agent-swarm.devagent-swarm.dev
Guides

Reviewing a potential customer service-level breach with AI agents

Agent Swarm gathers the agreement, timestamps and service records behind an alleged breach and hands a named owner a source-linked review packet. People decide.

Agent Swarm can be configured to gather the applicable agreement version, the relevant timestamps and the service records for one customer and one period, compare them against an explicitly defined contractual window, and hand a named owner a source-linked review packet. A person verifies the data and interprets the contract. Account and finance owners decide any customer communication or credit. What it removes is the assembly: finding the clause in force and the timestamps that bear on it across systems that each hold a piece, and naming where they disagree.

The scope is one alleged breach of a contractual service response-time commitment, for one customer and one period. A breach claim needs no outage behind it. Keeping delivery commitments visible in the first place is a different register, worked through in customer commitment tracking. A customer disputing a bill is a finance question, covered in customer billing disputes. Reviewing customer impact after an outage asks what a failure cost, which is a separate exercise from whether a response window was met.

When another tool fits better

Purpose-built service-level monitoring. A service desk or monitoring product that measures response windows natively and continuously already holds the clock. Where the commitment is a field on the ticket and both sides read the same timer, that product answers the question. Agent Swarm ships no service-level monitoring feature: no response-window clock and no alert as one closes.

Legal review. When the question is what the contract means, who is liable, or whether a formal claim follows, counsel reads the agreement. A review gate is not a legal opinion, and timestamps do not settle an interpretation.

Agent Swarm fits a different shape: the records are split across a ticketing system, an agreement store, an on-call log and email, and one person needs them together before anyone answers the customer. Choosing an AI operating model sets out that split.

An illustrative workflow

This is an illustrative design, not a built-in service-level feature, not a native contract integration and not a customer result. Integrations, permissions, contractual rules, timing checks and source access need configuration and testing by your team before any of it runs.

1. The alleged breach is logged. A customer writes that a response took longer than the agreement allows, or someone on the team flags a ticket that passed its window. The workflow starts from that claim, with the customer, the period and the commitment named.

2. Configured scripts retrieve the records. Custom connections your team builds pull the agreement version in force for the period, the ticket's created, first-response and resolution timestamps, and the on-call record for those hours. Scripts reach each system through script connections, with credentials from the credential broker. No documented integration returns a ticket's response timestamps measured against a contractual window: the tracker integrations bring issues in as tasks and post state back. No connector on the documented integrations list reads a contract repository or an on-call log, so each record pull here is custom work your team builds and tests.

3. A swarm-script compares each timestamp pair against the defined window. The window is a configured input, not something a model infers: the number of business hours, the time zone it is measured in, when the clock starts and whether it pauses outside business hours. The step emits JSON, so the same records and window produce the same result on every run. Reports from multiple sources documents that split between deterministic collection and an agent writing from it.

4. An agent assembles the packet. Every line carries a source link and the date of the record it rests on, so a reader can follow any timestamp back.

5. A person clears the gate. A human-in-the-loop gate pauses the workflow and resumes only on a human response. The named reviewer confirms the data, and the account and finance owners then decide whether anyone contacts the customer and whether a credit is offered. On reject the packet is archived, returned for more records or escalated.

What the review packet contains

Each item carries the date of the record behind it.

  • The agreement version in force for the period, citing the clause that sets the commitment and linking the document it comes from.
  • The window as configured, with the assumptions used: time zone, business hours, when the clock starts, whether it pauses.
  • Each event with its source and timestamp: ticket created, first response, resolution, and the on-call entry where one exists.
  • The computed result per event, within or outside the window, stated as arithmetic, not a conclusion.
  • Flags, each with its source: a missing record, terms that differ between two agreement versions or between an agreement and an order form, a time-zone assumption the records do not settle, and any contractual exception the review must weigh, such as scheduled maintenance, a customer-caused delay or a named exclusion.
  • Open questions for the reviewer, one per flag, with the person or team who can answer it.
  • One named owner for the review, plus a backup.

The packet is not a legal conclusion. It does not establish that a breach occurred, it does not trigger a refund or a credit, and it computes no service-credit amount.

Illustrative example: a response-time commitment for one customer

This scenario is fictional. It names no company or customer, and nothing in it is a measured result.

