If you are evaluating an AI voice agent for your business, the first decision is not which platform to choose - it is which direction the call travels. Inbound and outbound AI voice agents share the same underlying technology stack (VAPI, Bland.ai, ElevenLabs, Twilio, GPT-5.4 function calling) but they serve completely different business problems, carry different compliance obligations, and - critically for UK SMEs in 2026 - they deliver return on investment on very different timelines. This guide is based on our own project experience at Softomate, having built and deployed both types for clients across London and the wider UK.
An inbound AI voice agent answers calls that customers make to your business. An outbound AI voice agent makes calls to a list of contacts on your behalf. The direction of the call determines the trigger, the compliance obligations, the latency requirements, and - most importantly - where the ROI comes from.
Understanding this distinction early saves significant budget. We have spoken to SME owners who assumed a single agent could be 'switched' between inbound and outbound with minimal effort. In practice, while the same infrastructure (VAPI or Bland.ai connected to Twilio) can support both directions, the conversation architecture, the CRM integration logic, and the compliance framework are built differently from the ground up.
Here is a side-by-side comparison across the key decision factors:
| Feature | Inbound AI Voice Agent | Outbound AI Voice Agent |
|---|---|---|
| Primary use | Answer incoming calls, qualify, route, book | Make calls to a contact list, follow up, remind |
| Call trigger | Customer initiates the call | Business initiates the call |
| Call direction | Incoming to your number | Outgoing from your number |
| PECR applies | No (customer consented by calling) | Yes - for marketing calls; No - for service/reminder calls |
| Typical build cost | £5,000 to £18,000 | £6,000 to £25,000+ |
| ROI timeline | 30 to 60 days | 60 to 90 days |
| Complexity | Medium (routing, fallback, CRM write-back) | High (list management, compliance, retry logic, suppression) |
The latency profile also differs. Inbound calls can tolerate a response delay of 800 to 1,200ms without callers noticing - they are already waiting for a ring to be answered. Outbound calls are more sensitive: when your AI agent initiates the call and the contact picks up, any hesitation beyond 700ms reads as a robocall pause and leads to immediate hang-ups. We target under 700ms on all outbound builds, which influences platform selection and hosting region.
In the sections that follow, we break down the ROI case for each type, the real costs involved, the compliance picture, and which industries in the UK get the fastest payback from each direction.
Based on our project delivery at Softomate, inbound AI voice agents consistently deliver ROI faster for most UK SMEs. The reason is structural: inbound replaces an existing cost (a receptionist, an answering service, or missed calls), so the saving is immediate and measurable from day one. Outbound requires building a new pipeline before the return can begin.
This does not mean outbound is inferior - for the right business, the revenue uplift from outbound calling at scale dwarfs what inbound can produce. But the ramp-up time is longer, the compliance overhead is real, and the ROI calculation depends on conversion rate assumptions that take weeks to validate in practice.
Here are four real-scenario ROI calculations based on client builds we have completed or scoped. Company names are anonymised:
| Scenario | Before AI | After AI | Monthly saving / gain | ROI timeline |
|---|---|---|---|---|
| Inbound - East London dental clinic (3 chairs, 1 receptionist) | Receptionist at £28,000/yr + 40 missed calls/month (est. £80 each) | AI handles 85% of calls; 1 part-time receptionist at £14,000/yr | £1,500 labour + £2,400 recovered missed calls = £3,900/month | 35 days |
| Inbound - Essex property letting agency | Out-of-hours calls to voicemail; 22% callback rate; high tenant churn | AI answers 24/7, qualifies tenants, books viewings directly into Calendly | £1,200 answering service cancelled + 4 extra viewings/month at £900 avg LTV | 45 days |
| Outbound - South London vehicle finance broker (warm leads) | 1 sales rep calling 40 leads/day; 12% contact rate; 3% conversion | AI calls 300 leads/day; 34% contact rate; qualifies and books human callback | 4 additional funded deals/month at £350 commission each = £1,400 | 75 days |
| Outbound - Nationwide pest control (appointment reminders) | 18% no-show rate; admin calling manually; 2 hours/day labour | AI reminds 100% of bookings 24hrs prior; no-show drops to 6% | 12 recovered jobs/month at £85 avg = £1,020 + £1,200 admin saving | 60 days |
The inbound dental and letting agency scenarios reached positive ROI within 35 to 45 days because the cost displacement was immediate - from day one, the AI was doing work that previously cost real money. The outbound scenarios took longer because the AI needed 3 to 4 weeks of live calls to calibrate conversation flows before conversion rates stabilised.
