Our own reference for every voice agent conversation: what a client must have before it is worth starting, which use cases actually work, who supplies what, what it really costs, and how we deliver it. Not a client document, the honest version.
How to use this. Run the readiness check in the first call. If it fails, say so · a client who is not ready produces a failed project with our name on it. If it passes, use the matchmaker to narrow the field, then the delivery model to scope the work.
Four of these are hard gates: without them there is no project, only an expensive disappointment. The rest are things we can help build. Tick what the client actually has.
The obvious objection, and a fair one: if a human call costs about £6.26 and an AI call about £0.35, the tool is roughly fifteen times cheaper per contact. So why would low volume ever be a problem? Because the saving scales with volume and the cost of owning the thing does not.
Four minutes at $0.10 a minute is about £0.35 a call against £6.26 for a person. Even at 50 % containment the unit economics are overwhelming. This part always works.
£35k to £150k to build, plus half a person to own it forever, plus a £150k annual floor if they go managed. None of that shrinks when the client is small.
Automating half the calls frees half the phone staff. If that is less than the person you must add to own the system, the client has swapped a support role for a technical one and paid for the privilege.
5,000 contacts a month is roughly three people on the phone at four minutes a call. Automate half and you free about one and a half. Add back half a person to own the knowledge base and the error queue and you net about one role, worth some £32,000 a year. But the agent itself costs about £21,000 a year at that volume, because you pay for every call it attempts, including the ones it hands over. Net: around £11,000 a year against a build of £35,000 or more. That is a three year payback, and nobody signs that off on cost grounds.
So the honest version is less convenient than the round number: 5,000 is the floor at which the arithmetic stops being negative, not the point at which it becomes attractive. With typical four minute calls the case gets comfortable somewhere around 10,000 to 15,000 contacts a month. Below that, either the calls have to be long (eight minutes at 5,000 contacts nets about £58,000 a year and works fine) or the reason to do it has to be something other than saving money.
Assumptions. One agent handles about 6,720 talk minutes a month (8 hour shifts, 20 days, 70 % occupancy). Loaded cost per UK contact centre FTE about £32,000 a year. AI cost £0.35 per contact attempted, charged on every call including the ones that escalate. Owner overhead 0.5 FTE up to 20,000 contacts, then 1 FTE. Change any of these in front of a client, they are ours, not gospel.
The honest caveat. The original 5,000 figure came from a vendor blog and we could not find an independent source for it. The reasoning above is ours, built from the UK cost per call, agent capacity and the ownership overhead we know these projects need. Use the logic in front of a client, not the round number.
Maturity below is our reading of the field, not a vendor promise. The pattern is consistent: the more the agent has to decide rather than retrieve, the worse it gets.
| Use case | Complexity | Maturity 2026 | Risk | Needs | Best fit |
|---|---|---|---|---|---|
| FAQ & status | Low | High | Low | Knowledge base, order/record system | Either, the standard entry point |
| Triage & routing | Low, med | High | Low, med | ACD/CCaaS routing, CRM | Either |
| Appointment booking | Low, med | High | Low | Calendar, CRM | Either |
| Authentication / ID | Low, med | High | Med, fraud | Auth system, optional voice biometrics | Either; biometrics is a separate build |
| Order / contract changes | Med | Medium | Med, high, money moves | CRM, billing, often several systems | Managed, integration depth |
| Outage / fault reporting | Med | Med, high | Med | Monitoring, ticketing | Either |
| First notice of loss (FNOL) | High | Medium | High | Core system, claims, document upload | Managed / sector specialist |
| Payments | Med, high | Medium | High, PCI | Payment gateway, billing | Managed with PCI scope |
| Cancellation / retention | High | Low, med | Med, high | CRM, billing, retention rules | Agent-assist, not automation |
| Use case | Complexity | Maturity 2026 | Risk | Best fit |
|---|---|---|---|---|
| Reminders (appointments, payments, delivery) | Low | High | Low, med, consent | Either |
| Appointment confirmation | Low | High | Low | Either |
| Surveys / CSAT | Low | High | Low | Either |
| Renewals | Med | Med, high | Med, contract law | Either |
| Lead qualification | Med | High | Med, cold calling law, strict in DE (UWG §7) | Outbound specialist |
| Collections | High | Medium | High | Managed, compliance-heavy |
Lower risk, faster payback, no customer ever hears it. The only rigorous study we have found (Brynjolfsson et al., published in the QJE) measured +14 % resolved cases per hour, and +34 % for inexperienced agents, and it was about assisting humans, not replacing them. For a nervous client, this is the right first project.
| Use case | Maturity | Note |
|---|---|---|
| Post-call summary / wrap-up | High | Often already included in their CCaaS. Check before selling it. |
| Knowledge retrieval during the call | High | Cheapest visible win, and it builds the knowledge base we need anyway |
| Real-time copilot / next best action | High | Cresta, Observe.AI territory |
| Automated QA (100 % scoring) | High for compliance, medium for tone | DACH: co-determination applies. Employee monitoring, see the gate below |
DACH blocker, easy to miss. Under §87(1)(6) BetrVG the works council must agree to any system that is abstractly capable of monitoring performance, actual monitoring is not required (BAG, 8 Mar 2022, 1 ABR 20/21). Call logging and QA scoring qualify. Without agreement the rollout is void and the data unusable. Plan it as a project phase before contract, not after. The EU AI Act (Art. 26(7)) adds a duty to inform worker representatives before high-risk AI is used at work.
