Voice AI dialler modes: predictive, progressive or preview?

Every enterprise that runs outbound calls inherits the same question from the human contact centre: should the dialler run in predictive, progressive or preview mode? For decades the answer was driven by one scarce, expensive resource, the live agent, and the whole point of a predictive dialler was to keep that agent busy. A voice AI estate changes the premise, because the agent is now a concurrent software session that scales on demand rather than a person you are paying by the hour.

The stakes are not academic. In its 2025 State of AI report, McKinsey found that 88% of enterprises now use AI and 33% have moved it into production, yet the outbound voice AI agents sitting on top of that stack are often wired to a dialler still tuned as though live agents were the bottleneck. Get the mode wrong and you do not just waste capacity, you manufacture the exact calls the regulator is watching for.

Ofcom is explicit about the mechanism. Its January 2026 refresher on silent and abandoned calls states that over-dialling with automated calling systems leads to agents not being available when the recipient answers, and to calls being dropped. This guide compares the three dialler modes, explains how UK regulation treats the calls they produce, and shows why an elastic AI estate should usually not dial predictively.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What are predictive, progressive and preview dialling?

Predictive, progressive and preview are three ways an outbound dialler decides when to place the next call. A predictive dialler places more calls than there are free agents, predicting that some will not answer. A progressive dialler places one call for each available agent. A preview dialler shows the agent the customer record first and lets them trigger the call. Dilr Voice defaults to one call per available session, the progressive pattern.

The three modes sit on a spectrum of over-dial ratio. Preview dialling places fewer than one call per agent, because the agent pauses to read the record. Progressive dialling holds a strict one-to-one ratio. Predictive dialling is the only mode that deliberately dials at a ratio greater than one, and that single design choice is the source of the compliance risk covered below. In a human contact centre the ratio is tuned to keep expensive staff talking, because the more idle seconds you remove, the better the economics.

In an AI voice estate the picture inverts. The agent is a concurrent session on a platform like Dilr Voice, wired to telephony through Twilio and to the record in Salesforce or HubSpot, and you can add sessions elastically. The scarce resource that predictive dialling was invented to protect has, for an all-AI campaign, largely disappeared. That is why the mode question deserves a fresh answer rather than a copied contact-centre default, and why it interacts with everything from your natural language understanding to your DTMF and IVR migration plan.

Why does predictive dialling create abandoned and silent calls?

Predictive dialling creates abandoned and silent calls because it deliberately over-dials, placing more calls than there are agents ready to talk. When more people answer than the algorithm expected, the surplus calls have no agent to connect to. Ofcom classes the result as an abandoned call if a recorded message plays, and as a silent call if the recipient hears nothing at all. Dilr Voice avoids this failure mode by not over-dialling in the first place.

An abandoned call, in Ofcom's terms, is one where a party sets out to reach a live recipient but a recorded message plays instead because no agent was free. A silent call is worse: the recipient answers and hears nothing, faint noise, or a line that disconnects. Answer machine detection makes both more likely, because when the detector wrongly decides a live person is a machine, that person is dropped into silence. This is why the refresher messaging ties over-dialling and automated calling systems so directly to consumer harm.

The abandonment a caller waiting in a queue experiences is a related but separate problem, and we treat the return-leg case in the callback and virtual queue guide rather than repeat it here.

How is the abandoned call rate measured, and does a threshold apply?

The abandoned call rate is the number of calls abandoned to a live individual as a proportion of all calls where a connection with a live recipient was established, measured per calendar day and per campaign. Ofcom's current statement of policy sets no specific abandoned call rate that is automatically safe. The three per cent figure came from its 2010 policy as a way to prioritise enforcement, and the statement records that treating it as a safe harbour is incorrect.

The point is stated plainly in the persistent misuse statement of policy that Ofcom still publishes today:

We cannot create safe harbours allowing a certain amount of that conduct or unlawfully fetter our discretion to take action.

Ofcom goes further and records that its revised statement does not include any specific abandoned call rate as a prioritisation criterion at all, whether three per cent or a different rate. When it does calculate the rate, the denominator is every call where a connection with a live recipient was established, which Ofcom counts as live calls plus abandoned calls; a campaign is defined as calls for a single proposition to a single target audience across a calendar day, and silent calls, unconnected calls and calls handled by answer machine detection are excluded. For a voice AI team, the practical lesson is that there is no number to hide behind: the safest abandoned call rate is zero, and the DNC and suppression logic that governs who you may call is a separate duty layered on top.

Does elastic AI concurrency remove the reason to over-dial?

Largely, yes. Predictive over-dialling exists to amortise the cost of idle human agents, but a voice AI estate scales concurrency on demand, so there is usually no idle-agent cost to spread across extra calls. That is an architectural observation about where the bottleneck now sits, not a promise of infinite capacity. Dilr Voice treats concurrency as elastic, which removes the main economic reason a human contact centre chooses to over-dial at all.

