AI contract performance reporting.

What goes into a cleaning contract performance report

A cleaning contract performance report shows a client whether the service in the contract was delivered: cleans completed, tasks done, audit scores, complaint response, and corrective actions. The records already exist, spread across time and attendance, checklists, audit forms and the ticket system. The work is exporting, reconciling and explaining the misses.

That explanation is where most of an account manager’s time goes, and it is where a language model helps, provided it is never trusted with the arithmetic.

The reporting pipeline at a glance

Inputs

Rostered and attended shifts with start and finish times, checklist completions and photos, audit scores and comments, site issue tickets with resolution times, each contract’s KPI definitions and targets, and last month’s agreed actions.

Systems

The operations platform holding rosters, timesheets and checklists, the ticket system, audit forms, a reporting layer such as a spreadsheet or Power BI, and the client report template.

Steps

Pull the month’s records, compute each KPI with versioned code, detect exceptions, gather the records behind each exception, draft commentary and corrective actions that cite record IDs, check every number in the draft against the table, then route to the account manager.

Human review

The account manager reviews every exception explanation and decides what the client sees. An operations manager approves corrective action commitments. Anything about an individual cleaner is removed or approved separately.

Outputs

A client report (KPI table, trend, commentary, actions), an internal appendix with cleaner-level detail for coaching, and an evidence pack for disputes.

Measures

Hours per report, numeric mismatches caught by validation, client queries per report, reports delivered on time, and KPI disputes.

Risks

An invented reason for a missed clean, KPIs computed differently month to month, individual cleaners identified to the client, and ticked tasks without photos counted as done.

Illustrative example: one office contract, one month

Illustrative example: a nine-floor Wellington office building cleaned five weeknights a week by a crew of six, under a contract with three KPIs. September 2026 had 22 scheduled nights. All figures are synthetic.

The draft reports 95.5% as the raw result, then shows 100% excluding the client-caused access failure, labelled as such. Whether the exclusion applies is the account manager’s call under the contract, not the model’s. The audit dip becomes a corrective action: a level 6 kitchen checklist change and a supervisor spot check.

Assume an account manager spends 3 hours a month building each report by hand and 1 hour reviewing a computed draft. Across 25 contracts that is 75 hours down to 25, saving 50 hours a month, before counting fewer client queries.

The New Zealand rules that shape the report

Cleaner-level data is personal information, which changes what belongs in a client report:

• Use limits. Start and finish times are collected to pay staff and evidence service. Privacy principle 10 limits using personal information for other purposes, and ranking named cleaners for a client is a different purpose. Report site-level results to clients.
• Accuracy. Privacy principle 8 requires checking information is accurate before using it. A missed clock-out can be a phone battery, not a skipped shift, so confirm before anything is attributed to a person.
• Fair process. If report data leads towards a disciplinary step, Employment New Zealand’s good faith guidance says the employee must get the information and a chance to comment before a decision. Internal coaching flags are prompts for a conversation, not findings.

What the account manager keeps

• Whether a miss counts against the contract.
• Corrective action commitments, which are promises to the client.
• Anything about an individual cleaner.
• Sign-off, sending, and handling disputes or credits.

Where automated reports go wrong

• Invented explanations. The model fills a gap with “due to staff illness” when no record says so. Control: commentary may only cite record IDs in the evidence bundle, and uncited sentences fail validation. See structured outputs and validation.
• Number drift. 95.5% becomes “about 96%”. Control: an automated check that every figure in the text matches the table.
• Definition drift. A KPI is calculated one way in March and another in April. Control: versioned KPI code taken from the contract schedule.
• Tick-box inflation. Tasks ticked without photos count as done. Control: report completion with and without photo evidence side by side.
• Over-sharing. Names leak into the client version. Control: a separate internal appendix and a final redaction check.

A first contract to automate

Pick one contract and the KPIs already in its schedule. Build the computed table first; it is often useful on its own. Add commentary once the numbers are trusted. Operations platforms such as freshOps already record checklist completion with photo verification and start and exit times, which is the raw material; see the freshOps project in our portfolio. Our AI reporting service builds the computed layer and the review step, and the lesson on preparing management reports with AI covers the same principles at smaller scale.

Report quality depends on upstream workflows: complaint KPIs come from site issue triage, and fill rates from roster change communications. The cleaning industry guide shows how they fit together.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.