The Programme
By Operation
We design, build, and run multi-agent voice systems that pick up the call, understand the customer, and actually resolve the problem. Integrated with your CRM and telephony. First production call in about 8 weeks.
If it follows a process, has a clear answer, and lives in a system you already use, a voice agent can handle it. Support, bookings, payments, qualification, recovery. Inbound and outbound.
Answer every call, 24/7. Handle routine enquiries, create tickets in your CRM, authenticate callers, and hand off to human agents with full context when needed.
See Roof Maker case studyAppointment reminders, lead qualification, no-show recovery, payment collections, satisfaction surveys. Run thousands of concurrent calls with consistent quality.
Integrate with your calendar system to book, reschedule, and confirm appointments by voice. Ideal for clinics, dealerships, salons, and professional services.
PCI DSS compliant voice agents that take card payments, confirm balances, and handle collections conversations with fully logged, audit-ready transcripts.
See arrears recovery case studyAutomated candidate screening and structured interviews at scale. Transcribe, score, and route high-intent applicants to human recruiters.
See AI Voice Interviewer case studyHandle your main line: triage enquiries, route internally, take messages, book meetings, qualify leads. Replace or augment a traditional front desk.
Production deployments across regulated, high-volume, customer-facing operations.
Technical support hotlines that read serial numbers, log warranty claims, and create HubSpot tickets while the customer is still on the line.
Appointment reminders, prescription refill triage, and post-visit check-ins, with consent capture and PII redaction by default.
Polite, compliant arrears recovery and balance enquiry calls. Outbound and inbound. Structured outcomes back into your CRM.
Voice interviewers that ask 65+ adaptive questions, handle three languages, and decide on the fly when to probe deeper.
Lead qualification, viewing and test-drive bookings, dealer hotline triage, with calendar and CRM integration end to end.
Order status, returns, product questions, and recovery calls when a customer hangs up mid-checkout. Multi-language by default.
Oliver is our own multi-agent customer support system, the same architecture we deploy for clients in healthcare, finance, and customer service. He runs every conversation on this website by voice or text.
Behind Oliver: seven specialised agents working together. A root agent triages every incoming conversation; four domain specialists handle AI projects, discovery, general enquiries, and bookings; two deeper specialists under discovery cover team augmentation and general development. Each agent has its own instructions, tools, and end conditions.
Voice systems running today, handling real conversations at scale.
UK manufacturer deployed a multi-agent voice system handling technical support, product enquiries, and ticket creation 24/7. A router agent classifies every call and routes to one of seven specialists: technical support, sales, delivery, information, lead capture, issue tickets, or delivery timelines. Built on Google Agent Development Kit, LiveKit, Vertex AI Search, and HubSpot CRM.
Confidential HR Tech client. A multi-microservice voice interviewer that generates 65+ adaptive questions per role, conducts the conversation in three languages, and decides on the fly when to probe deeper or move on. Built on LiveKit, Google ADK, and a streaming LLM stack.
A polite, compliant outbound voice agent for a US consumer finance platform. Bidirectional outbound and inbound, with structured outcome capture and CRM integration.
Five layers, each using the strongest available tool for the job, working together in real time.
LiveKit for real-time audio, Deepgram or Google STT for speech recognition, ElevenLabs or Google TTS for voice synthesis. Streaming end to end.
The strongest LLM for each task: Gemini, Claude, or OpenAI GPT, selected per use case. Google Agent Development Kit and LangGraph for stateful workflow orchestration.
Router agent for intent classification. Specialised sub-agents per domain: support, sales, booking, billing, lead capture. Each with its own tools, prompts, and end conditions.
SIP trunks, Twilio, Vonage, Genesys, Amazon Connect for telephony. Salesforce, HubSpot, Zendesk, ServiceNow, and custom APIs. Reads and writes records during the call.
Google Cloud Run for auto-scaling serverless deployment. BigQuery for analytics. Cloud Logging for full audit trails. Deployed in your UK or EU region.
Every voice agent we ship runs on this same five-layer architecture. Roof Maker uses it. The HR Tech voice interviewer uses it. The arrears recovery agent uses it. The components shift to fit the use case, but the shape of the build is the same every time.
