Customers naturally bundle related requests: “Check my order status, also I was charged twice, and my loyalty points seem wrong.” Without explicit decomposition, the agent processes the first concern and forgets the rest. The more concerns in the message, the worse the miss rate.
The completion data
| Concerns in message | All addressed |
|---|---|
| 2 concerns | 85% |
| 3 concerns | 62% |
| 4+ concerns | 38% |
| Most commonly missed | Last concern |
The declining rate and “last concern missed” pattern reveal the agent identifies and acts on early concerns but loses track of later ones. This is a workflow design issue, not a capability issue — the agent HAS the tools for all concerns.
The fix: explicit decomposition step
Add a decomposition step at the start of the workflow: before taking any action, parse the full message to identify ALL distinct concerns, list them, then address each one. This transforms “respond to the first thing I see” into “catalog everything, then respond to each.”
One system reduced multi-concern errors from 40% to under 10% by adding this single step. The decomposition creates an implicit checklist that prevents concerns from being forgotten.
Independent vs dependent concerns
Independent (return + address update): handle in parallel for faster resolution. Different agents or tool calls, no ordering constraint. Unified response at the end.
Dependent (return → refund → purchase with credit): process sequentially respecting the dependency chain. The refund can’t happen before the return is approved. The purchase can’t use credit that doesn’t exist yet. The agent manages the chain and communicates status at each step — the customer shouldn’t have to track dependencies.
Mixed resolution and escalation
Customer raises 3 concerns: shipping status (resolvable), policy exception (needs escalation), address update (resolvable).
Correct: resolve concerns 1 and 3 immediately, then escalate concern 2 with a handoff summary that includes the resolutions already provided. The customer gets immediate answers for what the agent can handle. The human agent receives context about what’s already resolved — they don’t re-investigate completed items.
Wrong: escalate the entire interaction because one concern needs escalation. This wastes human time on concerns the agent already resolved.
Also wrong: handle only the resolvable concerns and ignore the one needing escalation. Even if the agent can’t resolve it, it should acknowledge it and route it properly.
Decomposition accuracy
Production data on decomposition quality:
- 78% correct (all concerns properly identified and separated)
- 8% over-decomposed (split one concern into two)
- 14% under-decomposed (merged two concerns into one → partial resolution)
The 14% under-decomposition is the bigger problem. Fix: few-shot examples showing correct separation of concerns that appear in the same sentence. “I was charged twice AND my shipping is late” = two concerns, not one. Examples teach the agent that co-occurring concerns in a single sentence are still independent items.
Unified response, not three messages
After resolving all concerns, synthesize a single unified response — not three separate messages. A unified response:
- Demonstrates the agent understood the complete request
- Lets the customer verify all concerns were addressed at once
- Avoids confusing interleaved messages in chat
Natural narrative, not robotic checklist
The response should be comprehensive AND natural. Keep the decomposition for completeness tracking, but synthesize into a conversational narrative:
“I’ve looked into all three items you mentioned. Your order is on track for delivery tomorrow. Regarding the double charge, I’ve issued a credit of $45.50 which will appear in 3-5 days. And I’ve updated your shipping address for future orders.”
Not: “Concern 1: Order status - on track. Concern 2: Double charge - credit issued. Concern 3: Address - updated.” Both are complete, but the first reads like a helpful person, the second like a form printout.
One-liner: Without explicit decomposition, only 38% of 4+ concern messages get fully addressed (last concern most often missed) — add a decomposition step, handle independent concerns in parallel, respect dependency chains, resolve what you can and escalate what you must, and synthesize into a natural unified response.