> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vh3.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# The harness is the product
> The runtime harness, quality loops, and deployment work that make field service AI reliable
Riccardo Vezza ยท August 2026
A capable model can summarise a job note, draft a briefing, and spot an obvious repeat visit. Live operations add inconsistent customer names, portal emails that do not match job records, worksheets that say "resolved" before the site books again, and actions that require a human before they reach a customer.
Weekly operational work tests more than model quality. This guide defines the runtime harness, the quality loops that measure it, the capability that compounds in an account, and the deployment work that improves the next release.
## What applied AI means in field service
The useful test is simple. Can the system complete operational work reliably enough for people to trust it in the weekly rhythm?
Field service organisations already use systems of record for jobs, finance, and CRM. They need a layer beside those systems that:
* Reads the operational picture across the tools they run
* Prepares context before a person or agent asks a question
* Watches for problems while the team works elsewhere
* Routes ambiguous or high-impact work for human review
* Leaves evidence that the business can inspect
VH3 AI runs beside the FMS. The FMS records job creation and closure. VH3 AI enriches the work, connects its history, and exposes the result to Connie, sentinels, reports, automations, and agents that use the same substrate.
In this guide, the **substrate** names the decoupled, persistent operational knowledge graph and vector index in the customer's account. VH3 AI stores enriched jobs, resolved entities, linked history, cases, sentinel definitions, and operating rules there. Model context and generated text belong to ephemeral inference. They can read the substrate, but they do not replace its persistent records.
Your team owns thresholds, cases, approvals, and judgement. VH3 AI supplies the runtime infrastructure that lets people and agents act on those decisions.
## The harness
A job-note summary demonstrates language generation. A production workflow ingests years of history, resolves "Tesco PLC" and "Tesco Express Watford" when the evidence supports one customer, watches SLA drift overnight, classifies a portal email, and gives an account manager a cited briefing thousands of times each week. The harness makes that workflow repeatable and auditable.
Treat an AI system as three layers.
1. **The model.** It supplies reasoning and language.
2. **The interface.** It presents chat, digests, Connect, Slack, or custom applications.
3. **The harness.** It supplies the runtime controls, data access, routing, and review paths that let the model work inside real systems.
The harness gives an operator a repeatable workflow against the same systems and history. It requests review when a task crosses a defined risk boundary.
The harness has five runtime responsibilities. It provides domain tools, connects third-party integrations, maintains persistent graph and vector memory, routes tasks deterministically, and enforces human oversight.
### 1. Domain tools
Models reason about a task. They need functions that retrieve operational facts and take bounded actions.
VH3 AI exposes domain tools for hybrid search, graph traversal, aggregates, customer knowledge, sentinel runs, report generation, case creation, connected finance lookups, and briefing assembly. Deterministic tools handle lookups and calculations. Model synthesis handles narrative and judgement. Connie uses these tools internally, and builders call the same capabilities through the API, MCP, and n8n.
### 2. Third-party integrations
Field service work crosses multiple systems. Jobs live in the FMS. Portal traffic reaches shared inboxes. Certificates sit in SharePoint or Drive. Invoices live in Xero or QuickBooks. Account notes sit in a CRM.
The harness connects those systems without requiring one replacement platform. BigChange supplies live operational ingestion today. Native connections add communication, storage, CRM, and accounting context. Managed ingestion and entity resolution attach incoming records to the right customer, site, and job.
### 3. Persistent graph and vector memory
An account review and a multi-day investigation continue after one model call. The substrate keeps the working record outside a chat session.
VH3 AI stores enriched jobs, resolved entities, graph links, vector search representations, cases with evidence, sentinel definitions, and company operating rules in the customer's account. People and agents can resume work across shifts. The model reads prepared records and performs ephemeral inference only for the part of the task that needs language or judgement.
### 4. Deterministic task routing
The harness selects a runtime path from the task type. Some work needs no model call.
| Work | Runtime path |
| ----------------------------------------------------------- | --------------------------------------------- |
| Precedent search, aggregates, entity lookups | Deterministic discovery |
| Repeat visits, SLA clusters, dormant accounts | Background sentinels on the graph |
| Portal email with high extraction confidence | Automated triage |
| Ambiguous triage, investigation narrative, Connie questions | Cited model synthesis |
| Irreversible customer-facing actions | Human approval inside cases and review queues |
Task routing keeps latency predictable, costs visible, and quality measurable. A single chat call should not handle work that a stored query or sentinel can complete.
### 5. Human oversight
Field service organisations need bounded autonomy. The harness sends low-confidence triage to review, assigns case owners, carries job references into Connie answers, and records the tools that each agent session used.