Every call produces a transcript, a structured outcome, and a tool-call trail. Each one lands in BigQuery alongside metadata: agent, intent, resolution, sentiment, time-to-resolve, and escalation point. Every data point is kept.
On top of that we run an automated QA layer: random sampling, rule-based checks, and LLM-as-judge scoring against your own quality rubric. Failures become coaching items for both the agents and your human team. Patterns become product changes.
The result: calls become data, data becomes improvements, improvements become better calls. You see exactly how the agents get sharper week over week, with the numbers in front of you.
Transcripts
100%
Stored and searchable
Structured outcomes
Per call
Intent, resolution, escalation
QA sampling
Automated
Rule + LLM-as-judge scoring
Improvement cycle
Weekly
Numbers, not vibes
The key facts
Softblues builds custom multi-agent voice systems for UK and Ireland businesses. Inbound and outbound. Integrated end to end with your CRM, your telephony, your back-office, and your compliance stack. Deployed on Google Cloud in your UK or EU region. Ship the first production call in about 8 weeks, then tune.
Eight live agents are already in production for Roof Maker. An HR Tech voice interviewer runs adaptive interviews in three languages. A US arrears recovery agent handles outbound and inbound calls with structured outcome capture. The architecture is the same every time, shaped around your operation.
Speech accuracy
97%
Measured on production calls across UK accents and noisy lines.
Live in production
8 agents
Roof Maker router plus seven specialists, handling real calls 24/7.
To first production call
8 weeks
Discovery, build, integrate, test. Real conversations, not presentations.
Always on
24/7
No queue, no out-of-hours gap, no overflow. The agent is always there.
Three phases. Fixed time per phase. You hear a real agent on a real call before you commit.
Every engagement follows the same shape: a short Pilot that proves it works on your top call category, a full Production build that covers your top three to five categories, and an optional Managed Optimisation Retainer that keeps tuning after go-live. You see the path before you sign, you hear the agent on real calls before you commit, and you keep the work whether you stay with us or not.
Prove it works on one call category
We take your highest-volume or highest-pain call type and build a working voice agent around it. Connected to your telephony, reading and writing to your CRM, tested on real calls. At the end of week four you hear the agent handle a live conversation and decide whether to go further.
Go or no-go decision with the numbers in front of you.
Scale to your top 3 to 5 categories
Router agent plus specialised sub-agents for each call type. PII redaction, recording consent, audit logs, and CRM writes all included. First production call typically lands within about 8 weeks of kickoff.
A multi-agent voice system handling real customer calls.
Optional, ongoing after go-live
Prompt tuning, automated QA scoring, compliance monitoring, new agent additions, and language expansion. Most clients take it because the second-month gains are usually larger than the first.
Continuous improvement without tying up your own team.
Three phases. Fixed time per phase. You know the timeline before you sign.
Single call category
3 to 4 weeks
Single-flow proof of concept on your top call category. Working voice agent on your data, your CRM, your telephony.
Multi-agent rollout
6 to 8 weeks
Multi-agent, CRM and telephony integrated voice system across your top 3 to 5 call categories.
After go-live
Ongoing
Tuning, QA, monitoring, compliance, and ongoing agent additions after go-live.
Fixed-time delivery for each phase. Timelines confirmed at the end of discovery.
How a Softblues custom voice agent compares to other options on the market.
| What matters | Softblues custom | SaaS voice platform | In-house team | Large SI |
|---|---|---|---|---|
| Time to first production call | About 8 weeks | Days to weeks | 6 to 12 months | 6 to 12 months |
| Fits your processes exactly | Yes, designed to your flows | Generic flows only | Yes, but slow | Yes, but expensive |
| Deep CRM read and write | Per-object writes during call | Limited | Yes | Yes |
| UK or EU data residency | Configurable per deployment | Vendor default | Yes | Yes |
| Pricing model | Fixed-price phases | Per-minute or per-seat | Internal cost | Time and materials |
| Specialised telephony or compliance | Supported | Limited | Possible | Supported |
| Ongoing optimisation | Managed retainer after go-live | Self-serve console | Internal team | Day-rate change requests |
Built for UK and EU compliance from day one, not added as an afterthought.
