The Agentic Handover Gate

A process passes five gates before agents take it: written down, measured, tested on the operation's own work, checked by people first, and reversible. Fail any gate and it stays with people.

The agentic handover gate is the set of tests a service or planning process must pass before its execution moves from people to AI agents. It treats the handover as a placement decision — the same decision an operation makes when it moves work between locations, delivery arrangements or suppliers — applied to a supply node whose attributes are unusual: near-zero ramp time, a quality profile that must be measured in the domain rather than read from a benchmark, and eligibility constraints that may bar it from some work outright. The gate has five tests: documented, instrumented, measured here, overseen first, and reversible and logged. A process that passes all five is in the specified band of the three bands of work; one that passes the first two but not the fourth is overseen; one that fails the first stays discretionary. The gate is the fourth stage of the agentic journey map. It is a node-specific companion to the six-test placement check in Placement Engine Architecture, run alongside it rather than added to it, and it is narrower than Agentic AI Workforce Planning, which covers the whole capacity model for agentic supply — lifecycle, cost, risk and metrics; this page covers only the decision to move one process across.

Why a gate rather than a strategy

Most organizations approach agents as a technology program with its own roadmap, business case and governance, separate from the workforce planning that places every other kind of capacity. The gate rests on a different premise, taken from placement architecture: an agentic capability is a node. It has a cost, a quality profile, a capability set, a ramp profile, a continuity profile and a set of things it may not do. "Should a person or software do this" is therefore the same question as "should this sit in one location or another," asked of a node with unusual attributes. Two consequences follow. The decision to hand a process over needs no placement machinery separate from the one that places every other node — only the agentic node's attributes measured on the same instrument. An organization still needs the capability roadmap that Agentic AI Workforce Planning describes; what it does not need is a second placement process. And the question of how far a human population can fall, and over what horizon, becomes an output of the same planning machinery, bounded by the same constraints, rather than a target announced in advance.

The gate is where that premise becomes operational. It adds one node type to the placement register and a parallel set of node-specific tests, run beside the six-test check every placement proposal already passes.

The five tests

#	Test	What passes	Why it is there
1	Documented	The process has a step table and work instructions to the L0–L3 standard, and the step table's tools column names the systems each step touches	The tools column is the permission map the agent will need. An undocumented process cannot be handed over because nobody can state what "done" means
2	Instrumented	Every objective the process serves — cost, experience, revenue, compliance — has a measured attribute on the agentic node, on the same instrument as the human nodes	A multi-objective placement model optimizes only the objectives its nodes can measure. An objective with no instrument on the agentic node collapses out of the comparison silently (see Instrumenting the Objective Before Building the Model)
3	Measured here	The node's quality profile is measured on this work in this domain, not read from a vendor benchmark or a generic evaluation	Agent performance is strongly domain-dependent and degrades under repetition. On a standard tool-agent benchmark the same model scores roughly half its retail-domain performance in the airline domain, and the probability of succeeding on all of eight independent trials of the same task falls below a quarter in retail and lower still in airline. The ratio and the shape transfer; the headline does not (see AI Agent Evaluation and Benchmarking)
4	Overseen first	A catch rate has been published from an overseen phase, using error injection and sampled deep verification, before the process moves to specified	Human verification of automated output degrades toward rubber-stamping when the automation's reliability is constant and the verifier carries other tasks — automation complacency, in the human-factors literature. Error injection and sampled deep verification are the countermeasure Service Chain Decomposition and Node Sourcing already prescribes for the verification step of a chain; here they make the catch rate measurable
5	Reversible and logged	The process can move back a band; the decision, its owner, its date and the conditions under which it is re-posed are recorded in a trigger log	A handover is a dated transition, not a permanent state. The domain, the platform and the catch rate all move, and each movement re-poses the question

The five outcomes of the placement check apply: pass, fail, fail unless a stated price is paid, pass until a stated date, and cannot run — the last never silently skipped, because it names the definitional or data gap that stopped the test. For agentic handovers pass until a stated date is the normal answer, and cannot run is the normal result of test 2 when an objective has no instrument on the node. A process handed over is handed over until its next re-verification, and the log says when that is.

The tests are cumulative, not independent

Test 1 is the entry condition; nothing is evaluated without it. Tests 2 and 3 can only be run on a documented process, because the objectives and the definition of done come from the step table. Test 4 requires tests 2 and 3, because the catch rate is measured against the definition of done on the instrument the human nodes use. Test 5 is a property of the record, not of the process, and is cheap once the others exist. In practice the gate is therefore a sequence: document, instrument, measure, oversee, log — and the overseer is the role that runs the fourth step.

Choosing the first process

The first handover matters more than any later one, because it forces the documentation standard, the objective instruments, the measurement view and the overseer role to exist for at least one process before any of them is declared standard. Good first candidates share three properties: they are internal to the workforce function, so failure is contained; they have low stakes per run, so the overseen phase can be short; and they require exactly one object to be canonical across the organization, so the definitional work they force is work the organization must do anyway. Assembling the inputs to a monthly capacity outlook is a typical example — collation rather than judgment, with the decision that follows staying discretionary. Publishing schedules is a strong second but touches every worker, so its overseen phase must be long. Intraday reallocation within guardrails is strategically central but usually gated by whether the routing platform exposes a runtime write-back.

The ironies-of-automation argument applies to the choice: the processes easiest to automate are the ones whose failure the remaining humans are least practiced at handling, so the first handover should be one whose failure mode is visible and recoverable.

Failure modes

• Benchmark substitution. Test 3 passed on a published evaluation rather than on the operation's own work. The airline-versus-retail ratio is the reminder that domain can shift the answer by half.
• Skipping the overseen phase. A process moved from documented straight to specified because it "obviously works". Without a catch rate there is nothing to review when it stops working.
• Objective collapse. Test 2 returned pass on cost alone when it should have returned cannot run. The agentic node then wins every comparison on the one attribute it can measure, and experience and compliance objectives disappear from the model.
• Permanent handover. Test 5 treated as a formality; no re-verification date; the process runs on after the domain has shifted beneath it.

Maturity Model Position

The gate presupposes Level 3 — a documentation standard and rule-based automation already in place — and is the mechanism by which an operation reaches Level 4, where agentic nodes are placed by the same machinery as every other node. At Level 5 the gate runs continuously as part of the operating loop rather than as a project milestone.

See Also

• Three Bands of Work — the segmentation the gate moves processes between
• The Agent Overseer — the role that runs the overseen phase and publishes the catch rate
• Agentic AI Workforce Planning — the full capacity model for agentic supply, of which the handover decision is one part
• Placement Engine Architecture — the six-test placement check the gate runs beside, and its five outcomes
• Service Chain Decomposition and Node Sourcing — the verification-step countermeasure test 4 reuses
• AI Agent Evaluation and Benchmarking — domain ratio and reliability under repetition
• Instrumenting the Objective Before Building the Model — why test 2 exists
• The Agentic Journey Map — the sequence of stages the gate sits within
• The Automation Analyst — the role that carries a process through tests one to three