Operations & Reliability
Order book, OTIF drift, equipment readings, downtime clusters, PO slippage, and material cover.
OperationsReliabilitySupplier Risk
See the examples →Agents
The agents we run
on your behalf.
Named AI programs, each owning a domain of your operation. We configure them, run them, and tune them over time — they watch, flag, draft, and act on schedule or trigger, while your team retains final say on every consequential action.
Overview
What your agents run.
A Master Agent synthesises across departments. Department Agents each own one function — operations, quality, procurement, reliability, estimating. Sub-Agents are the specialists within each department. The hierarchy is deny-by-default: a sub-agent in procurement cannot reach operations directly — coordination happens via a shared event bus, asynchronously, with a full audit trail.
Margin, Procurement & Inventory
Quote margins, job actuals, inventory cover, consumption drift, and threshold metrics.
Estimating & MarginInventory IntelligenceMetrics & Threshold
See the examples →Quality, Customer & Compliance
Delivery performance, defect spikes, lot genealogy, compliance certs, and customer complaints.
Customer AgentQuality & CAPATraceability
See the examples →How an agent works
Five faculties. One governed loop.
Every agent perceives, reasons, acts, stays inside the approval rules, and remembers — the same loop whatever the domain.
01
Perception
It watches your operation — continuously.
Connected to live operational data from day one — every record, threshold, and condition monitored without interruption.
Scheduled cadence
Morning briefings, weekly scorecards, monthly statements. The agent runs on the rhythm your operation already uses.
Event-driven
Fires the moment a record is created, updated, or a watch condition is met — inside the same minute.
On demand
An operator triggers a run manually for exception handling, scenario checks, or urgent investigations.
02
Reasoning
It reasons — not just reacts.
When something changes, the agent maps it against operational baselines, prior decisions, and KPI history scoped to its role. An LLM reasons over this structured context — with rationale produced before any action is taken. When the conclusion spans departments, it emits an event on the shared event bus; the affected department agent picks it up on its next tick, asynchronously and fully audited.
01PerceiveReads current state across all connected systems
02ContextualiseMaps change against baseline and domain memory
03PlanDetermines what action is needed and why
04Select ToolChooses which tool contract to invoke
05ActExecutes through granted tools or escalates for approval
06LogRecords the full trace — trigger, reasoning, outcome
03
Action
It acts through your tool library.
Every action the agent takes is executed through a granted tool — a connection to a system your operation already runs on. The grant model is deny-by-default: an agent with no granted tools has zero action capability. You expand what it can do by granting tools explicitly.
Built-in actions
InternalRead and update your operational data, generate intelligence outputs, trigger monitoring conditions, route escalations.
Analytical scripts
ScriptRun complex calculations, scoring models, or data transforms in an isolated environment. Credentials are injected securely.
External systems
WebhookPush updates to your ERP, legacy systems, or any internet-accessible endpoint via outbound HTTP.
Your own infrastructure
MCPRun the connection inside your own environment. No data leaves your trust boundary.
250+ connected apps
250+ appsSalesforce, Jira, Slack, Gmail, ServiceNow, and 245+ more. Authentication managed. No integration engineering from your team.
Test mode
TestSimulate agent behaviour before going live. Full trace, no real actions taken.
04
Governance
You control how much it does on its own.
Every tool carries a side-effect class. The runtime derives the approval tier from that class automatically — no per-agent configuration needed. Low-risk actions execute and are logged. Consequential actions queue for your team with the agent's reasoning attached, so you know exactly what it is proposing and why before you approve.
AutomaticUpdate an internal status, log an anomaly, tag a record for follow-up.Happens immediately. Recorded in full.
Draft for reviewGenerate the Monday briefing, build a risk scorecard, draft a quotation.Ready when you are — not blocked.
Notify and approveEscalate to a team member, open a support ticket, send an internal alert.Queued for your confirmation before it goes.
Admin-gatedCustomer comms, ERP or financial system updates, any payment-impacting step.Nothing customer-facing or financially material leaves without explicit approval.
05
Memory
It carries everything forward.
