We have built AI voice agents for estate agents in Essex, legal firms in Canary Wharf, and dental practices across Greater London. In every single project, the make-or-break factor was never the technology - it was the script. At Softomate, we say it plainly to every client: poor scripts cause roughly 80% of AI voice agent failures. The AI can be flawless, the voice can sound indistinguishable from a human, and VAPI can route calls perfectly - but if the words are wrong, callers hang up and the project gets scrapped.
This guide is everything we have learned from writing UK-specific voice agent scripts across dozens of deployments. We cover Ofcom compliance, PECR rules for outbound calling, the three layers of script architecture, real objection handling, and the exact mistakes that sink most projects. If you are building a voice agent for a UK business - or hiring someone to do it for you - this is the playbook.
A voice agent script is not a phone script. It is a decision tree, a system prompt, a conversation flow, and a set of function calls - all working together. Writing it like a human call script is the first and most costly mistake we see, and it accounts for far more project failures than any technical issue.
When a human reads a script, they improvise around it. They pick up on tone, skip sections that are irrelevant, and recover naturally from unexpected tangents. An AI voice agent - even one powered by GPT-5.4 with dynamic objection handling - cannot do that without a carefully structured framework underneath.
The average UK full-time receptionist cost in London is £28,000-35,000/year in salary plus 25-30% employer on-costs (National Insurance, pension, holiday pay), totalling £35,000-45,500/year. AI receptionist setup costs range from £1,500-8,000 with monthly running costs of £100-400. The break-even point versus a human receptionist is typically 3-5 months. UK businesses using AI receptionists report 92% of calls answered within 2 seconds (versus 35% for human receptionists during peak hours). After-hours call capture rates improve by 40-65% when businesses deploy AI receptionists: UK trade businesses receive 28% of their enquiries between 6pm and 8am, 40% of which were previously missed. AI receptionist booking accuracy (correctly captured name, number, and appointment slot) averages 96% for clearly spoken calls. UK dental practices using AI receptionists report 18% reduction in DNA (did not attend) rates due to automated 24-hour SMS reminders. The ICO confirmed in 2024 that AI receptionist deployments are lawful under UK GDPR when callers are informed of automation at call start.
There are three distinct script layers in any AI voice agent deployment:
Most businesses we talk to have only ever thought about the middle layer. They hand us a script that looks like something a call centre manager wrote in 2019, and they are surprised when the voice agent sounds robotic and fails to hold a conversation.
The second major difference is latency. Voice AI on platforms like VAPI has a typical response latency of 600 to 900 milliseconds. That pause after the caller finishes speaking is real, and if your script is not written to account for it, callers interpret it as confusion or a bad line and they either repeat themselves or hang up. We write pause-aware scripts - short affirmation phrases like 'Of course' or 'Absolutely, let me check that for you' that fire in the latency window to hold the caller's confidence.
The third difference is interruption handling. Real callers do not wait politely for the AI to finish. They interrupt, they change their mind mid-sentence, they speak over background noise. Scripts need interrupt recovery paths - short, graceful recovery phrases that pull the conversation back on track without making the caller feel ignored.
Finally, a voice agent script must account for the unpredictability of real speech. Callers mumble, use slang, give partial answers, and sometimes go entirely off-topic. The system prompt layer needs explicit fallback logic: what happens when the AI cannot confidently classify the caller's intent? A good fallback is not silence or an error message - it is a clarifying question that sounds completely natural.
Ofcom's 2025 guidance on automated calling systems is explicit: any AI system making or handling calls must identify itself as automated within the first five seconds. This is not optional, and failing to comply exposes UK businesses to significant regulatory risk. We include the compliance declaration in every script we write, and we build it into the system prompt as a hard constraint the AI cannot override.
The Ofcom identification rule is the most widely known, but it is not the only requirement that affects script design. Here is what every UK AI voice agent script must include:
| Compliance requirement | What it means for the script | Where it goes |
|---|---|---|
| AI identification within 5 seconds (Ofcom 2025) | The opening line must state clearly the caller is speaking to an AI assistant, not a human | Opening turn, first sentence |
| Business name disclosure | The script must name the business the AI is acting on behalf of | Opening turn |
| Opt-out mechanism (PECR, outbound only) | Outbound scripts must offer the caller a way to opt out of future calls | Mid-call and close |
| Call recording notification | If the call is recorded for training or quality purposes, this must be disclosed | Opening turn or IVR before call connects |
| Purpose disclosure | The AI must state the reason for the call within the first exchange | Opening turn |
One thing we are frequently asked is whether the identification must use the word AI. Ofcom's guidance does not mandate specific wording - it requires that the automated nature of the call is clear and unambiguous. In practice, we use phrases like 'I am an AI assistant for [Business Name]' or 'You are speaking with an automated assistant from [Business Name].' Vague phrasing like 'This is a virtual representative' does not meet the standard and we never use it.
