is a second reason to be explicit early. New hires are expensive to get wrong, and the cost of hiring is not a simple line item. Paychex's breakdown of hiring costs notes that employers must account for recruitment, screening, onboarding, training and lost productivity, and that cost per hire is the top HR metric for a large share of small businesses in its survey. Indeed's cost of hiring guide puts average hiring cost at about $4,700 in 2023 per data from the Employers' Council, while noting wide variation by role and process. QuickBooks' employee cost calculator exists because payroll taxes, benefits and overhead push the real number well above salary. Every week a new hire spends on a task the agent could draft is money already spent. That is the strongest argument for deciding the split before day one rather than after month three.

Frequency is the simplest axis. A task that happens once a quarter is not worth wiring an agent to; a task that happens several times a day is a candidate. The threshold of "at least weekly" is a judgment call, not a law, but it has a useful property: at weekly cadence a review step adds less than a minute of overhead per run for a task that takes several minutes by hand.

Error cost is the axis teams skip. The question is not "how bad would a wrong output be in theory" but "what does it cost to catch and fix a wrong output in practice." A wrong internal summary costs a rewrite. A wrong number in an invoice or a customer-facing message costs a correction and sometimes a relationship. Silent errors are the expensive kind, as the delegation framework above notes: failures that nobody checks are exactly where delegation goes wrong.

Reviewability is where most teams overestimate themselves. A reviewer can judge an output cheaply when there is a rubric, a source document and a short list of failure modes. A reviewer cannot judge an output cheaply when the task requires holding ten constraints in their head or re-deriving the answer from scratch. If reviewing takes as long as doing, delegation buys nothing.

Here is the scoring sheet, with the thresholds applied to typical office tasks.

The rule reads off the table: two green rows go to the agent from day one, one row is a draft-and-approve, and two stay with humans. That is a real onboarding plan, and it takes an afternoon to write. If the new hire's job description is mostly red rows, the agent will not be doing much, and you should say so rather than pretend.

Once a task passes the rule, the next question is how much authority the agent has. The cleanest way to answer it is to borrow the levels managers already use with people. Metal Toad's adaptation of the six levels of delegation maps them onto agents, from "look into it and report, I'll decide" up to "take action, no further contact required," and makes the point that people both under-delegate and over-delegate to agents in practice. Its practical warning is worth repeating: if you want ideas only, say so explicitly, because models default to pushing toward action.

The decision rule for levels is that the review step caps the authority. If the rubric is strong and the reviewer is fast, level 4 (act, but say so, and stop at a human gate for anything unusual) is safe for a weekly task. If the rubric is thin, start at level 1 or 2 for the first two weeks, collect the agent's outputs alongside the human's, and compare. That comparison is also the best possible training material for a new hire: the first week's work is reading ten examples of what good and bad agent output look like for this specific task.

The research on human oversight supports the staged approach rather than the switch. A systematic review of human-in-the-loop AI describes oversight as human feedback and decision-making inserted at various points in the pipeline, not as a single approval at the end, and organizes approaches by where the loop sits and how granular the interaction is. Translate that to onboarding: decide whether the human reviews every run, a sample, or only flagged runs, and write that down next to the task.

The piece that makes the whole arrangement safe is a checkpoint with a written rubric. Three things go in it.

The rubric. Six to ten yes/no questions the reviewer answers for every output. For a status draft: are all figures drawn from the linked export, is every section present, does anything assert a cause that the data cannot support, is the tone consistent with prior updates. Grading should take a couple of minutes, which is what "reviewable" means.

The sampling policy. Every run for the first two weeks, then one in three, then one in ten for stable tasks, with a rule that any output touching an external audience gets full review regardless. A sampling policy that is never written down becomes no review at all.

The trace. For every output, the reviewer should be able to see the input, the prompt, the model response and who approved it. This is the same discipline enterprise teams apply to agent infrastructure. Elementum's guide to agent coordination documents the failure modes that appear once pilots meet production: errors compounding across handoffs, token cost that is hard to predict because each step is billed separately, and observability collapse as agents spin up sub-agents. Its recommendation to build the orchestration and governance layer before the agent count scales applies just as well to one new hire with one agent. Its note that a growing number of organizations are experimenting with agents while fewer than 10 percent have moved them into production at scale is a fair warning that the review step is where most deployments quietly stall.

A minimal checkpoint artifact is worth sketching, because teams overbuild it:

Task: weekly status draft
Delegation level: 4 (act, report, human gate on exceptions)
Rubric: figures trace to export | all sections present | no unsupported causal claims | tone matches prior 3 updates
Sampling: all runs (weeks 1-2) -> 1 in 3 (weeks 3-4) -> 1 in 10
Reviewer: new hire
Sign-off: manager approves anything leaving the team
Rollback: if two graded failures in a week, drop to level 2 until rubric updated

That is the whole governance layer for a single task, and it is enough. A reviewer can run it against a printed rubric, and the rollback rule gives the new hire a way to raise a problem without escalating.

The onboarding literature is mostly about helping a person learn a process. If the agent already runs part of the process, the onboarding content changes. Indeed's onboarding checklist organizes the first weeks around clarity of expectations, speed to productivity, retention and progress tracking, with milestones at 30, 60 and 90 days. The HR Consulting Group's 30-60-90 template frames the same idea around measurable development rather than a compliance checklist. The Management Center's goal-setting template is a useful model for the goal-writing half of the plan.

Rewrite the plan so each task carries its delegation state. A practical sequence:

• Week 1: the hire reads the rubric, grades ten archived agent outputs against it, and discusses disagreements with the manager. This is faster than shadowing and produces a calibrated reviewer.
• Week 2: the hire reviews every agent run, logs failures, and proposes rubric edits. The manager approves external-facing outputs.
• Weeks 3 to 4: sampling drops to one in three. The hire owns the task end to end, with the manager spot-checking.
• Weeks 5 to 8: the hire owns the rubric and the prompts, and reports a single quality number per task (for example, failures caught per hundred runs).
• Weeks 9 to 12: the hire proposes the next task to delegate, using the same three-axis score. That proposal is a better ninety-day review artifact than a status presentation.

Two caution notes. First, a person is not an agent, and dressing up delegation to a person in agent language can blur employment questions. If the "agent plus person" arrangement is actually a contractor doing the work, classification is governed by control and independence tests rather than by the tooling; the IRS's worker classification guidance lays out behavioral, financial and relationship factors, and Paychex's contractor versus employee guide covers the same ground for employers. Indeed's guide to contracting independent workers notes that contractors generally choose how, when and where they work, which is hard to square with tight agent workflow rules. Second, generalists are common first hires at early-stage companies, and role definition is a known failure point for startups, as the Founder Institute's hiring resource points out. Intro's collection of founder advice on early hires emphasizes range and a bias toward action, which is the profile that thrives when part of the process is already automated. Write the agent boundary into the role description and you will attract people who want to own review and improvement, not people who want to run a manual checklist.

If the plumbing is the blocker, one API key across text, image and embedding models removes a week of integration work before the first hire starts; you can read what that covers on Quanta Inference. And if you would rather pressure-test the three-axis score against your own task list before committing, a free 30-minute call with Quanta Consulting is a reasonable way to spend an afternoon.