⚙️
You have a real workflow
Multi-step processes that touch internal systems — pricing engines, inventory, custom databases, proprietary CRMs. The agent needs to do the work, not just answer questions.
SERVICE 03 · CUSTOM AI AGENTAn agent built
An agent built
around your
workflow.
For teams whose work doesn't fit a generic widget. Our engineering team designs, builds, and continuously operates a bespoke agent wired deeply into your stack — 4-6 weeks from kickoff to live.
✓ 4–6 week delivery✓ Embedded engineering✓ Managed forever
Your agent · architecture● LIVE
Who this is forWhen the off-the-shelf
When the off-the-shelf
agent won't cut it.
🏛️
You're in a regulated space
Healthcare, finance, legal, insurance — where compliance is non-negotiable and you need audit trails, signed conversations, and policy guardrails.
🔌
You need deep integrations
If your stack involves SAP, mainframes, custom APIs, or a database that isn't on our integration list — we'll write the connectors and own the maintenance.
🏆
It's a competitive moat
When the agent is part of your product, not just supporting it. We'll work as an embedded team, ship to your release cadence, and stay out of your competitors' deployments.
The buildSix weeks from scope
Six weeks from scope
to shipped.
A senior engineer, a designer, and a customer-engineering lead — assigned to your build, on a Slack channel with your team, shipping daily.
Wk 1
Discovery & spec
Two on-site (or zoom) sessions with your team. We map the workflow, identify the data sources, design the conversation, define success metrics. You leave with a written spec, a Figma flow, and a fixed timeline.
Workflow mapData inventorySuccess criteria
Wk 2-3
Integrations & data layer
We build the connectors to your CRM, database, APIs and knowledge sources. Set up the retrieval layer, write the tools the agent can call, and stub out the conversation skeleton.
Custom connectorsTool definitionsRetrieval pipeline
Wk 4
Conversation engineering
Prompt design, model selection, voice tuning. Build the eval harness against your real historical conversations. Tune until the agent passes a quality gate your team defined.
Prompt engineeringEval suiteVoice / persona
Wk 5
Shadow mode
Agent runs alongside your team in production but doesn't send anything outbound. We grade every conversation against what your team did — and surface the gaps until the agent is genuinely better or equal.
Live trafficSide-by-side reviewNo customer impact
Wk 6
Phased launch
Start at 5% of traffic. Watch the metrics. Tune. Move to 25%, 50%, 100% over two weeks. Human escalation always available. You can roll back instantly if anything looks off.
Gradual rolloutAuto-rollbackSLA monitoring
∞
Operate & evolve
Your CSM watches the agent daily. We tune weekly, ship model upgrades quarterly, and add new capabilities as your business changes. Monthly written report, in plain English.
Weekly tuning99.99% SLANamed engineer
Case studyWhat a custom build
What a custom build
actually looks like.
CASE · NORTHWAVE SOLARFrom 14% lead
From 14% lead
contact rate to 96%.
Northwave gets ~2,000 solar enquiries a week. Their inside sales team could only get back to 14% of them within an hour — the window where conversion is 5× higher.
We built a custom voice agent wired to their pricing engine (rooftop sizing, panel inventory, install windows by postcode) and CRM. The agent calls every new lead inside 60 seconds, quotes, books a site visit, and updates Salesforce.
Build took 5 weeks. Payback in 7 weeks. Their sales team now only takes the qualified, ready-to-buy 22% — and revenue per rep is up 3.4×.
Lead contact rate
14%→96%
Revenue per rep
3.4×
Payback period
7weeks
Under the hood
Production-grade infrastructure, hidden inside.
Model-agnostic routing
OpenAI, Anthropic, Google, Meta, Mistral, or your own fine-tuned model. We pick the right one per task based on cost, latency and quality.
GPT-5ClaudeGeminiLlama
Nightly regression suite
Every model change runs through 500+ test conversations from your real history. If quality drops on any metric, the change is auto-rolled-back.
Trace every decision
Full conversation transcripts, model inputs/outputs, tool calls, and reasoning chains — searchable, exportable, and integrated with Datadog or your own observability stack.
Enterprise security default
SOC 2 Type II, ISO 27001, GDPR, signed BAA / DPA, region-locked data, customer-managed encryption keys on request.
99.99% uptime SLA
Multi-region failover, automatic fallback models, graceful degradation on partial outages. Status page wired to PagerDuty.
Your cloud or ours
Hosted on our SOC 2 infrastructure in AU / US / EU, or deployed inside your own VPC. On-prem available for the regulated industries that need it.
Custom build pricingFrom $25k
From $25k
for the build.
Plus a Scale plan for ongoing operation, model costs, and managed support. Most builds land between $25k–$80k upfront, with monthly fees scaled to your conversation volume.
Discovery + spec
Week 1 fixed fee, refundable
$5,000
Standard build
4-6 week delivery, fixed scope
$25k–$80k
Ongoing operation
Scale plan + usage
$3k+ / mo
FAQ
Common questions.
Yes. Custom prompts, evals, integrations and any fine-tuned models built for you remain your IP. You can export everything and run it anywhere — we just hope you don't, because operating it yourself is the boring part.
Let's scope it.
Tell us about the workflow. 30 minutes, no pitch — at the end you'll have a rough timeline, rough price, and a written next-step.