Patrick Michael logoPatrick MichaelVoice AI Engineering

We build and improve Voice AI for business calls

Whether you're starting from scratch or improving an existing Voice AI system, we help you handle calls reliably, give callers the help they need and pass clear information to your team.

Discuss your call workflow

Where are you with Voice AI?

Inspired by Calloptima

CHALLENGE 01

"We want to use Voice AI, but don't know if it will actually work."

We start with one call type and agree on what success looks like. We build a working pilot, test it with simulated calls, then test it on real calls.

We measure accuracy, task completion, operating costs and customer satisfaction (CSAT). The aim is to prove the system works without compromising the caller experience.

Includes

  • Working voice agent
  • Documented call flow
  • Simulated-call tests
  • Scored call results
  • Cost estimates
  • Customer-experience findings
  • A clear go/no-go recommendation for production
CHALLENGE 02

"We've built a voice agent, but it isn't working reliably."

We start by mapping the business objective against what the agent actually delivers. Then We assess whether the existing technology stack is suitable for the job.

We bring together call recordings, transcripts, tool logs, telephony data and CRM outcomes to build a clear picture of where and why calls fail.

That evidence directs the fixes. Poor transcription may require speech-model changes. Slow responses may point to tool latency. Failed bookings may reveal integration problems. We address the causes, not just the symptoms.

We then retest the failed scenarios and compare results against the original objectives, including task completion, reliability, response times and caller experience.

Includes

  • Objective-to-outcome assessment
  • Technology stack review
  • Consolidated call analysis
  • Prioritised failure report
  • Targeted fixes
  • Regression tests
  • Before-and-after performance evidence
CHALLENGE 03

"Our Voice AI handles a lot of calls, but it's costing more than we expected."

We examine where the money is going. Depending on the system, this may include telephony, speech recognition, language models, voice generation, tool calls and platform fees.

We trace unnecessary costs to their source, whether that's telephone billing increments, long conversations, redundant tool calls, repeated processing or expensive model choices.

We pursue all savings systematically, starting with those with the greatest impact. That could mean leaner prompts, more efficient tool calls, shorter workflows, better call routing, caching, different models or, where volumes justify it, moving to self-hosted infrastructure.

We measure the savings against the original costs, while protecting task completion and customer satisfaction.

Includes

  • Call-cost breakdown
  • Prompt and tool-call efficiency analysis
  • Cost-per-successful-outcome analysis
  • Identified savings opportunities
  • Architecture assessment
  • Prioritised recommendations
  • Projected operating costs
CHALLENGE 04

"We'd like to try Voice AI ourselves, but don't know where to start."

We provide reusable Voice AI recipes that give you a working starting point without having to build everything from scratch.

Start with a simple message-taking agent. Configure your business details, connect your voice service and choose where messages should be delivered.

Use the included demo to make test calls, check the information captured and confirm that messages arrive correctly. Once you're satisfied, you can connect a phone number and put the agent to work.

The recipe is yours to inspect, modify and deploy, with documentation to help you understand how it works.

Includes

  • Reusable Voice AI recipe
  • Configuration guide
  • Working demo
  • Message-delivery integration
  • Test scenarios
  • Deployment instructions
CHALLENGE 05

"We want to use Voice AI, but we're concerned about data security, compliance and control."

We start by understanding your data protection requirements and the information your calls handle.

We examine the complete data journey, including telephony, speech processing, language models, recordings, transcripts, integrations and third-party providers.

We assess the findings against applicable legislation and gazetted regulations, then prepare a strategic document with recommendations for processes, platforms, providers and data retention. This provides a technical foundation for review by your legal team.

Includes

  • Data-flow mapping
  • Regulatory references
  • Identified risks
  • Process and platform recommendations
  • Provider options
  • Data-retention recommendations
  • A strategy document for legal review

How we approach the work

From problem to proven production system

  1. Discover

    Understand the business problem

  2. Scope

    Define the workflow and success criteria

  3. Build Pilot

    Test the concept with a working solution

  4. Evaluate

    Measure results and validate feasibility

  5. Production Build

    Engineer for reliability, integration and scale

  6. Deploy & Improve

    Monitor performance and refine

Pilot and evaluation establish whether production is justified. Production Build is a separate commitment.

Work and evidence

Helping a UK Voice AI company get agents working for its clients

Since February 2025, I've worked under contract with a UK Voice AI company, helping take incomplete or failing Vapi agents through troubleshooting, testing and deployment.

My work goes beyond what the agent says during a call. I investigate why calls fail, improve the conversation, fix tool and integration problems, and make sure the information collected reaches the right people and systems afterwards.

I've also worked on adapting and improving these agents for UK businesses and their callers.

The client relationship is ongoing and confidential, so I can't share the company's identity or its client deployments.

