Using Voltara Intelligence · Lesson 9 · plant corporate admin

Voltara Agents and the AI assistant

What the agents do on their own, how to handle the jobs and decisions they put in front of you, and what the AI assistant can and cannot answer.

By the end of this lesson you can recognise a job an agent raised and decide what to do with it, use the Decision Board correctly, and judge any AI-labelled answer or draft before acting on it.

Warm-up, from memory: (1) Which daily figures does the nightly data check leave unmarked? (Lesson 8) (2) Who may approve a work order you raised yourself? (Lesson 2)

(1) Days typed by hand in the Logbook. (2) Nobody: the raiser can never approve their own job.

Everything below was done on 18 October 2026 in the training copy, running the current code. We ran the agents' night shift and their act step by hand, exactly as the server runs them, then used the screens as Sara (supervisor), Tariq (technician) and Carla (corporate).On the live server the agents' night shift runs once a day and their act step every 30 minutes. The training copy only runs them when told to.

1 · What the agents are

Voltara Agents are 14 programs, each watching one area of the data: performance, data quality, HSE, PM, spare parts, warranty, invoices, attendance and others. They are shown like colleagues, with portraits, job titles and “years declared”, but they are software reading the same records you do.The years of experience on each card are a setting, not a fact about anyone.

AI Agents tab: cards for the Commercial and Energy Claims Manager (deemed energy claim agent, on demand, 10 yrs declared, L2 autonomy), Report Validation and Certification Gatekeeper (Reports AI, 14 yrs declared, L1 autonomy), Portfolio Asset Manager (Asset Manager agent, 22 yrs declared, L1 autonomy), Senior HSE Manager, Instrumentation and Calibration Lead, Insurance and Risk Manager, each with what it produces, counts awaiting you, authorised and rejected, and the next check. — Voltara Agents → AI Agents, as Carla. Every card says “L1 autonomy”, which reads as “it prepares, you decide”. That is not what happens next.

2 · They raise work orders by themselves

Each agent writes up what it finds. Then, unless an administrator has switched it off, it acts: it opens a work order, picks a priority and assigns someone at the plant, without asking. One run in the training copy opened 27 jobs across seven plants. They are never safety actions (no permit, lock, switching or energising), but they are not all repairs either:

Work Order Center rows for Aden Solar: WO-2026-1016 Invoice validation, WO-2026-1021 Preventive maintenance 1 released unfinished, WO-2026-1022 Asset Manager briefing 7 findings, WO-2026-1029 Stale feed, each type CM, priority medium, assigned to Khalid Al-Rashidi or Mohamed Farouk, status Assigned, with the line Raised automatically by the named agent. — Agent jobs in Sara's Work Order Center. Each says *“Raised automatically by the … Agent”* under its title. All are corrective (CM), medium priority, and most were assigned straight away.

What the agent found	Job it raised
Unpaid customer invoices	“Invoice validation: … 31 invoices unpaid after 30 days”
Other agents' open findings	“Asset Manager briefing: 7 findings from the other agents open about this plant”
A plant feed that stopped	“Stale feed: the daily KPI table has no data for 3 days”
Too few toolbox talks	“HSE watch: toolbox talks below cadence”
A part out of stock	“Spare parts: 1 item out of stock (Inverter cooling fan 230 V)”

Agent jobs skip the approval ladder 19 of the 27 went straight to assigned, with no approval record, although every job a person raises waits for approval first (Lesson 2). They also count as corrective work, so Work orders MTD went from 15 to 42 in one run. Treat an agent job as a request: read it, then either do it, pass it to the right person, or cancel it with a reason. This has been reported as a fault.

3 · The Decision Board

Decision Board for Sara: 7 decisions waiting on you, start with Arjun Mehta's feed faults; Can you trust today's numbers: Aden 20%, feed health below 50% means every figure is provisional; what is on your desk: critical 2, high 2, normal 3; all raised today; first item from the Data Quality agent, critical, blocks others: stale feed and missing days at Aden Solar. — Voltara Agents → Decision Board. The board puts the findings in order and says which to decide first. “Can you trust today's numbers?” is the same warning as the dashboard banner (Lesson 8).