The concession matters, because AI concurrency is neither free nor unlimited. Model providers impose concurrency ceilings, telephony carriers rate-limit, and a badly planned burst can exhaust capacity just as a call centre runs out of seats, which is why model failover and provider redundancy and disciplined outbound scheduling windows still matter. The difference is that these are capacity-planning problems you solve by provisioning sessions, not by deliberately abandoning members of the public. The same operating discipline sits behind our AI operating model consulting, which treats compliant defaults as an engineering requirement rather than a policy afterthought.

Who is liable if a voice AI dialler makes silent or abandoned calls?

The organisation making the calls is liable. Under the Communications Act 2003, Ofcom can take action against any person who persistently misuses a network or service, and its January 2026 refresher confirms that call centres and the businesses they call on behalf of are all in scope, with penalties of up to two million pounds. Using a vendor such as Dilr Voice does not move that duty; the calling organisation stays responsible for how the campaign runs.

Ofcom is clear that a business which engages a third party to make its calls should still be judged on the arrangements it made to select that party and the compliance obligations it imposed on it. In other words, buying a dialler does not outsource the risk. This is distinct from the obligations that fall on your telephony carrier: communications providers face separate duties under General Condition B1 on the efficient use of numbers, where non-compliance can attract a penalty of up to 10% of turnover. That is a different regime binding a different party, and conflating the two is a common and expensive mistake. Where Dilr Voice operates the dialler for a client, the enterprise remains the responsible caller and our job is to make the compliant configuration the default, the same principle behind our AI execution office.

How should you configure answer machine detection and the information message?

If a call is abandoned, Ofcom expects an information message to play as quickly as possible and no later than two seconds after the recipient picks up, identifying who is calling and giving a number the recipient can ring to decline further calls, with no marketing content in it. Answer machine detection should be used sparingly, because a wrong detection is itself a silent call. Dilr Voice plays the information message inside the two-second window by default.

Three configuration rules follow from the policy. First, keep answer machine detection conservative or off, since false negatives, where a live person is mistaken for a machine, are one of the main sources of silent calls. Second, if a recipient has already received a call amounting to misuse, do not re-dial that number within the following 72 hours unless you can guarantee a live agent, because Ofcom treats a repeat inside that window as an aggravating factor. Third, measure the abandoned and silent call rates continuously, because you cannot manage what you do not instrument, which is why outbound campaigns belong under the same observability and tracing discipline as the rest of the estate, and why the output guardrails that shape what the agent says apply to the information message too.

Which dialler mode should an enterprise voice AI estate use?

For most enterprise voice AI outbound, progressive or one-to-one dialling is the right default: it places one call per available AI session and never over-dials, so it does not manufacture the abandoned and silent calls Ofcom treats as misuse. Preview dialling suits high-value or regulated calls that need a human check first. Predictive dialling earns its place only in blended estates where a human agent remains the scarce resource. Dilr Voice ships progressive by default.

The honest verdict depends on four criteria: regulatory exposure, the value of each call, whether a human is the endpoint, and whether human agents are still your bottleneck. Judged this way, an all-AI estate has almost no reason to reach for predictive dialling, and platforms across the market, including Vapi, Retell AI, Bland AI, Synthflow and PolyAI, are increasingly built around concurrent sessions rather than seat utilisation. The concession is genuine: in a blended operation that still routes answered calls to a bank of human agents, a carefully capped predictive mode can lift human utilisation, and a mature dialler can hold abandonment very low. For an estate where the AI is the endpoint, that concession simply does not apply, and choosing predictive means accepting regulatory risk to solve a scarcity problem you no longer have.

Does progressive dialling eliminate abandoned calls?

Progressive dialling removes the over-dial that manufactures most abandoned calls, because it never places more calls than there are sessions ready to take them. It does not guarantee zero silent calls, since latency, a failed transfer or a wrong answer machine decision can still leave a recipient in silence. Dilr Voice pairs progressive dialling with a fast information message so that any residual dropped call is handled the way the policy expects.

Is predictive dialling banned in the UK?

No. Ofcom considered proposals that respondents argued would effectively ban predictive dialling and decided not to prohibit the technology. What it declined to do was create a safe harbour, so there is no abandoned call rate that is automatically compliant. Predictive dialling remains lawful in the UK, but it is the mode most likely to generate the abandoned and silent calls that Ofcom prioritises for enforcement, which makes it a poor default for a voice AI estate.

Does the abandoned call rate apply per campaign or per call centre?

Ofcom calculates the abandoned call rate on either a per campaign or a per call centre basis, and may aggregate or disaggregate the data case by case. A campaign is defined as calls for a single proposition to a single target audience. This matters for a voice AI platform running many concurrent campaigns, because each is assessed on its own, so one misconfigured flow can attract scrutiny even when the estate average looks healthy.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read about our approach to placing AI inside enterprise systems.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. This piece sits in our voice AI series and builds on our enterprise voice AI agents guide. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.