Deployed on Google Cloud in your chosen UK or EU region. Audit-ready logging end to end.
Consent capture at the start of every call. Configurable retention. PII redaction before storage.
Aligned with UK GDPR, the UK Data Protection Act, ICO AI guidance, and the EU AI Act.
PCI DSS controls for card-not-present voice flows, with DTMF redaction and tokenisation.
Healthcare-grade configurations with BAAs and segregated environments where required.
Every transcript, outcome, and tool call lands in BigQuery for QA, BI, and model improvement.
Three facts you can verify before you book a call.
Oliver, our own voice and chat assistant, handles every conversation on this website. Seven agents, one router, the same architecture we deploy for paying clients. You can talk to it right now.
Roof Maker (8 agents, 97% accuracy), an HR Tech voice interviewer (300+ interviews per week, 3 languages), and a US arrears recovery agent (outbound and inbound, CRM-integrated). Real calls, real outcomes, real data.
Every engagement is fixed-time: Pilot in 3 to 4 weeks, Production in 6 to 8 weeks, optional Retainer after that. Built on your CRM, your telephony, your data. UK or EU residency by default.
Everything buyers, IT leads, and compliance teams ask before signing.
Book a discovery call. We spend 30 minutes mapping your top call categories, current telephony, CRM, and the rules a voice agent would need to follow. If there is a real fit, we move into a 3 to 4 week Pilot on a single call category. You see a working agent on your data, your CRM, and your numbers before any commitment to a full production build.
An AI voice agent holds a natural, real-time conversation. It understands intent, asks clarifying questions, retrieves data from your systems, takes actions such as booking, payment, or ticket creation, and hands off to a human with full context. IVRs are rigid menus. Chatbots are text-only. Softblues builds voice agents that combine speech recognition, large language models, and real-time voice synthesis to sound and behave like a trained agent.
Three phases, with a fixed time per phase. The Pilot on a single call category takes 3 to 4 weeks. The full Production build covering your top 3 to 5 call categories takes 6 to 8 weeks, with the first production call typically live within about 8 weeks of kickoff. An optional Managed Optimisation retainer runs continuously after go-live, covering tuning, QA, monitoring, compliance, and ongoing agent additions. You know the timeline before you sign.
Deployments run on Google Cloud in the UK or EU region of your choice. Consent capture happens at the start of the call, recording retention is configurable per use case, and PII redaction is applied automatically before storage. We align with UK GDPR, the UK Data Protection Act, GDPR, ICO AI guidance, and the EU AI Act. PCI DSS controls for payment voice flows and HIPAA-eligible configurations for healthcare are both supported.
We connect through SIP trunks for PSTN, or directly with Twilio, Vonage, Genesys, Amazon Connect, Zoom, or Microsoft Teams. CRM and ticketing integration is built for Salesforce, HubSpot, Zendesk, ServiceNow, and custom APIs. The agent reads and writes records in real time during the call, including creating tickets, updating opportunities, and logging structured call summaries.
End-to-end response latency is typically under 800 milliseconds, within the range of a natural human conversation. We use streaming speech-to-text, low-latency LLM inference, and streaming text-to-speech so the agent starts speaking as soon as it has a response, not after the whole reply is generated.
Yes. Inbound agents answer customer calls, handle support, take bookings, and route to humans. Outbound agents run campaigns: appointment reminders, lead qualification, payment collections, satisfaction surveys, no-show recovery. Both share the same underlying multi-agent architecture.
Our stack supports 40+ languages and regional accents, including direct audio-to-audio translation. We choose the speech-to-text, text-to-speech, and LLM combination that best fits your target markets. Deployments in English, German, French, Spanish, Italian, Polish, Ukrainian, Japanese, and Mandarin are all in production.
Book a discovery call. We map your highest-volume call type, build a working voice agent around it, connect it to your CRM and telephony, and put it on real test calls. You hear it work before you commit to anything else.
Google Cloud Partner. Production voice AI across manufacturing, HR, and consumer finance.
