The delegation gap: what you can hand off to your agent

Anthropic's 2026 Agentic Coding Trends Report landed with a finding that stopped a lot of engineering teams cold: developers use AI assistance in roughly 60% of their work, but they can fully delegate only 0-20% of tasks. The authors named this the delegation gap.

The report was about software engineers, but the gap shows up almost identically in customer experience. Your support team probably uses AI to draft responses, summarize conversations, and surface relevant knowledge. But when you count the conversations where the agent runs start-to-finish without a human in the loop? That number is usually much smaller than you'd expect, even for the simple cases.

That gap is the problem worth solving. Not "should we use AI" but "what can we actually hand off."

What full delegation means for CX agents

Full delegation means the agent handles a conversation from the first message to resolution without a human touching it. No review queue. No draft approval. The agent reads the message, decides what to do, calls the tools it needs, and sends the response.

That's different from AI-assisted, where a human still clicks send. It's different from partially automated, where the agent handles some steps and escalates others.

Full delegation is only appropriate for some conversations. The goal isn't to push everything there. The goal is to correctly identify which conversations belong there and close the gap on those.

Why the gap stays wide

The gap persists because of three fixable problems: insufficient test coverage, vague escalation policy, and no real-time quality monitoring.

Most CX teams that have deployed agents describe the same frustration. The agent handles routine questions well. But the moment a customer adds one wrinkle, "my order is delayed AND I want to change the delivery address AND I'm a premium subscriber," the agent either halts, escalates unnecessarily, or, worse, does something that has to be manually unwound.

This happens because the agent's test coverage doesn't match production variety. Teams test the happy path. Customers bring edge cases. When the edge case hits and the agent fumbles it, teams pull back delegation across the board rather than fixing the specific failure.

The escalation policy is usually vague too. "Escalate to a human when needed" sounds reasonable but gives the agent no clear criteria. The agent either escalates too readily, defeating the point of automation, or too rarely, leaving customers in dead-end loops.

And without monitoring that catches quality degradation early, the team often learns about problems through customer complaints rather than instrumentation. By the time the signal arrives, the damage is already done.

The gap isn't fundamentally about AI capability. It's about coverage, clarity, and observability.

A framework for the delegation decision

Not every conversation is a candidate for full delegation. Four questions, asked in order, sort any conversation type into "ready" or "not yet."

The questions matter in order. A predictable, low-stakes conversation with no test coverage is still not ready. You need all four.

Conversations that typically clear all four bars

FAQ and policy questions. "What's your return window?" "Do you ship internationally?" The answer is deterministic, the stakes of a wrong answer are low, and these questions are easy to cover in scenario testing. This is where most teams start, and rightly so.

Order and account status lookups. "Where's my order?" "When does my subscription renew?" The agent retrieves the record and reports back. The outcome is predictable. The risk is low. Coverage is straightforward to build.

Appointment reminders and confirmations. Outbound messages with no ambiguity in the content. The agent sends the reminder, handles "can we reschedule," and books the slot. This is also where the "tasks that wouldn't have been done otherwise" effect shows up. Proactive outreach that was previously too expensive to run at scale becomes routine when the agent handles it fully.

Conversations that need conditions before delegation

Multi-step transactional flows. "I want to return this item and use the refund to buy a different size." This is delegatable if you've tested the combined flow, but not if you've only tested returns and exchanges in isolation. Composition failures, where two tested-separately scenarios break when they appear together, are the most common source of production surprises.

Complaint handling. A customer who is frustrated but has a clear, solvable problem is often a good candidate for full delegation, if the agent's tone is calibrated and you've tested it on charged language. But complaints that include explicit escalation requests need clear policy so the agent knows exactly when to hand off.

Account modification requests. Changing an address, updating payment info, canceling a subscription. These are operationally routine but carry enough weight that most teams want monitoring in place before they fully delegate. A mistake isn't catastrophic, but it's friction you'd rather avoid.

Conversations that should stay human-assisted

Anything with significant financial, legal, or compliance implications. Disputes, refunds over a threshold, regulatory requests. The cost of an agent error is too high, and the outcome is too variable. AI can assist, surfacing the relevant policy and drafting the response, but a human should confirm.

