A solar administrator opens an installation file, finds an empty document field and writes another follow-up email. Later, they read a long message thread to work out who promised to send the document. These small tasks are sensible places to explore AI assistance because the team can define the inputs, inspect the result and decide what happens next.
Start with one queue your team already understands. Keep the original records available, make unfinished work visible and measure the effort required to produce an accepted result. The following workflow is a practical proposal for a controlled trial, rather than a claim about results any business will achieve.
Use fixed rules for fixed checks
If your procedure requires a document link before a job enters review, a conventional rule can flag an empty field. A scheduled reminder can trigger when an assigned follow-up date arrives. Neither task needs a language model to decide whether the condition is true.
AI becomes a candidate when the task involves language: drafting a request from a confirmed list of gaps, summarising approved correspondence or suggesting a document label for someone to check.
Keep three statuses separate: missing, received and checked. A populated attachment field establishes that something was uploaded. It does not establish that the document is correct, current or sufficient for its intended purpose.
Have the process owner define required fields and exceptions. A model should not invent those rules or silently change a job to “complete”.
Give each assistive task a reviewer and evidence
Choose a small set of uses with a clear review step. This table is an original starting point for a solar admin team:
| Use case | Proposed assistance | Human reviewer | Evidence to retain |
|---|---|---|---|
| Missing-document request | Draft wording from the verified gap list | Assigned administrator checks recipient, requested item and tone | Gap list, approved message and sending record |
| Internal handover summary | Summarise selected correspondence with source references | Coordinator checks dates, promises and unresolved questions | Source messages and accepted summary |
| Suggested document label | Propose a label from the supplied document | Administrator checks the source and job match | Original file and confirmed label |
| Follow-up reminder | A fixed date rule creates the reminder; AI may draft its wording | Queue owner checks whether a reply already arrived | Due date, current status and next action |
During the trial, keep generated text in a draft area. Give reviewers access to the source beside the suggestion so checking does not depend on memory.
Walk a missing document through the process
Consider a hypothetical installation file whose document index has no linked installer statement. The team's agreed rule flags the gap. The administrator checks the approved storage location and recent correspondence before requesting it, avoiding a duplicate chase if it has already arrived elsewhere.
If it remains missing, an approved AI tool could draft a short request using the item name and the team's template. The reviewer checks the wording and adds any permitted job details through the approved workflow. A useful instruction is: “Use only the supplied facts. Leave unknown details blank. Do not state that the installation is compliant.”
The responsible person supplies the actual document. The administrator records receipt and routes it to the required reviewer. Any inconsistency goes back for clarification; the model cannot supply the missing declaration or reconstruct evidence.
The Clean Energy Regulator sets out installer attendance evidence and written compliance statement requirements for rooftop solar systems under the STC scheme. Those responsibilities remain with the relevant people. Read CER's installer and designer guidance. For organising the surrounding file, see our STC evidence pack guide.
Approve the information flow before using customer records
The OAIC advises businesses to assess an AI product's suitability, access to information, privacy risks and human oversight. It recommends, as best practice, keeping personal information, especially sensitive information, out of publicly available generative AI tools. See the OAIC's commercial AI guidance.
Begin testing with fictional jobs and approved generic templates. Before using real records, have the responsible person decide what may enter the selected tool, who may access it and how outputs will be stored. Check the actual service terms and configuration, including data retention, training use and third-party access.
Give the team a written example of an acceptable input. For a drafting task, a verified item name and an approved template may be enough. Avoid uploading the whole job folder simply because it is convenient. Removing a customer's name alone should not be treated as proof that a record is anonymous.
Give the VA authority to question and stop the draft
A trained solar VA can own the daily review queue: checking suggested text, finding source records, following up replies and escalating uncertain information. Allocate work according to capability and authority, with technical questions going to the appropriate qualified person. Our solar VA task guide helps define that scope.
Write down what triggers escalation: conflicting identifiers, an unsupported customer promise, uncertain document relevance or a request outside the approved procedure. Make rejection easy and allow time for review.
The National AI Centre's adoption guidance supports named accountability, testing, monitoring and human intervention appropriate to the use. Read the AI adoption foundations.
In this trial, keep sending messages and changing authoritative job statuses under human control. Continue using the normal manual process whenever the tool is unavailable or its output needs more investigation.
Measure accepted work across the pilot
Record a baseline from a representative set of existing tasks, including straightforward and awkward cases. Then trial the chosen assistance on comparable work. Keep the checklist and acceptance standard consistent so a process change does not masquerade as an AI improvement.
Measure total handling time, including source preparation, checking and correction. Record unsupported details, duplicate follow-ups, wrongly labelled documents and missed gaps. Also count how many outputs reviewers accept without substantive changes.
Set a review date and stopping conditions before starting. An invented technical assertion should trigger investigation, even if average drafting time improves. Expand only when the team can explain the results and maintain quality; a small pilot provides local evidence, not a universal productivity forecast.
Contact Nexus about solar administration support to define a workflow, reviewer responsibilities and a practical first trial for your team.
