24/7 AI incident triage and dispatch

What problem was this solving?

Incidents arrive at any hour, in whatever form the reporter can manage — a phone description, a typed message, a photograph of a spill. Someone had to read each one, judge severity, work out which depot was nearest and reachable, and decide whether it warranted waking a duty principal. That judgement was consistent when the experienced people were on shift and less consistent when they were not.

What did we build?

Continuous AI incident triage across four Irish depots: natural-language and image severity classification, Eircode geolocation, dispatch routing and duty-principal escalation.

• Natural-language severity classification of free-text incident reports against the operator's own severity ladder.
• Image classification for photographed incidents, used as corroborating evidence rather than as the sole decision input.
• Eircode-based geolocation resolving a reported location to coordinates and to a serving depot.
• Dispatch routing across four depots, accounting for depot coverage rather than straight-line distance alone.
• A live operational map giving the duty team a single current view of open incidents.
• Duty-principal escalation triggered by severity threshold, by low model confidence, or by incident type.
• An insurance-claim pathway so that incidents with a claims dimension carry the evidence trail from first report.

What does it integrate with?

The integration surface technical buyers ask about.

• Inbound incident capture from multiple report formats, including photographic evidence.
• Eircode resolution feeding depot assignment.
• Escalation to on-call staff outside business hours.
• Evidence retention structured for downstream insurance claims.

How does the system actually work?

The decision path in order. Client-confidential detail is deliberately excluded.

• 1. Report received — Free text, structured fields or photograph, from any hour of the day.
• 2. Classify — Severity inferred from language; photographs classified as corroborating evidence.
• 3. Locate — Eircode resolved to coordinates and matched to the serving depot of four.
• 4. Route — Dispatch assigned by depot coverage, with the live map updated for the duty team.
• 5. Escalate or act — Severity threshold or low confidence wakes a duty principal; otherwise the job proceeds.
• 6. Evidence retained — Report, classification and imagery retained for the insurance-claim pathway.

What made it hard?

The real engineering constraints on this build.

• Operational, not advisory: an unhandled failure means an incident sits unactioned, so every AI decision path has a deterministic fallback to human triage.
• Low confidence must escalate rather than guess. The system is designed to be wrong in the direction of waking someone up.
• Runs continuously, including nights and weekends, so the design target was predictable behaviour under low supervision rather than peak-hour throughput.
• Photographic evidence quality is uncontrollable — a phone photo at night in rain is a normal input, not an edge case.

What was the outcome?

The system has been in continuous service across four Irish sites. Verdé have not authorised publication of internal response-time or cost figures, so we do not publish any — what we can state is that triage, depot assignment and out-of-hours escalation now follow one written decision path rather than the judgement of whoever is on shift. No measured performance figures are published for this engagement because the client has not authorised their release.

What capability does this demonstrate?

Engagement band: Production AI systems band.

• Multimodal classification (text + image) in a production decision path
• Geospatial resolution and multi-site dispatch routing
• Confidence-gated human escalation
• 24/7 operational reliability engineering
• Evidence chain suitable for insurance use

What happens when the AI misclassifies an incident?

Misclassification is treated as an expected event, not a defect. Severity thresholds and confidence gating are set so that an uncertain classification escalates to a human duty principal rather than resolving itself quietly. The failure mode we engineer for is an unnecessary escalation, never a missed incident.

Does the AI make the dispatch decision on its own?

It makes the routine ones and defers the rest. Routing to the serving depot is automated because it is a deterministic geographic question once the Eircode resolves. Severity judgement above the escalation threshold always reaches a person.

How do you build something like this without breaking live operations?

We run a shadow-mode pilot first. The system processes real incidents in parallel with the existing human process and produces its recommendation without acting on it, so the operator can compare decisions on their own data before anything goes live. That method is described in full on our delivery page.

What does a system of this scope cost?

Work of this class sits in our production AI systems band. Published engagement bands, with stated assumptions and timelines, are on the AI integration pricing page — we publish the ranges rather than hiding them behind a call.

A note on what is published here

Case studies on this page describe work Digital Bridge delivered. Client-confidential detail, credentials and internal data structures are deliberately excluded. Where a client has not authorised publication of measured figures, outcomes are described qualitatively rather than quantified — we do not publish numbers we cannot evidence. Last reviewed 2026-08-18.

Related reading

• How we de-risk a production AI build
• AI engagement bands and pricing
• AI integration services
• Funding routes for AI projects

Other builds in the estate

Fifty-plus shipped tools on one platform, including two proprietary scoring algorithms and production generative AI running server-side. A full web-to-print platform with a live design configurator, print-file automation and twenty-nine-plus structured product pages. Generative redesign from a single photo, paired with a live multi-retailer price engine that keeps the result inside the user's budget. Jo McAteer, Forever Home Sanctuary, Riversdale House, AI Triangulate, Equi Consignment, RCP Psychotherapy and Time To Change — grouped by what the engineering actually required.

Could this pattern work in your operation?

Every production build starts with a seven-day shadow-mode pilot, so you see the system's decisions against your own data before it touches anything live.

What does the incident triage system for Verde Environmental Group do?

The system takes incident reports as they come in and classifies them by severity and required response, so the reports needing immediate attention are surfaced first rather than sitting in the same queue as routine paperwork. In an environmental services operation, the cost of a serious incident being noticed hours late is genuinely high, so the triage logic was built conservatively: anything ambiguous is classed as higher priority rather than lower, on the basis that a false alarm reviewed quickly costs far less than a real incident missed.

What are the safeguards and limits of an automated triage system?

No incident is closed or resolved by the system itself; triage only ever changes the order in which a person sees reports and flags which ones need attention fastest, with every classification reviewable and overridable by staff. Incident data is handled under GDPR with a defined retention period appropriate to the sensitivity of the material involved. The honest limit, stated plainly to the client from the start, is that an automated triage system reduces the time to notice a serious incident; it does not and should not make the judgement about how to respond, which stays entirely with trained staff following the organisation's own procedures.