What we know about voice AI support

Three things worth knowing

• Most of these projects fail, and not for technical reasons. Scope and governance are what kill them. Every one of those causes is a consulting problem, which is the whole argument for us being in the room.
• The vendor numbers are not the real numbers. Containment is sold at 80 to 90 per cent. Independent work puts it at 45 to 65. We plan with 50 to 55 and are more credible for it.
• The knowledge base is the bottleneck, not the platform. It usually does not exist, because it lives in the heads of the best agents. No vendor builds it for you.

Failure is the normal outcome

This is the finding that reframes every conversation. It is not a reason to stay out of the market. It is the reason clients need someone who has read the post mortems.

• Sinch reports that 74 per cent of companies had to roll back a live customer facing AI agent. Vendor source, so read it critically, but the direction matches everything else we have seen.
• Klarna and McDonald's are the two public cautionary tales, both walked back after going too wide too fast.
• In the UK, 83 to 85 per cent of customers say they would rather speak to a person, and frustration with AI agents rose from 54 to 59 per cent in a year. An agent that blocks the route to a human makes the original complaint worse.

Sources: Gartner (analyst), Sinch 2026 (vendor), 8x8 (vendor, trend corroborated elsewhere).

Three gates. No gate, no project

We worked these out so we can say no early, which is cheaper for everyone. All three have to be open.

Missing data is not a gate. If a client cannot name their top ten call reasons, that is a paid contact analysis first, then the agent.

Why the volume threshold, worked out ourselves

The running costs are never the problem. A human call costs about £6.26, the same call handled by AI about £0.35, a factor of fifteen. The problem is the fixed cost that does not scale: £35k to £150k to build, plus half to one full time owner forever, plus an annual minimum of around $150k on a managed platform.

Our own model, 9 August 2026. Note the AI cost applies to all contacts, including the ones that escalate to a human.

Numbers we can stand behind

Sorted by how much weight they will carry in a room. The dashed tags are the ones to quote carefully or not at all.

Two things to carry out of that table. First, the only peer reviewed effect in this whole field is for agent assist, meaning AI helping a human, not replacing one. For a nervous client that is the right first project. Second, headline per minute prices of $0.05 to $0.07 only cover orchestration. They do not include the model, speech to text, speech synthesis or the line.

There is no independent payback benchmark anywhere. That is the biggest hole in the market, and it is also our most honest line: we build the business case from the client's own numbers, not from an industry average.

Fifteen vendors, four layers

Infrastructure (models and speech), toolkit (build it yourself), managed (they run it), and the suite you may already own. The average score below is the least useful column. Cognigy leads and is still wrong for a small unregulated client. Always read the column that matches the constraint the client actually has.

• Only three vendors publish real prices. Salesforce Agentforce at $2 a conversation, Genesys at $0.06 a bot minute plus seats, Amazon Connect at $0.008 an AI minute. Everything else is a third party estimate and should be labelled as market observation, never as fact.
• Two names land on shortlists by mistake. Assembled and Verint are workforce management and quality management. They are not voice agent vendors.
• The existing suite matters more than the shortlist. If the client already runs Genesys, Amazon Connect or Salesforce, that constrains the choice faster than any feature comparison. It is discovery question number one.

What we get from talking to them

Desk research gives you the brochure. The calls give you the shape of the deal. We are working through the shortlist and writing down what each one says, so the comparison is ours rather than theirs.

Still to do: Sierra, ElevenLabs and Vapi, so the list is complete. Cognigy and Parloa came to us cold, which is worth remembering when we need a fast conversation.

Compliance is the product design

There is no separate AI law in the UK. The existing FCA rules apply to an AI conversation exactly as they apply to a human one, which is stricter than most people assume.

The DACH blocker nobody writes about in English

Section 87(1) no. 6 of the German Works Constitution Act bites on the abstract suitability of a system to monitor performance, not on whether you actually use it that way. The Federal Labour Court confirmed this on 8 March 2022. Call logging and QA scoring fall under it. Without works council agreement, the rollout is void, and there is a prohibition on using what you gathered. This belongs in the plan as its own phase, before the contract, not as a footnote.

This summary comes from web research, not from reading primary law. Have it checked before it goes to a client.

Three phases, and one of them is the business

Phase two is the whole argument. It is the part vendors deliver worst, because it is not a software problem. The knowledge usually does not exist in writing at all: it lives in the heads of the people who have been answering the phone for six years. Nobody sells a product that fixes that.

What we still do not know

• There is no independent payback benchmark in this market. Every ROI figure in circulation was paid for by a vendor.
• All the volume and cost benchmarks we have are American or British. We have nothing DACH specific, which matters because the German market is where most of our pipeline sits.
• The Checkatrade work is our strongest voice credential and it is still not written down anywhere. Jacob Haynes can close this, and he is the person who actually did it.
• Vendor pricing is third party estimate throughout, apart from the three that publish. We should say that out loud rather than imply precision we do not have.
• We have not yet tested what a vendor does when there are no historical transcripts, which is exactly the situation when a client is opening a phone channel rather than replacing one.