When one misheard place name stopped a property search

During a test call, Vale, my estate agent voice assistant, heard "Barrydale" as "Barridale". The agent couldn't find the property the caller was asking about.

I traced the problem to speech recognition, adjusted the settings and tested the same enquiry again. This time, Vale recognised Barrydale and opened the correct property page.

Listen to both calls and see what changed.

The recording was correct, but the saved message was wrong

In another test, the caller's recording was clear, but the message saved by the agent was missing a word that changed its meaning.

I added a confirmation step so the agent would read the message back before saving it. In the next test, the caller confirmed the intended message and the saved record was correct.

Examine the call and the correction.

Try a conversation

Choose an agent to open its demo.

Miles Snr

Talk with Miles

Start a live browser call with Miles, Patrick's AI Voice assistant. Ask about missed calls, qualification, or booking, and hear how a practical Voice AI system handles the conversation.

Ready

Browser call

Ready for a browser call.

Vale

Property enquiries

Find a demo property, ask questions and request a viewing.

Try Vale

Leah

Private tours

Ask about Cape and Karoo tours, or leave a message for the guide. English or German.

Try Leah

Declan

Diner questions

Ask about the published menu, opening hours and visiting the diner.

Try Declan

Insights and research

Voice AI doesn't operate in isolation

A voice agent depends on phone networks, speech and language models, business software and the people responsible for what happens after a call. Decisions about data, infrastructure and costs shape how those parts work together.

I write about what I learn from building and testing Voice AI, and research the wider technical, commercial and regulatory questions involved in putting it into production.

Keep Voice AI Callbacks Short: YouTube video thumbnailLoad video

Moving a callback from Voice AI to a contact form

Miles confirms a caller's callback number, then opens a pre-filled contact form while the voice conversation stays connected.

The caller can type details that speech recognition may get wrong and make a separate, optional choice about marketing emails. My website code confirms whether the form opened, rather than leaving the voice agent to guess.

Watch the YouTube videoRead the technical article
Explore articles and research

Collaboration

A collaboration that changes what's possible

I'm proud to be collaborating with people who have built and operated Voice AI at enterprise scale.

For me, this is a significant professional milestone. In a field crowded with bold claims, having my expertise recognised by people with this depth of experience is both an honour and a source of genuine pride.

What excites me most is what we can achieve together.

My experience spans Voice AI engineering, systems integration, business processes and technical delivery. Combined with my collaborator's experience in high-volume infrastructure, production operations and cost engineering, this creates opportunities to take on projects of a very different scale and complexity.

The team, responsibilities and delivery approach are agreed around each project's requirements.

Questions before you start

Can this work with our existing phone and business systems?

You don't necessarily need to replace your existing systems.

Voice AI can connect through call forwarding or SIP routing, and exchange information with your CRM, calendar or other business software through supported integrations and APIs.

We'll establish which connections your existing systems support, and whether anything needs to change, before committing to the build.

What happens when a caller needs a person?

The caller can ask to speak to someone, or the agent can be instructed to transfer certain types of calls automatically.

Where your telephone system supports live transfers, the call can be routed to the right team, with the information already collected by the agent.

If nobody is available, the agent can take a message or callback request and send it to the appropriate person.

What happens if the Voice AI service goes down?

A production system needs a fallback that doesn't depend on the AI service being available.

Depending on the architecture, we can use backup speech and language-model providers, or configure telephone routing that sends new calls to your existing team, call queue or voicemail when the AI service is unavailable.

These failure scenarios are tested before deployment, rather than waiting for a real outage to discover what happens.

Where do our call recordings and customer details go?

That depends on how the system is built, and it's a decision you should have control over.

We can identify exactly which providers process call audio, transcripts and personal information, where that data is stored, who has access and when it is deleted.

For businesses with stricter security or regulatory requirements, we can explore running parts of the Voice AI system in your own cloud or a specified region, with controls for recording consent, sensitive information and data retention.

The resulting architecture and data-handling arrangements can then be reviewed by your legal and security teams.

Will we be locked into one platform or provider?

No. Provider independence should be considered from the start, whether we're building a new system or improving an existing one.

For a new build, we can design the architecture so telephony, speech recognition, language models and voice generation are separate components. Providers can then be changed as costs, performance or business requirements evolve, without rebuilding the entire system.

If you're already using a platform such as Vapi or Retell, we can work with what you have rather than forcing a migration.

Different providers have different capabilities, so switching may still require integration changes and testing. The aim is to keep those changes contained, with control over provider choices and infrastructure remaining with your business.

Have a Voice AI problem? Let's discuss it.

Book an initial consultation to talk it through, or send me a written brief if you prefer.

Book a discovery callSend a written brief