Demos are easy. Systems that hold up under real traffic, real data and real governance are not. The examples below show the kind of applied AI we design, ship and operate — the problems we take on and the way we approach them.
A large share of our production deployments serve automotive businesses across APAC, and New Zealand especially — from national dealer groups to used-import specialists.
Challenge — A New Zealand dealer group with sites nationwide was drowning in fragmented vehicle data: OEM spec sheets, used-import histories from Japan, finance and insurance products, and shifting Clean Car and NZTA compliance rules. Sales staff couldn't answer buyer questions confidently, and compliance risk sat with individual salespeople.
Solution — We built a grounded RAG assistant over inventory, vehicle history and regulation, plus an agentic layer that checks Clean Car obligations, drafts finance-ready summaries and flags import-compliance gaps — all tuned for right-hand-drive, used-import realities and hosted with NZ data residency.
“They understood the New Zealand market — used imports, Clean Car, the lot. That local knowledge is why it actually works on the floor.” — General Manager, national dealer group
A cited RAG assistant sitting on top of 120k compliance documents — so a lending team stops hunting through PDFs and gets grounded answers with sources.
Compliance analysts spent hours tracing rules across regulator circulars, internal policy and product manuals. Answers were inconsistent, hard to audit, and impossible to scale as the rulebook changed weekly.
We built a retrieval pipeline with section-aware chunking, hybrid semantic + keyword search and a re-ranking stage, then constrained the model to answer only from retrieved passages — every sentence carries an inline citation to the source clause. A nightly ingestion job keeps the index current, and an evaluation harness gates each release against a labelled question set.
“We stopped guessing where a rule lived. The assistant hands us the clause and the citation — audit trail included.” — Head of Compliance, lending platform
A tool-using agent that triages shipment exceptions and drafts resolutions across five internal systems — with a human approving anything that moves money.
Ops coordinators juggled a TMS, a WMS, a carrier portal, email and a ticketing tool to clear each stuck shipment. Context-switching was the job; exceptions piled up overnight and SLAs slipped.
We designed a planner-executor agent with typed tools for each system, strict input/output schemas and a policy layer that decides what it may do autonomously versus escalate. It reads the exception, gathers state across systems, proposes a fix, and either executes low-risk actions or routes a pre-filled resolution to a human. Every step is logged and replayable.
“It does the boring detective work across five tools and hands us a decision. My team runs the exception queue, not the other way round.” — Director of Operations, freight network
A generative pipeline that turned a spreadsheet of raw specs into 40k on-brand product descriptions — with human review kept exactly where it earns its keep.
A fast-growing catalogue shipped with thin, inconsistent copy. Manual copywriting couldn't keep pace with new SKUs, and outsourced batches drifted off-brand and off-tone.
We built a structured generation pipeline that pulls attributes per SKU, generates copy against a brand voice guide and per-category templates, then runs automated checks for tone, banned claims and SEO structure. Anything low-confidence or high-visibility is queued for a human editor — the rest publishes on approval, with every generation versioned for audit and rollback.
“Our editors stopped writing from scratch and started approving. Same voice on every page, at a scale we couldn't touch before.” — Head of Content, online retailer
A copilot that drafts visit notes from the consultation and cites the clinical guideline behind each recommendation — giving clinicians their evenings back.
Clinicians lost hours a day to documentation, and note quality varied. Any AI here had to be grounded, conservative and never invent a clinical claim — accuracy and traceability were non-negotiable, and nothing could be autonomous.
We combined structured note generation with retrieval over an approved guideline library. The copilot drafts a SOAP note from the encounter, grounds every recommendation in a cited guideline, flags anything uncertain, and hands a full draft to the clinician to edit and sign. Nothing enters the record without a human signature, and every draft is evaluated against clinician-reviewed gold notes.
“It writes the first draft, I stay in control of the record. The citations are what got it past our clinical governance board.” — Chief Medical Officer, digital health provider
A RAG-backed agent that answers what it can from the docs and routes the rest — with priority, product area and a suggested reply already attached.
A rising support volume buried the team in repetitive questions, while genuinely urgent tickets waited in the same queue. Canned macros felt robotic and mis-routing sent issues to the wrong pod.
We deployed an agent that first attempts a grounded answer from the help centre and changelog, citing its sources. When it can't resolve confidently, it classifies the ticket by product area and urgency, drafts a reply for an agent to approve, and routes it to the right team. Confidence thresholds and an eval loop keep deflection honest — it never bluffs an answer it can't ground.
“Deflection that doesn't annoy customers, plus triage that actually lands tickets in the right place. Our CSAT went up, not down.” — VP Customer Experience, B2B SaaS
Clause extraction and risk flagging across thousands of contracts — turning a shared drive of PDFs into a queryable, reviewable knowledge base.
A legal team faced renewal deadlines and audit requests with no way to know which contracts held which terms. Manual review of thousands of agreements was slow, and non-standard clauses slipped through until they became a problem.
We built an extraction pipeline that identifies parties, key dates, liability caps, indemnities and termination terms, then flags clauses that deviate from the team's playbook. A retrieval layer lets a lawyer ask natural-language questions across the whole portfolio and jump straight to the governing clause — every extraction links back to its exact location for human verification.
“We went from ‘we think that clause is in there somewhere’ to a searchable answer with the paragraph attached. Renewals stopped ambushing us.” — General Counsel, enterprise software firm
Regulated, high-throughput teams where accuracy, auditability and uptime are not optional.
We don't call a system done because a demo impressed someone. We call it done when it moves a number the business already tracks — and keeps moving it in production.
Illustrative of the outcomes and working relationship our clients value.
“Netatum was the first team that talked about evaluation before they talked about models. That's why the thing actually works six months later.”
VP Engineering, lending platform
“They shipped a working proof-of-value in weeks, then were honest about what to automate and what to leave to people. Rare and refreshing.”
Director of Operations, freight network
“The citations are what won over our governance board. We could see exactly where every answer came from. That built the trust to roll it out.”
Chief Medical Officer, digital health provider
Bring us the workflow, the data and the metric you want to move. We'll tell you honestly whether AI is the right tool — and if it is, we'll ship it.