The single most useful thing we can tell a client: these suppliers are not comparable to each other. Putting Deepgram next to Sierra on a shortlist is a category error, and most client shortlists have one.
Speech to text, text to speech, telephony. Building blocks, no conversation logic. They never compete with a platform, they sit underneath one.
You assemble the agent. Cheapest per minute, fastest to prototype, and you own compliance, evaluation and maintenance forever.
Enterprise compliance and support included, six figures a year, real lock-in. The right answer when the client has no engineering capacity to spare.
If they already run a contact centre platform, its own AI layer is often the cheapest path. Always check this before proposing anything else.
| Vendor | Level | Inbound | Outbound | Assist | Reach for it when |
|---|---|---|---|---|---|
| PolyAI | Managed | ✅ voice-first | ○ | ○ | UK client, voice quality matters, 40+ languages. UK company, easiest data residency answer |
| Parloa | Managed | ✅ voice-first | ✅ | ○ | EU data residency, telephony audio quality is the problem, insurance (Swiss Life) |
| Cognigy / NiCE | Managed | ✅ | ✅ | ✅ | On-premise or private cloud required. Banking, health, government. ERGO as insurance reference |
| Decagon | Managed | ✅ strong | ○ | ○ | Digital-first B2B SaaS support, fast time to value. No insurance track record found |
| Sierra | Managed | ✅ outcome-priced | ○ | ○ | Consumer brands with no internal AI capability. Pay per resolution appeals to CFOs |
| Regal | Managed | ✅ | ✅ strong | ○ | Revenue-adjacent calling: sales, collections, B2C at volume |
| Retell | Toolkit | build | ✅ batch | ○ | Outbound campaigns, fast and cheap, engineering team on hand |
| Bland | Toolkit | build | ✅ extreme volume | ○ | Very high outbound volume, own infrastructure, developer-led |
| Vapi | Toolkit | build | build | ○ | Prototyping and unregulated work. ⚠️ EU hosting still "planned", no public SOC 2, do not put in front of a regulated client |
| Cresta | Assist | ○ | ○ | ✅ strong | Sales-heavy contact centres, humans stay in the loop |
| Observe.AI | Assist | ○ | ○ | ✅ strong | 100 % QA scoring and compliance-heavy environments |
| ElevenLabs | Infra | voice layer only | Best-in-class voice under someone else's agent. EU residency since 2026 | ||
| Speechmatics | Infra | speech to text only | Accuracy, diarisation, on-premise. UK company, useful when residency is the blocker | ||
Two names that get onto shortlists by mistake: Assembled is workforce management (forecasting and scheduling), not conversational AI. Verint is workforce optimisation and speech analytics, not a voice agent vendor. Both are good at what they do and neither belongs in this comparison.
Sourcing caveat we should state out loud: almost every comparison online is written by a competitor. Parloa publishes "Sierra alternatives", Retell publishes "Bland vs Air AI". We use them for orientation, never as evidence.
Answer as the client would. The output is a starting position for the conversation, not a recommendation we would put in writing before seeing their data.
Every answer changes the shortlist. The reasoning is shown, so you can argue with it.
Sold as three separate pieces so a client can stop after any of them. In practice phase two is where we earn the reputation, and phase three is where the client keeps it.
Build or buy, which platform, which calls are in scope and which are explicitly not. Ends with one working agent on one or two narrow use cases.
Transcript mining, shadowing, extraction interviews, and reading what exists. The output is a versioned, owned knowledge base, not a wiki nobody updates.
Weekly sampling plus every conversation with a bad signal, an error taxonomy by severity, golden conversations as regression tests, and reporting the failures to the client.
Our starting position, to be calibrated against their volume and system complexity. Never quote before the readiness check.
| Phase | Model | Range | Basis |
|---|---|---|---|
| Scoping & implementation | Fixed sprint | 15,000-20,000 GBP | Two consultants, two to three weeks |
| Knowledge base | Per week | 8,000-12,000 GBP / week | Two to three people, scales with team size and system count |
| Continuous refinement | Monthly retainer | from 5,000 GBP / month | Roughly two consultant days a month, scales with volume |
Typical first engagement therefore lands around 35,000-55,000 GBP plus retainer. Derived from our existing rate card, not yet tested in the market.
Two things clients consistently get wrong: they believe the headline per-minute price, and they forget everything that is not the platform licence. Here is the whole bill.