Our honest recommendation: if you are a UK SME with a receptionist, an answering service, or a measurable missed-call problem, start with inbound. If you have a warm lead list larger than 500 contacts per month and a human sales team already converting, then outbound compounds fast once the pipeline is running.
An inbound AI voice agent built by Softomate starts at £5,000 for basic call deflection - answering FAQs, collecting name and enquiry type, then either ending the call or transferring to a human. A full-integration build with live CRM write-back, appointment booking, multi-department routing, and a custom ElevenLabs voice costs up to £18,000.
The cost range is wide because complexity compounds quickly. A basic inbound agent needs three things: a phone number (Twilio), a conversation engine (VAPI or Bland.ai), and a simple script. A production-grade inbound agent needs all of the following:
What does it replace? For most UK SMEs, an inbound AI voice agent eliminates or dramatically reduces one or more of these existing costs:
The total cost of ownership matters too. On VAPI, per-minute call costs run at approximately $0.05 to $0.09 per minute including the LLM and TTS components. A business handling 500 inbound calls per month at an average of 2 minutes each pays roughly £55 to £90 per month in usage costs on top of the build fee. That ongoing cost is usually 10% to 15% of what the answering service cost previously.
For clients who are nervous about AI replacing human warmth, we always recommend a hybrid model: the AI handles the first 60 to 90 seconds of every call (greeting, intent capture, basic qualification) and then warm-transfers to a human with a spoken handoff summary. This approach typically reduces receptionist call time by 40%, which translates to one fewer full-time hire for a growing practice. You can read more about how we structure these builds on our AI development service page.
An outbound AI voice agent starts at £6,000 for appointment reminder campaigns - a single-purpose agent that calls a list, reads a reminder, and logs the response. A full sales qualification build that integrates with your CRM, handles objections, books callbacks, and manages suppression lists costs £25,000 or more, depending on compliance requirements.
The ROI calculation for outbound is different from inbound. Instead of cost displacement, you are building a new revenue channel, and the ROI depends on three variables:
Worked example: an outbound agent making 300 calls per day to a warm finance lead list achieves a 34% contact rate (102 conversations), a 60% conversation rate (61 completions), and a 12% conversion rate (7 qualified leads per day). At 20 working days per month, that is 140 qualified leads passed to human closers. If the human closes 20% of those at an average deal value of £500 commission, monthly revenue is £14,000. With a £12,000 build cost, payback is achieved in the first full month of operation.
The honest caveat: outbound ROI projections are highly sensitive to list quality. We spend significant time in scoping calls understanding where the list came from, how recently it was generated, and whether the contacts have any prior relationship with the business. A stale or purchased list degrades contact rates to under 10% and often breaks the ROI case entirely.
Platform selection also matters more for outbound. We use Bland.ai for high-volume outbound builds because its infrastructure is optimised for concurrent outbound dialling with low per-call overhead. VAPI is excellent for inbound and lower-volume outbound but becomes expensive at scale. Twilio's dialler integration supports parallel calling, which means our outbound agents can run 20 to 50 simultaneous lines, compressing a 1,000-call campaign into under two hours.
Outbound AI calling in the UK is regulated by the Privacy and Electronic Communications Regulations (PECR), enforced by the ICO. Getting this wrong exposes your business to fines of up to £500,000, and we always build compliance into the architecture rather than treating it as an afterthought.
The key distinction is between marketing calls and service calls. PECR applies to automated marketing calls - calls where the purpose is to promote a product or service. It does not apply to service calls, such as appointment reminders, delivery notifications, or debt repayment reminders for accounts in arrears, provided the contact already has a relationship with you.