For inbound calls - where the caller rings the business and the AI answers - the same identification rules apply. Many businesses assume inbound AI is less regulated. It is not. The caller still has a right to know they are speaking to an automated system, and the business still has an obligation to disclose it clearly.
We also build a human handoff path into every script. This is not a regulatory requirement, but it is best practice and it is increasingly expected by callers. If a caller asks to speak to a human, the AI should acknowledge the request immediately and either transfer the call or offer a callback. Scripts that trap callers in an AI loop generate complaints and damage brand trust.
For more on call recording compliance and GDPR obligations that run alongside these scripts, see our guide on GDPR and PECR compliance for AI voice agents.
The opening of a voice agent script sets the entire tone of the interaction. Get it wrong and you lose the caller in the first 10 seconds. Get it right and you build enough trust in those first seconds to carry the whole conversation. UK callers respond to clarity, politeness, and brevity - not enthusiasm, not Americanised sales energy, and not robotic formality.
The formula we use for UK inbound voice agent openings is: identification, business name, purpose, and invitation. In that order, in under 20 words, with a natural cadence. Here is what that looks like in practice for an estate agent inbound line:
Estate Agent Inbound AI Voice Agent - Opening Script (Softomate build)
Hello, thank you for calling Hargreaves Property. I am an AI assistant and I am here to help with your enquiry. Are you calling about a property you have seen listed, a valuation, or something else?
[Caller responds]
Understood. I can help you with that. Could I take your name, please, so I have it to hand?
[Caller gives name]
Thank you, [Name]. Just to let you know, this call may be recorded for quality and training purposes. Now, tell me a bit more about what you are looking for - are you buying, renting, or looking to sell?
[Call branches to relevant flow based on response]
Notice several things about this opening. The AI identifies itself in the first sentence without making it feel like a legal disclaimer. The question at the end of the opening is open enough to capture multiple intent types but structured enough that the AI can reliably classify the response. The name capture is positioned naturally in the second exchange, not at the very start where it feels like a data-harvesting exercise.
The recording disclosure comes after the rapport has begun - not at the very start where it can feel alarming to some callers. We have tested both positions and the mid-opening placement consistently produces lower hang-up rates.
UK tone rules we apply to every opening:
On VAPI specifically, we configure the voice latency buffer with a short filler audio file that fires when the function call or LLM response takes longer than 800ms. This prevents the dead-air pause that makes callers think the line has dropped. ElevenLabs voices - particularly the UK-accented voices - are our default choice for UK business deployments because they handle British pronunciation of place names and industry terms far better than generic US-trained voices.
Objection handling is where most AI voice agent scripts collapse. The instinct is to write a rigid response for every possible objection - but that produces robotic, unconvincing exchanges that callers see through immediately. The correct approach is to define a framework of objection categories, write a natural response pattern for each category, and then let the LLM - GPT-5.4 in our current builds - generate the specific wording within those guardrails.
We still write the base script. GPT-5.4 is not writing the script; it is filling in the natural variation within a structure we control. This is a critical distinction. Letting the LLM improvise entirely produces inconsistent, occasionally off-brand, and sometimes non-compliant responses. The system prompt defines the rules; the conversation flow defines the structure; the LLM handles the natural language variation.
Here is how the most common UK business call objections should be handled:
| Objection | Poor script response (fails) | Good script response (converts) |
|---|---|---|
| I am not interested | I understand, but could I just tell you about our offer before you go? | That is absolutely fine. Is there anything that might make it worth a conversation at a different time, or would you prefer I remove you from our list? |
| I do not want to speak to a robot | I assure you I am very capable of helping you today. | That is completely understandable. I can arrange for one of our team to call you back - what time would suit you best? |
| How do I know you are legitimate? | We are a fully authorised company and you can check our website. | That is a fair question. I am calling on behalf of [Business Name], registered at [Address]. I can send you our details by text or email right now if that helps. |
| I am busy right now | This will only take two minutes of your time. | Of course - I can call you back at a better time. Would morning or afternoon work better for you this week? |
| Is this a sales call? | It is not really a sales call - I just wanted to... | It is a call about [specific purpose]. I am not here to pressure you into anything - I just wanted to make sure you had the information you need. Can I give you a quick summary? |
| I already have someone for that | We are probably better value than whoever you are using. | That is great to hear. If that ever changes, or if you want a second opinion, we are always here. Can I leave you our details? |
The column headings reveal the underlying principle: poor responses are defensive or pushy; good responses are validating and offer genuine choice. UK callers in particular respond negatively to any script that feels like it is trying to manoeuvre them past their objection rather than address it honestly.