Build it yourself and you pay for five layers. Prices below are current published rates, August 2026.
| Layer | Low | High | Plan with | Notes |
|---|---|---|---|---|
| Speech to text | $0.006 | $0.012 | $0.008 | Deepgram Nova-3 streaming $0.0077, AssemblyAI $0.0075, Speechmatics $0.007-0.012 |
| Language model | $0.001 | $0.020 | $0.006 | Cheap on mini models. The spread comes from context size and missing prompt caching, not token rates |
| Text to speech | $0.016 | $0.050 | $0.030 | Google/Azure neural at the bottom, ElevenLabs multilingual ~$0.10 at the top |
| Telephony | $0.007 | $0.025 | $0.015 | Telnyx ~$0.007, Twilio DE landline $0.015 / mobile $0.025. ⚠️ German-native trunks hit ~$0.13/min on mobile |
| Orchestration & infra | $0.005 | $0.015 | $0.010 | Self-hosted and amortised. LiveKit-class managed orchestration runs ~$0.077/min all in |
| Total per minute | $0.035 | $0.122 | ≈ $0.07 | Quote $0.05-0.10 in conversation. Dev platforms land at $0.11-0.30 all in |
| Vendor | Model | Typical | Published? |
|---|---|---|---|
| Salesforce Agentforce | Per conversation or flex credits | $2.00 / conversation, or ~$0.10 / action | Yes, official |
| Genesys Cloud | Bot minutes + seats | $0.06 / bot minute; seats $75-240 / user / month | Yes, official |
| Amazon Connect + Lex | Pay as you go | $0.008 / AI minute; Lex ~$0.05-0.06 / call | Yes, official |
| Cognigy / NiCE | Custom enterprise | ~$115k/yr average, $350k+ enterprise | Estimate only |
| PolyAI | Per minute, custom | from ~$150k/yr | Estimate only |
| Sierra | Per successful resolution | $1.00-2.50 / resolution; from ~$150k/yr, year one $200-350k | Estimate only |
| Parloa | Per successful conversation | from ~$300k/yr, sweet spot 500k+ calls/yr | Estimate only |
| Decagon | Per conversation / resolution | median contract ~$386k/yr (range $95k, 590k+) | Estimate only |
Ask the pricing model before asking the price. Per-minute, per-conversation and per-resolution scale completely differently. At high volume, outcome pricing can end up two to three times more expensive than per-minute, which is exactly when the vendor's incentive stops matching the client's.
Connecting CRM, telephony and the system of record is most of the work and can lift total cost by 20-50 %. The demo runs on a spreadsheet; production does not.
Six weeks for something basic, twenty or more for enterprise. Telephony integration and testing alone takes two to four weeks and cannot be compressed.
Plus $500-2,000 a month while actively iterating. This is the line item that decides whether build or buy is cheaper, and it is the one clients leave out.
Also routinely forgotten: concurrency limits on pay-as-you-go tiers (Vapi caps at 10 simultaneous calls, fatal for a campaign), compliance floors (LiveKit charges $500/month minimum once you need HIPAA or SOC 2 regardless of usage), redundancy across two speech vendors (+10-15 %), each additional language (+10-20 %), and regression testing after every model update.
Change the two numbers. Everything else follows the assumptions listed underneath, which are ours and are meant to be argued with in front of a client.
Assumptions. Build: $0.07/min variable ($0.055 above 50k calls with commitments), one-off $50k / $80k / $150k by size, ownership $18k / $70k / $230k per year. Toolkit: blended $0.15 / $0.12 / $0.10 per minute plus $5k / $20k / $40k setup. Managed: $0.25 / $0.15 / $0.10 per minute, enterprise discounts only arrive with volume commitments, against a $150k annual floor plus $50k setup. Derived from published component prices and third-party contract estimates; the managed figures are the softest, because no vendor in that tier publishes real prices.
There is no credible independent payback benchmark. Every ROI figure in circulation comes from a vendor or a vendor-commissioned study. Say that plainly and build the case from the client's own cost per contact instead, in the UK the defensible anchor is £6.26 per inbound call (ContactBabel), up 47 % in five years.
74 % of enterprises have had to roll back or shut down a live AI customer communications agent. Among those with fully mature guardrails, the figure is 81 %.
Sinch research, 2026 · Sinch sells communications infrastructure, so read the framing critically. The direction is corroborated elsewhere.Gartner puts it more precisely: over 40 % of agentic AI projects will be cancelled by the end of 2027, and the cause is not the technology. It is scope, governance and strategy, decisions made in the first fortnight.
Replaced around 700 support roles with an AI assistant in 2023 and automated roughly two thirds of enquiries. By May 2025 they were hiring people again. The CEO's own summary: they focused too much on cost, and quality suffered. Now a hybrid model, AI for routine, humans for anything that matters.
Two-year drive-through trial across 100+ restaurants, ended in June 2024 after order failures spread online, including bacon added to ice cream. The technology mostly worked; the failures were the only part anyone saw.
What this means for how we sell. Every one of these failure modes is a consulting problem, not a software problem. That is the argument for hiring us rather than buying a platform and hoping, and it is also why we should turn down clients who fail the readiness gates.