Here is what PECR requires for automated marketing calls specifically:
For more detail on exactly how PECR and GDPR interact in AI voice calling scenarios, read our dedicated guide: GDPR and PECR for AI voice agents in the UK.
On outbound builds, we add a compliance layer that costs between £500 and £1,500 depending on complexity. This includes: automated TPS screening via API before each dial, consent-flag validation against the CRM before the call is initiated, a compliant call opening script with mandatory disclosures, and a suppression list write-back when a contact says 'remove me'. For appointment reminder builds that are not marketing calls, this layer is simplified - TPS screening still applies for Ofcom compliance on nuisance call rules, but the consent requirement is reduced to 'legitimate interest' with a documented basis.
Inbound AI voice agents have essentially no PECR exposure. The customer is calling you. They have initiated the contact. There is no automated dialling. The primary compliance obligation for inbound is call recording disclosure (which we build into every opening script as standard) and standard GDPR data retention for transcripts.
Ofcom separately regulates silent or abandoned call rates for outbound dialling campaigns. Your outbound agent must not generate more than 3% abandoned calls per campaign (calls that connect but the agent fails to speak within 2 seconds). Our Bland.ai builds monitor abandoned call rates in real time and throttle dialling if the rate approaches the threshold. This protects your Ofcom standing and avoids nuisance call complaints.
The inbound-versus-outbound decision often resolves itself when you map it to the operational reality of your sector. Some industries have an obvious missed-call problem that makes inbound the natural starting point. Others have a large addressable contact pool and a conversion bottleneck that makes outbound the faster path to revenue.
Based on our project experience and the sectors we work with most frequently across London and the UK, here is our industry matching assessment:
| Industry | Better for inbound | Better for outbound | Why |
|---|---|---|---|
| Dental and healthcare practices | Strong | Moderate (reminders only) | High inbound call volume; missed calls = lost patients; reminders reduce no-shows without PECR risk |
| Property letting agencies | Very strong | Moderate | 24/7 tenant enquiry need; viewings booked instantly; outbound useful for expired listings re-engagement |
| Trades and home services (plumbers, electricians, roofers) | Very strong | Weak | Every missed call is a lost job; outbound rarely applies unless following up on quotes |
| Financial services and brokers | Moderate | Very strong | Large warm lead lists; qualification at scale transforms conversion economics; inbound useful for support |
| Legal (solicitors, conveyancers) | Strong | Moderate | High-value missed calls; outbound useful for case update calls (service, not marketing) |
| Vehicle and asset finance | Moderate | Very strong | Lead volume is the constraint; AI outbound qualification unlocks sales team capacity |
| Pest control and cleaning services | Strong | Strong (reminders and re-engagement) | Inbound handles new enquiries; outbound reminder campaigns recover no-shows and drive repeat bookings |
| Recruitment agencies | Moderate | Very strong | Screening large candidate pools is exactly what outbound AI does efficiently; inbound handles inbound applicants |
| E-commerce and retail | Weak | Moderate (abandoned cart, delivery) | Most customer contact is digital; outbound works for abandoned cart voice nudges where consent exists |
| Accountancy and bookkeeping | Strong | Weak | Client calls are high-trust; inbound qualifies new enquiries; outbound rarely justified |
A pattern emerges: industries where the bottleneck is answering existing demand (trades, healthcare, property, legal) benefit most from inbound AI first. Industries where the bottleneck is reaching enough of a large addressable market (finance, recruitment, debt recovery) benefit most from outbound AI first.
The most common build we deliver is a hybrid for a service business that starts with inbound (answering calls, booking appointments) and adds outbound 6 to 12 months later once the inbound system is proven and the team is comfortable with AI-handled calls. This sequencing reduces risk and builds internal confidence before tackling the more complex outbound compliance and campaign management requirements.
If you want to compare the specific platforms we use for these builds - VAPI, Bland.ai, and Retell AI - read our detailed breakdown: Bland.ai vs VAPI vs Retell AI: which platform for UK businesses in 2026?