BYOK keeps model spend on the organisation's provider account. Access rules scope each request to a tenant. Approval rules identify actions that need a person. Cases retain the evidence and decision after the workflow ends.
Model improvements can remove brittle extraction prompts and manual calculations. The harness still needs to act inside real systems, persist state across days, connect imperfect APIs, and show who approved each consequential action.
## Quality loops
VH3 AI builds checkable quality into the platform from day one. The team measures the work against real examples, then widens autonomy only when the results hold.
**Start from real work.** Onboarding loads up to five years of history so the first discovery sprint runs against the customer's jobs, notes, and relationships. People who know the operation grade the day-one output.
**Build reference examples with the people who own the work.** Operators approve portal email mappings, agree the risk story for selected accounts, and define what counts as a repeat visit for each fault class. Those examples guide enrichment, triage, briefings, and sentinel thresholds.
**Raise the grade before widening autonomy.** High-confidence triage can create work automatically. Ambiguous extraction stays in a review queue with ranked suggestions. Operators tighten sentinel thresholds after they confirm the signal.
**Keep the loops running after go-live.** New job types, renamed customers, seasonal patterns, and portal format changes create drift. Continuous sentinel runs, cited Connie answers, and case outcomes show whether the workflow still meets its grade.
The standard stays operational. A briefing must cite the right jobs. Triage must match the right customer and site or reach review. A sentinel must fire on a real pattern, or the team must adjust its threshold.
[AI you can check โ](/guides/ai-you-can-check)\
[Agent observability โ](/guides/agent-observability)
## What compounds
Field service contractors need to keep customer relationships, commercial exceptions, and safety-adjacent work inside their own account. The substrate stores the capability that a deployment builds.
* Enriched job history and resolved entities
* Sentinel definitions tuned to the operation
* Cases, evidence, and decisions that form searchable precedent
* Triage taxonomies and review habits
* Automations and agent kits that the team wires into its weekly rhythm
* The field intelligence lead's judgement about the next workflow
An agent may complete a workflow once in a demonstration. The harness and the quality loop keep it useful across changing job types, portal formats, and seasonal load.
Intelligence compounds in the customer's account and remains portable. Model spend stays on the organisation's provider key. If the organisation leaves, its enriched operational data comes with it.
[Compounding operational capability โ](/guides/compounding-operational-capability)
## Deployment should feed the product
Hard deployments begin as discovery work. No playbook matches every field service operation, so early work must sit close to dispatch, account management, and the inbox.
Custom engineering creates value when it reveals the next platform capability. Examples include a better triage confidence tier, a sentinel definition, a case pattern, an agent kit, or a Connect surface. The work loses value when the team answers every customer request with a one-off fix that never returns to the product.
VH3 AI uses a product-led deployment loop.
1. **Discover on real history.** Connect the FMS, enrich the record, and let operators grade what the layer surfaces.
2. **Start with standard surfaces.** Use Connie, sentinels, reports, cases, triage, n8n templates, and agent starter kits.
3. **Give the customer one bridge role.** The [field intelligence lead](/guides/2027-blueprint) stays close to the work, tunes thresholds, and ships the next workflow.
4. **Absorb repeated pain into the platform.** A recurring problem becomes a capability that the next tenant can use.
Implementation work still matters. Connecting Slack, wiring a digest, and configuring exclusions execute a known shape. Discovery work watches the first enrichment miss a local job-type convention, reviews a messy QBR pack with the account team, and identifies portal emails that must never auto-create work. Keeping those activities distinct shows whether the team invests in product capability or grows a permanent delivery organisation.
Each deployment should shorten the next one.
## What pulling ahead looks like
Field service generates operational signal and operational judgement. People have kept much of that value in their heads, inboxes, and FMS records.
Organisations that pull ahead use a harness on the stack they already run.
* Prepared context before inference
* Tools and integrations that reach the real systems of work
* Oversight matched to operational risk
* Quality loops that prove the work still holds
* A deployment motion that turns field pain into shared product capability
VH3 AI provides the intelligence layer beside the FMS, the harness that people and agents share, and the account record that compounds as models change.
## What to do next
How prepared relational, semantic, structured, and temporal memory work together.
How humans and agents share one foundation on the stack you already run.
Why evidence and definitions matter more than confident prose.
How the field intelligence lead turns deployment learning into owned capability.
What accumulates in your account as the learning loop runs.
Connect real history and grade what the layer produces on your own data.
Start a bounded discovery sprint with one workflow. Load representative history, define ten reference examples, and require human approval for every customer-facing action until the workflow meets its agreed quality grade.