One provider team supports one customer under an agreement committing to a first response within four business hours, Monday to Friday, 09:00 to 17:00 in the customer's time zone. The review covers one month: did any ticket fall outside that window?

TicketOpenedFirst responseElapsed business hoursResultSource
NW-4412Mon 12 Jan, 09:40Mon 12 Jan, 12:152.6Within the windowTicket history export, 12 Jan
NW-4418Tue 13 Jan, 16:05Wed 14 Jan, 09:501.8Within the window, clock paused overnightTicket history export, 14 Jan
NW-4427Thu 15 Jan, 14:20Not recordedNot computedFlagged: no first-response timestampTicket export, 20 Jan
NW-4431Fri 16 Jan, 23:10 (customer time)Mon 19 Jan, 08:05 (team time)0.1Within the window once both records are read in the customer's time zone: 09:05 on Monday, five minutes after the clock starts. Flagged: the two records are stamped in different time zonesTicket export (team time) and the customer's message (customer time)
NW-4436Tue 20 Jan, 10:30Wed 21 Jan, 11:008.5Outside the window by 4.5 business hoursTicket history export, 22 Jan

Three rows go to the reviewer, not to a verdict. NW-4436 is the only ticket the packet reports outside the window: it links the export and carries both timestamps, 10:30 on Tuesday and 11:00 on Wednesday morning, which come to 8.5 business hours against a window of four. The named reviewer confirms both timestamps against that export and reads the agreement for an exception that could stop the clock, such as scheduled maintenance, a customer-caused delay or a named exclusion, and checks the on-call log for those hours. On NW-4427 the missing timestamp is a gap in the export, and the packet names the mailbox where the reply may sit. On NW-4431 the customer's message dates the ticket in customer time and the export stamps the first response in the team's time zone, one hour behind. Read in the customer's time zone, the response lands at 09:05 on Monday, five minutes after the clock starts, and the Friday night to Monday morning gap never enters the count, because the clock runs only Monday to Friday, 09:00 to 17:00. Account and finance owners then decide whether to reply to the customer and whether a credit is offered.

Sensitive data, stale records and escalation

Least privilege is a setting, not a guarantee. It depends on the connections, credentials and data handling the team decides. Script connections are scoped global (the default), agent or repo, and the team chooses. The credential broker keeps raw secrets out of script source for allowlisted hosts, protecting the secret, not which records are read.

Two limits. RBAC gates user-token requests, not operator-key or agent calls. Shared memory is visible to every agent, so keep ticket, agreement and customer correspondence content out of it or scope it.

The team keeps the record. Its governance entry states who may read which record and who grants access.

No records pasted into chat. A record the swarm cannot reach is named with the owner who can grant access, never obtained by asking someone to paste it into a thread.

Stale or incomplete records are flagged, not filled. When a ticket export stops before the period ends, an agreement carries no effective date, or the clock-start rule is unset, the packet states the record, its last-updated date and the gap.

Conflicting terms are both reported. When two agreement versions state different windows, or an order form names a shorter response time than the master agreement, the packet shows both with dates and links. People decide which governs.

Escalation when the gate is not cleared. The gate carries a timeout and a fallback path, as HITL gates recommends: a reminder, a named backup owner, then an explicit hold so the claim stays on the open list. It must never proceed on silence, because a customer reply can carry an admission and a credit at once.

What stays human-owned

Five decisions stay with people in every version of this workflow.

  • Validating the data. Whether a timestamp is the right one, and whether the packet is complete enough to act on, is a person's judgement. The agent presents sources and certifies none.
  • Interpreting the contract. Whether a clause covers this commitment, what the window means in practice and whether an exception applies are readings of the document. The packet quotes terms and interprets none.
  • Deciding the customer communication. The agent never messages the customer. A person writes or approves the reply, including any acknowledgement of a missed window.
  • Any credit, refund or commercial remedy. The agent never issues one and never moves money. The workflow has no path to a payment instruction and ends at a decision a person recorded.
  • Changes to the record. The agent does not edit the ticket, the agreement store or the account record. Someone with write access does it, inside the system that owns it.

Gate the irreversible steps, as the HITL guide states. Gathering records and writing the packet are reversible. A customer reply and a credit are not.

Building blocks

Each piece below ships in agent-swarm and is documented. Pointing them at this review is the illustrative part.

Two neighbouring reviews use the same shape with a different owner: customer commitment tracking and customer billing disputes.

On this page