Button	What it does	What it does not do
Authorise	Agrees with the finding; recorded with your name	Raise the job: on an agent in Act mode it already exists
Reject	Disagrees, with a reason the agent keeps (a reason is required)	Cancel the job. Sara rejected a finding; its job WO-2026-1036 stayed assigned to the technician
Seen it	Acknowledges, so the agent stops chasing	Agree or disagree
Snooze 24h	Hides it for a day	Anything else

So when you reject a finding, open its job (the finding names it) and cancel the job too, with the same reason.Draft purchase requests and warranty claims that an agent proposes are different: those wait for you, and Accept, Decline and Reverse really do act on the draft. Avoid Authorize all pending: one click agrees with every finding at once, read or not.

4 · The AI assistant and AI drafts

The assistant (the chat panel opened from the round button at the bottom right) answers from the figures on the page you are on. It does not search everything:

We asked	It said
“What is the average availability of the fleet?” (a suggested question)	99.8%, with fleet PR 82.1%: the page's own figures, including any wrong typed days (Lesson 8)
“What was the PR at Aden Solar on 9 October?”	“I cannot answer … Treat this as no answer, not as a negative finding.”
“Can I open combiner box CB-04 without a permit if I am quick?”	The same refusal, and no safety advice. Safety questions go to a person and the permit rules (Lesson 4)

That honesty is good: it says when it cannot answer instead of making something up.On the live server the external AI model is switched off, so free questions get this refusal there too. Every question and answer is logged with its source.

AI Generate WO in the Work Order Center is meant to draft a job from your description. Two warnings:

The button does not work at the moment: it fails with “CSRF token validation failed”, and nothing appears. This has been reported as a fault.
When the AI model is off, the draft comes from a fixed template for that kind of equipment, labelled Static Template. For “DC insulation fault … smell of burnt plastic near combiner box CB-04” it proposed high priority and a cooling-fan checklist. A burning smell is an emergency. Always rewrite the draft from what you know.

AI alarms start with “AI:” (for example “AI: Failure Risk - Central Inverter Block A”, “about 25 in 100 chance of a fault in the next 12 months”). They are statistical estimates from age, alarms and history. Open the asset's Health & AI tab to see what the estimate is based on before you act on it.

5 · Practice

What do you do?

From memory: (1) How do you recognise a job an agent raised? (2) You reject an agent's finding. What else must you do? (3) The assistant says “I cannot answer”. What does that tell you about the fault you asked about?

(1) “Raised automatically by the … Agent” under its title in the Work Order Center. (2) Open its job and cancel it with the same reason, because Reject does not touch the job. (3) Nothing at all: it is no answer, not a sign that all is well.

Your part, by role

You are	You do
field Technician	When an agent job is assigned to you, read it before going out: some are paperwork. Ask your supervisor to cancel jobs that are not real work.
plant Supervisor, plant manager	Check agent jobs daily; keep, reassign or cancel each with a reason. Decide findings one by one on the Decision Board; cancel the job when you reject.
corporate	Never use “Authorize all pending”. Remember agent jobs are in the work order counts you report.
admin	Set each agent's mode (Observe, Propose or Act) from its card (Open profile) or Brain Admin, to what the plants can handle. One fleet-wide setting stops them all acting at once.

Do next (10 minutes), in the app Open the Work Order Center and search for “Raised automatically”. For each open agent job at your plant, decide: real work, someone else's paperwork, or cancel. Then open the Decision Board and decide the first item, reading its finding in full.

Questions? Ask me in the chat, for example “which agent raised this job and why?” or “how do I set an agent to Propose only?”.

Reference: Agents and AI card · Glossary · Previous: Lesson 8 · Next: Lesson 10, Administration