Every run is appended to an audit trail — trigger source, reasoning, tool used, approval status, outcome. That history feeds the agent's context on the next run. Baselines sharpen. Exception patterns build. The agent running in month six is informed by everything that came before it.
Append-only action log
Every action, every approval, every dismissal. Trigger, reasoning, and outcome recorded in full — traceable forever.
Context snapshots
Synthesised memory documents the agent carries into each run: what changed, what was decided, what the current baseline is.
Knowledge items
Durable facts, risks, and anomalies identified and retained — available to every future run in the same workspace.
Agents vs Automations
Same runtime. Very different jobs.
Automations do the deterministic, repeatable work — and Zipdata runs them as first-class objects. Agents take on what automation can't: watching continuously, reasoning in context, carrying a decision through several steps, and staying inside the approval rules. You want both.
AutomationAgent
Trigger modelOne declared trigger — one webhook, one cron, one event — wired to one action.Fires from scheduled cadences, entity events, and watch conditions — evaluated continuously through one runtime.
Context awarenessReceives the inputs declared on the trigger payload. Stateless across runs by design.Carries operational baselines, prior decisions, KPI history, and role-scoped entity context across every run.
Decision behaviourDeterministic branch logic. Edge cases not anticipated by the rules fall through.Reasons over current state against the baseline — outputs a decision with the rationale attached.
Output shapeA single completion event — record updated, message sent, webhook posted.A multi-step outcome in one cycle: detect, prioritise, draft, route, and escalate.
GovernancePermissioning lives outside the runtime — in IAM, in the called system, or in the script's secrets.Every tool carries a side-effect class. The runtime enforces the approval tier automatically — no per-agent config.
Adding capabilityEngineering work — a new script, a new trigger, a new run history to maintain.Grant a tool to the role. The loop is unchanged; the action surface expands. Governance tier inherited automatically.
Kinds of agents
Five jobs in the intelligence loop.
Named agents on this site are manufacturing beats — operations, reliability, margin. Underneath, every agent is one of five kinds: edges that move data in and artifacts out, monitors that watch day to day, higher-order agents that synthesise, and agents empowered to act under governance.
01Input edge
Document Processing
Gets unstructured signal into the live picture.
Invoices, claims packs, field photos, multi-format exports — extracted, validated, and written as typed records. Expands what OI can see beyond ERP and spreadsheets.
Scheme-claim ingest · KYC packs · shop-floor photo capture02Output edge
Content Generation
Turns reasoned state into finished artifacts.
Briefings, scorecards, CAPA drafts, customer updates, audit packs — authored for a named audience and cadence. The deliverable ships; the team does not recompile it.
Morning OTIF briefing · 8D draft · held customer update03Day-to-day watch
Monitoring
The junior analyst: tracks, flags, does not editorialize.
Watches metrics and conditions against thresholds. Surfaces deviations with context — who owns it, what crossed, when. Prepares the queue; leaves the judgment call to people.
OTIF drift · blade-wear bands · material cover below X days04Pattern of patterns
Higher-order / trend
The senior analyst: watches the watchers.
Synthesises across monitoring agents and departments. Names what the cluster means — systemic risk, recurring failure mode, margin leak that is not one job. Briefs leadership, not just the floor.
Cross-line downtime pattern · supplier risk compounding OTIF05Empowered to act
External action
The analyst that can do, not only report — under governance.
Writebacks, work orders, supplier notices, ERP updates — executed through granted tools. Low-consequence steps auto-apply; consequential ones queue with reasoning attached for one-tap approval.
Promise-date writeback · CMMS draft · held supplier escalationA single named program often combines kinds — a Reliability Agent monitors readings (03), drafts a briefing (02), and queues a CMMS work order (05). The kinds explain the faculties; the examples below show them on a live beat.
Agents at work
Production control on the floor.
Orders, equipment, and supply — watched continuously so the morning stand-up is a queue of decisions, not a list of surprises. These beats combine monitoring, content generation, and governed action on the live plant. See the five kinds →
Your agents watching.
Your team deciding.
We design the agent roles, configure the trigger model, grant the tools, and set the approval tiers for your operation — not a generic demo. Your first briefing arrives before the meeting.