In VAPI, we implement objection handling as intent classifications in the conversation flow with branching paths. The system prompt includes an explicit instruction: if the caller expresses resistance or reluctance, always acknowledge it before offering any alternative. This single system prompt instruction consistently improves conversion rates on objection paths in our A/B testing.
Outbound AI calling is one of the most tightly regulated areas of UK telecoms. The Privacy and Electronic Communications Regulations (PECR) govern who you can call and what you must include in the call. Getting this wrong is not just a compliance risk - the ICO has the power to issue substantial fines and enforcement notices. We take this seriously in every outbound script we write.
The core PECR requirements for outbound AI voice agent scripts are:
The opt-out mechanism in the script must be genuinely easy to use. Hiding it at the end of a long message, or making the caller press multiple options to opt out, does not meet the PECR standard. Here is an example of a compliant opt-out line that we include in all outbound scripts:
PECR-Compliant Opt-Out Line (Softomate standard outbound script)
Before we continue, I want to let you know that you can ask us to stop calling you at any point - just say 'remove me' or 'stop calling' and I will take you off our list immediately. You can also call us on [number] or email [address] to request this. Is it alright if I continue?
[If caller says 'remove me', 'stop', or equivalent at any point:]
Understood. I have marked your number as do not call. You will not receive any further calls from us. Have a good day.
[Call ends - CRM updated via function call in real time]
The function call at the end of the opt-out path is not cosmetic. We integrate VAPI with the client's CRM using VAPI's function calling capability so that the opt-out is written to the database the moment the call ends - not queued for manual processing. This is essential for PECR compliance because the regulation does not allow for a grace period between the opt-out request and the suppression being applied.
We also build a suppression list check into the pre-call logic. Before VAPI initiates any outbound call, a function call queries the CRM to verify the number is not on the suppression list and is not TPS-registered via a TPS API integration. Calls that fail this check are dropped before they connect.
For a detailed breakdown of the legal framework covering call recording alongside outbound calling, see our guide on GDPR and PECR compliance for AI voice agents in the UK. The ICO's official guidance on PECR is also essential reading for any business considering outbound AI calling: ICO Guide to PECR.
After building voice agents from £5,000 starter builds to £30,000 enterprise deployments, we have seen every script mistake there is. Most of them are avoidable with the right framework. Here is the full list of what we see regularly, and what to do instead:
| Common mistake | Why it fails | Correct approach |
|---|---|---|
| Copying a human call script verbatim | Human scripts rely on improvisation and emotional reading - AI cannot do this without structure | Write three separate layers: system prompt, conversation flow, function calls |
| No AI identification in the opening | Ofcom 2025 non-compliance; callers feel deceived when they realise they are talking to AI | First sentence must identify the system as AI, clearly and unambiguously |
| US-style sales language | UK callers respond negatively to high-pressure openers and superlative claims | Use measured, factual, polite language with genuine choice offered throughout |
| No interrupt handling | Callers who interrupt get ignored or trigger a restart, causing frustration and hang-ups | Build interrupt recovery paths into every major script turn |
| Ignoring latency | 600-900ms pauses feel like confusion or a dropped line; callers start repeating themselves | Add filler audio and short acknowledgement phrases to bridge latency gaps |
| One-size-fits-all objection responses | Rigid responses to objections feel scripted and untrustworthy | Write objection categories with LLM-guided natural language variation within guardrails |
| No fallback path for unclear speech | When the AI misclassifies intent, it goes off-track; without a fallback, the call derails | Every intent classification needs a fallback clarifying question |
| No opt-out for outbound calls | PECR breach; ICO enforcement risk; brand damage | Opt-out offer in the opening, repeated mid-call, with immediate CRM update via function call |
| Long, clause-heavy sentences | TTS voices struggle with complex syntax; callers cannot process long sentences aurally | Maximum 15 words per sentence in the core script; shorter in the opening |
| No human handoff path | Callers who want to speak to a human and cannot will complain | Any request for a human triggers an immediate offer of transfer or callback |
The mistake we see most often in scripts written by non-specialists is the absence of a clear intent classification structure. Businesses write a linear script and do not account for the fact that callers will not follow the expected path. A voice agent without solid intent classification is brittle. It works on the demo call and fails on the first real customer interaction.
