At twenty past eight on a Wednesday evening, a woman in Sheffield rings her insurer about a claim that has been open for three weeks. The line picks up on the second ring. The voice is calm and unhurried, and its first question suggests it already knows which claim she means. It is not a person. Although illustrative, enterprise voice agents are increasingly moving from pilots to production, and most customers only notice the difference when the conversation reaches something the software was never told how to handle.
That is roughly where the industry now sits, which is why OpenAI’s move into customer operations matters less as a product launch than as a change of subject. On 22 July 2026, the company introduced Presence, a platform for building, testing and running voice and chat agents that can look things up in company systems, take approved actions and hand a conversation to a human. It is not self-serve. Deployments run through OpenAI’s forward-deployed engineers or select global systems integrators, and at launch the company had not publicly disclosed pricing or contract terms. BBVA is exploring voice support for banking in Mexico, SoftBank is testing Japanese-language conversations, and the Australian insurer IAG is looking at high-demand periods. OpenAI says its own English-language phone line now resolves 75% of inbound issues without human assistance, a figure the company reports itself and nobody has independently audited.
That last figure is doing two jobs. It is a proof point, and it is also an admission: the claim implies OpenAI had to run a phone line, define what counted as resolved, watch where the handoffs clustered and rebuild the thing until the number was worth quoting. Nobody sold them that work.
What enterprise voice agents can actually do now
The capability jump is genuine, though it is worth naming precisely rather than generously. Presence supports real-time voice and chat, and an agent can retrieve relevant account information from connected company systems while the conversation is still running, then stop short of any action it has not been authorised to take. OpenAI says the platform includes policy guardrails, an approval step for consequential actions, simulation, evaluation tools and production signals, so that teams can test proposed changes before a controlled rollout.
What the market is buying, though, is forecasts rather than results. Gartner expects conversational AI to handle roughly 70% of enterprise customer support interactions by the end of 2027, and back in 2022 the same firm predicted that conversational AI would reduce contact-centre agent labour costs by $80 billion in 2026. Both remain projections, read out in board papers as though they were floors rather than ceilings, and the distance between a forecast and a working service is where the work lives.
The seven links a single call has to survive
Follow the Sheffield call end to end and the shape of the problem becomes obvious. Voice, identity, CRM lookup, qualification, action, escalation if that action is refused or contested, then an outcome someone can measure. Seven links. On any reasonable reading of what OpenAI describes, the platform supplies the first and part of the fifth. Everything else is a decision the business already made, badly or well, years before anyone mentioned AI.
Identity is a policy question, not a technical one: what proof is enough to discuss a claim, and what happens when a caller fails it twice. The CRM lookup only works if the record is where the agent expects it, which is often not true after a decade of merged fields, free-text notes and three competing definitions of “customer”. Qualification depends on commercial rules that often live in a team leader’s head, escalation needs a named human with capacity at 8.20pm, and “outcome” has to mean something firmer than a call that ended. None of these are exotic failure modes; they are the ordinary operational risks of any service line.
This is the work that Riophany helps businesses do: mapping which fields the agent may read, deciding what it may write back, agreeing with the revenue operations side what a resolved conversation looks like in the reporting layer. Companies that skip it do not end up with a broken agent. They end up with a fluent one that books the wrong appointment and logs it as a success.
Where enterprise voice agents run into UK law
British firms now have a second reason to be careful about the escalation step. Core reforms in the Data (Use and Access) Act 2025 took effect on 5 February 2026. For solely automated significant decisions using non-special-category personal data, the rules are more flexible and ordinary legitimate interests may be available as a lawful basis, while special-category data remains subject to tighter limits. The trade-off is a set of safeguards that are now mandatory rather than advisory: tell people an automated process was used, let them make representations, give them meaningful human intervention, and let them contest the outcome.
The operative word is meaningful. Meaningful intervention requires a person with genuine authority to reconsider and change the decision, and rubber-stamping is not enough. That single distinction turns an escalation queue from a cost line into a compliance control, and it is why the accountable human in any serious deployment of enterprise voice agents should be more visible on the org chart than the model behind it.
Not one accent, and not one market
The same architecture behaves differently depending on where the call lands. OpenAI’s own pilot list makes the point: Spanish-language banking in Mexico is not the same problem as Japanese-language support in Tokyo. Push further and the gaps widen. A Lagos service line has to cope with Nigerian Pidgin and mid-sentence code-switching, patterns documented well enough to sit in public language corpora, which means an agent needs explicit testing against Pidgin, local languages and code-switching rather than an assumption that English coverage will carry it. South Africa recognises twelve official languages, including South African Sign Language, and a bank that handles English and Afrikaans well and isiZulu poorly has not automated its service; it has automated a subset of its customers and quietly downgraded the rest.
Channel habits diverge just as sharply. In many markets WhatsApp is a major service channel rather than a secondary one, which changes what identity verification and escalation even mean. Treating all of that as one “African market” is how enterprise voice agents produce good containment figures and bad customers.
What the people on the other end actually think
The evidence on public appetite is unflattering and consistent. YouGov research for Pega, fielded among 2,222 UK consumers, found 68% were not very or not at all confident in how businesses deploy generative AI, half said AI-only interactions rarely or never produced a successful outcome, and 80% said they always or often got better results when dealing with a human. A separate UK survey found 83% would rather speak to a real person, though more than 30% said lower prices would make them more likely to use an organisation automating interactions with AI.
That resistance may reflect prior experiences of poorly designed automated service. It sits alongside a quieter finding: the UK Customer Satisfaction Index reached 78.2 in January 2026, its highest since July 2022, with a record 83.2% of customers saying their experience was right first time, which the Institute identifies as a key driver of the rise. The data do not establish that speed was irrelevant, only that first-time resolution mattered.
That is the standard the technology will be judged against. The argument now doing the rounds is that Presence and its competitors have made the conversational layer close to a commodity; if that holds, what separates deployments is whether the business behind one has decided, in writing, what a good ending looks like. The woman in the Sheffield scene does not care who built the voice. She cares whether the claim closes, and whether the person she is eventually put through to has the authority to close it.
What does Riophany do
Riophany helps businesses automate work without losing control of important decisions. We connect CRM, messaging, finance and operational systems, give the AI worker a defined job, and decide exactly what it can do itself and what must go to a person.
For voice calls, a natural AI voice is only the beginning. Riophany can design and connect AI support workflows to customer context, qualification, CRM logging, defined next actions and human escalation — while your team keeps control of important decisions.