Technically yes - VAPI and Bland.ai both support inbound and outbound call handling on the same account and the same phone number. However, in practice we build them as separate conversation flows because the logic is fundamentally different. An inbound flow must handle any intent a caller raises. An outbound flow is purpose-built for one objective (reminder, qualification, survey) and the conversation tree is much tighter. Running both through the same flow creates a bloated, slow agent that does neither task well. We typically provision two separate agents sharing the same CRM integration but with distinct conversation designs. The cost saving of combining them is small; the performance cost is significant.
With Bland.ai and a properly configured Twilio account, we typically run outbound campaigns at 20 to 50 simultaneous concurrent calls. At 50 concurrent lines with an average call duration of 90 seconds, you can complete approximately 2,000 call attempts per hour. Not all of those result in conversations - contact rates of 28% to 38% on warm lists mean roughly 560 to 760 actual conversations per hour. For most UK SME campaigns, we run at 20 to 30 concurrent lines to stay well within Ofcom's abandoned call rate limits and maintain call quality. A campaign to 1,000 warm leads typically completes within 2 to 3 hours at this dial rate.
For most UK SMEs, no - and we never recommend positioning it that way to your team. The AI handles the high-volume, repetitive part of receptionist work: answering standard questions, taking caller details, booking appointments, and routing calls. A human receptionist adds value in complex situations - a distressed caller, a complaint that needs empathy, a relationship-critical client who expects to speak to someone by name. The realistic outcome is that an inbound AI agent reduces receptionist call time by 40% to 60%, which either frees your existing receptionist for higher-value tasks or allows you to hire a smaller team as you grow. Full replacement works only for businesses whose calls are almost entirely transactional (e.g., a high-volume appointment-only clinic).
Every inbound AI agent we build includes a graceful escalation path. When the agent encounters an intent it cannot handle - a question outside its training, a caller who explicitly asks for a human, or a situation where the sentiment analysis flags distress - it triggers a warm transfer. The agent says something like: 'Let me connect you with one of our team who can help with that - I will let them know you have been asking about X.' It then transfers the call to the designated human line and sends a real-time webhook to your CRM or a Slack channel with the caller's name, number, intent summary, and transcript. If no human is available (out of hours), it takes a message and books a callback slot. The caller never hits a dead end.
Yes - this is one of the strongest capabilities of modern outbound AI, and one of the most underused. Using GPT-5.4 function calling connected to your CRM, the outbound agent can open the call with a reference to the contact's specific situation: 'Hi Sarah, I am calling from Apex Finance about your enquiry on 12 May regarding a £15,000 car finance agreement.' This personalisation increases conversation rate significantly - in our A/B tests across client campaigns, personalised openers improve conversation completion rates by 18% to 24% compared to generic openers. The agent can also vary the script based on CRM data: different tracks for hot leads versus cold, for previous customers versus new prospects, for contacts who previously asked for a callback versus first-touch calls.
Outbound is significantly more complex from a GDPR and PECR perspective. For inbound, the primary obligations are: disclose that the call may be recorded at the opening (standard), ensure transcripts are stored securely and deleted per your retention policy, and respect any SAR or erasure request. For outbound marketing calls, you need explicit prior consent for each contact, a TPS screening process before every dial, a suppression list with real-time write-back, a documented lawful basis for processing, and a record of consent with timestamp and source. For outbound service calls (reminders, case updates), the consent bar is lower but legitimate interest must be documented and a balancing test completed. We add a compliance module to every outbound build that handles TPS screening, consent validation, and suppression list management automatically - this adds £500 to £1,500 to the build cost but is non-negotiable.
AI receptionists handle unlimited simultaneous calls - there is no queue, no hold music, and no missed calls during peak periods. A UK trade business receiving 40 calls simultaneously on a Monday morning (post-weekend enquiry surge) will have all 40 answered within 2 seconds by the AI. Human receptionists typically handle 1-2 simultaneous calls. For UK businesses where peak call volumes cause missed enquiries (trades, healthcare, estate agents, hospitality), the simultaneous handling capability alone justifies AI receptionist investment: each missed call in a UK trade business represents £150-500 average job value.
Written by the Softomate Solutions AI Development Team, Barking, East London. Sources: ICO Guide to PECR - Automated Calls; Ofcom; VAPI Documentation.
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