On the technology side, we use Bland.ai for some high-volume outbound campaigns where cost per call matters more than voice quality, and VAPI with ElevenLabs for inbound and relationship-critical outbound work. The scripts differ between platforms because the function calling syntax is different and the interrupt handling works differently. A script written for VAPI will not port directly to Bland.ai without modification. This is something to clarify with any vendor before signing a contract.
If you are evaluating platforms and want to understand the full technical and commercial comparison, our AI development services page covers how we approach platform selection for UK clients.
You can use ChatGPT to draft a first version of your script framework, and we do use LLMs in our own scripting process. However, a ChatGPT-generated script will not include Ofcom compliance language, PECR opt-out mechanisms, VAPI-specific function call syntax, or interrupt recovery paths - all of which are essential for a production deployment. It will also default to US English conventions and sales language patterns that do not land well with UK callers. Use AI-generated drafts as a starting point, not a finished product. Every script we deploy goes through at least three rounds of manual review and live call testing before it is considered production-ready.
Scripts should be reviewed at minimum every 90 days. In practice, we recommend reviewing after every 500 calls using the call transcript data that VAPI and most other platforms provide. Look for: calls that go off-track at the same point repeatedly, objections that are not being handled effectively, and intent misclassifications where the AI asks a clarifying question it should not need to ask. Regulatory updates from Ofcom and the ICO should trigger an immediate compliance review regardless of the call volume cycle. UK telecoms regulation is moving fast in 2025-2026 and scripts that were compliant six months ago may not be today.
With GPT-5.4 as the underlying LLM, yes - to a controlled degree. The system prompt defines hard constraints: topics the AI must not discuss, information it must not share, actions it cannot take. Within those constraints, GPT-5.4 can generate natural language variation, handle unexpected tangents gracefully, and respond to novel objections it has not seen before. What it cannot do is reliably stay on-brand and compliant without a well-written system prompt. We describe this as guided improvisation: the AI has latitude to be natural, but not to go off-script on anything that matters. Full improvisation without guardrails is how you end up with a voice agent that makes promises your business cannot keep.
This depends entirely on the industry and brand. For legal, financial, and healthcare clients, we write formal scripts: full sentences, measured pace, no humour. For hospitality, retail, and consumer services, we write warmer scripts with contractions and a friendly register. The one constant across all UK scripts is that we never use Americanised sales enthusiasm - phrases like 'Absolutely!' as a generic affirmation or 'Awesome!' as a response to any statement. These patterns are immediately recognisable to UK callers as synthetic and erode trust faster than any technical issue. Match the formality level to the brand's existing marketing tone, then apply UK-specific language rules on top.
Testing has three phases. First, a table read: go through the script manually with a colleague playing the caller, following every branch including every objection path and every fallback. This catches logical gaps before you spend time on deployment. Second, a shadow test: run the deployed voice agent on a small number of internal test calls and review the full transcripts. VAPI provides call transcripts by default; use them. Look for intent misclassifications, unexpected fallbacks, and places where the AI's response does not match what the script specifies. Third, a live pilot: run the agent on a defined subset of real calls - typically 50 to 100 - before full deployment. Review transcripts daily during the pilot. In our builds, we do not consider a script production-ready until it has passed all three phases with a misclassification rate below 5% on the live pilot calls.
AI receptionist setup costs for UK businesses range from £1,500 for a basic phone answering system to £8,000 for a fully integrated solution with CRM, diary booking and WhatsApp. Monthly running costs are £100-£400 depending on call volume. This compares to a full-time human receptionist at £28,000-£35,000 per year including NI and pension.
Yes. AI receptionists integrate with Google Calendar, Outlook, Calendly, and industry-specific booking systems to book appointments directly during calls. The AI checks real-time availability, offers the caller 2-3 slots, confirms the booking and sends SMS and email confirmations automatically. No human involvement is required for the booking process.
More from the blog
Let us help
Talk to our London-based team about how we can build the AI software, automation, or bespoke development tailored to your needs.
Softomate Solutions Limited
Deen Dayal Yadav
Online