First contact on enterprise or high-value accounts. The relationship risk of an autonomous response on a high-value account usually outweighs the efficiency gain. Use the agent to prepare the human, not to replace them.

Novel situations the agent hasn't seen. If the customer's situation doesn't match any known pattern, the agent shouldn't improvise. A well-built agent recognizes when it's out of depth and escalates cleanly. This is harder to build than it sounds, and it's worth testing explicitly as its own scenario type.

How test coverage closes the gap

The delegation gap shrinks when confidence goes up. Confidence comes from test coverage. These are two ways of saying the same thing.

When you can run the agent against a representative library of conversations from the category you want to delegate, see how it handles the edge cases, measure its resolution rate and tone, and compare that to your human team's baseline, you have evidence. That evidence is what makes it safe to say "this category is ready."

The teams that close the gap fastest treat scenario libraries as infrastructure, not as a one-time pre-launch check. They build coverage that grows with the product. When a new edge case surfaces in production, it becomes a scenario. When a policy changes, the scenarios get updated. The coverage becomes a living definition of what the agent can handle.

This is what Chanl's Scenarios feature is built for. Rather than ad-hoc test scripts, you define simulation personas that cover the realistic range of your customer base, run them against the agent, and use the results to make the delegation call with actual data rather than gut feel.

For a deeper look at how the handoff works when the agent does need to escalate, Handoff is the new prompt covers the context package in detail.

What monitoring tells you after you delegate

Monitoring tells you whether the delegation decision you made last month still holds today. Customer behavior shifts, model updates land, and the edge cases your test suite covered may no longer match what's arriving.

Real-time monitoring does three things.

It catches off-script conversations early. When a conversation starts deviating from the patterns the agent was tested on, surfacing it quickly means you can intervene before the customer has a bad experience. Waiting for a CSAT survey is too slow.

It grades quality on resolved conversations. A scorecard that evaluates whether the agent got the right answer, maintained the right tone, and resolved the issue efficiently gives you a signal independent of customer rating. Some customers don't respond to surveys. Some rate poorly even when the interaction was handled well. An automated scorecard gives you a consistent measure.

It tracks the escalation rate by category. If the agent starts escalating 40% of appointment scheduling conversations when it used to escalate 8%, something changed. That's a diagnostic signal, not just a performance number.

Score

Good

0/100

Tone & Empathy

94%

Resolution

88%

Response Time

72%

Compliance

85%

Chanl's Scorecards run AI-powered quality grading on completed conversations. Monitoring surfaces the real-time patterns. Together they form the feedback loop that lets you expand delegation to new categories over time, rather than treating the initial rollout as a fixed boundary.

The 27% effect in CX

One finding from the Anthropic 2026 report that tends to get overlooked: about 27% of AI-assisted work consists of tasks that wouldn't have been done otherwise. The productivity gain isn't just "same work, faster." Some of it is entirely new work that was previously too expensive to attempt.

That same dynamic appears in CX when you close the delegation gap. When your agent can fully handle appointment reminders, you don't just speed up reminders that humans were already sending. You start sending reminders for appointment types that previously got no outreach. When your agent handles post-purchase check-ins, you do them at a scale that was never economical before.

The delegation gap isn't only an efficiency problem. It's the boundary of what's possible.

For more on where the human-AI line in customer experience is shifting, Conversational vs. agentic AI draws the distinction in detail.

Closing the gap

The delegation gap doesn't close because the model gets better. It closes through better process: the test coverage that builds confidence, the escalation policy that gives the agent clear criteria, and the monitoring that tells you whether the delegation decision is still holding.

The teams that move fastest don't wait for perfection. They pick one conversation category, build coverage for it, add monitoring, delegate it, and measure it. Then they do it again for the next category. That's how Chanl approaches the Build-Connect-Monitor loop: each pass through the cycle earns the right to delegate a little more.

The gap closes one conversation type at a time.

El briefing de Signal

Un email por semana. Cómo los equipos líderes de CS, ingresos e IA están convirtiendo conversaciones en decisiones. Benchmarks, playbooks y lo que funciona en producción.

Subscribe

500+ líderes de CS e ingresos suscritos