Decagon Alternative UK: What to Test on the Phone Line

Decagon calls itself "the AI concierge for every customer", and in 2026 it has built a visible UK presence. It opened a London office in November 2025, said in June 2026 that the London team had passed 50 people, and is holding its first European Dialogues event in London on 12 November 2026. On 1 October 2026 it launched Voice 3, a rebuilt voice agent that speaks through Chord, a speech model Decagon trained itself.

This guide is written for the Head of Customer Operations who owns the phone line and has been asked whether Decagon should take it. That person is measured on calls resolved, transfers handled well, complaints avoided and an audit trail that survives a regulator's question, not on how good a demo sounded. So the guide stays on the phone line: how Voice 3 behaves on a live call, how a Decagon agent reaches your numbers, how it checks who is calling, who changes it after launch, and what to get in writing on data and calling rules, the same ground our voice AI vendor checklist covers in general form.

It deliberately leaves three things to other pages. The broader question of whether one multi-channel agent should own the phone at all is covered in our Sierra alternative guide, which argues that case in full. Decagon's funding rounds and what vendor valuations mean for a buyer sit in our voice AI valuation signals guide. Price comparisons live in our best AI voice agent guide. Nothing here states a price for any vendor.

This guide is shipped by the team behind Dilr Voice, a multi-agent voice AI platform that chains specialised agents into a single phone call. Or see DATS, our five-stage AI consulting system for placing AI inside enterprise operations.

What is Decagon, and what is its UK presence in 2026?

Decagon is a United States AI customer-service company that sells AI agents for chat, voice and email from one platform, built through natural-language workflows it calls Agent Operating Procedures. Its UK presence is recent: Decagon opened a London office in November 2025, names European customers such as Deutsche Telekom and Rituals, and publishes a UK voice case study with Checkatrade.

The Decagon homepage groups the product into channels (voice, chat and email), build tools (Agent Operating Procedures, usually shortened to AOPs, and integrations), optimisation tools (experiments and testing) and scale tools (insights, an always-on quality reviewer called Watchtower, and knowledge suggestions). Duet is described as an AI partner for building and improving agents. The pitch is that one intelligence layer sits behind every channel, so an agent built once can answer a chat, an email or a call.

The UK footprint is recent and real. Decagon announced its London office in a post dated 4 November 2025, and a June 2026 update said the London team had reached 50+ people and named Deutsche Telekom, Hunter Douglas, Rituals and Oura among regional customers, with Hunter Douglas running localised agents in the UK, the US and Australia. The same post quotes Decagon's VP of EMEA, Sam Schofield, saying EMEA buyers want visibility and control and are "not looking for black-box automation". That is a fair summary of what a UK operations lead should demand from any vendor, Decagon included, and the rest of this guide turns it into questions. If you are new to the category, our enterprise voice AI agents guide sets out the moving parts first.

The most useful UK evidence is the Checkatrade case study, because voice is Checkatrade's primary customer contact channel and the study lists Voice as the Decagon product used. We return to it below, because what it says about ownership and testing matters more than its headline numbers.

What does Decagon Voice 3 change on a live call?

Decagon Voice 3 replaces a step-by-step speech pipeline with what Decagon calls a duplex architecture: a fast conversational model listens and speaks while a more powerful model reasons, calls tools and enforces guardrails in parallel. Decagon says this lets the agent keep talking during a long lookup, take follow-up questions mid-task and switch among 70+ languages.

The Voice 3 launch post, dated 1 October 2026, explains the problem Decagon set out to fix. Most voice agents run a cascade of speech-to-text, a language model and text-to-speech, and each stage adds delay, so a short "mhm" can be heard as an interruption and a slow lookup leaves the caller in silence. Voice 3 splits the job into two layers running at the same time. The Decagon voice page adds interruption handling, a choice of hundreds of voices, escalation to a human agent with a concise summary, outbound calling and guardrails that enforce your AOPs and policies.

Chord is the other half of the launch. Decagon says it is the first voice model trained by its own Decagon Labs for customer conversations, shaping speech phrase by phrase so it slows down for a confirmation code or a phone number. The post says Chord is trained on licensed data and consented voice talent, never on customer-owned data. In Decagon's own blind test across three pairs of recordings, the post says that on average approximately 90% of listeners could not pick the human. That is a vendor test with three voices, not an independent study, and it is a reason to listen for yourself rather than a measurement of your calls.

None of this tells you how the agent handles your callers. A duplex design should reduce dead air and clumsy turn-taking, and the way to find out is to replay your own recordings: callers who talk over the agent, callers on a building site, callers spelling a surname or reading a long reference number, scored against the test set our voice AI accuracy guide describes. Test, too, the languages your callers actually speak rather than the count on a product page.

How does a Decagon agent reach your phone numbers?

A Decagon agent reaches your phone numbers in one of three ways named on its integrations page: through a contact-centre or communications platform such as Amazon Connect or RingCentral, through SIP trunking from your existing telephony, or by forwarding inbound calls to the agent. The route decides who owns the number, the recordings and the transfer path.

The Decagon integrations page describes CPaaS platform integrations "like Amazon Connect or RingCentral" for voice calls with handoffs, SIP trunking to integrate with existing telephony infrastructure, and call forwarding with transfers to human agents when escalation is required. It also lists help desk and CRM connections such as Salesforce, Intercom and Zendesk, knowledge syncs from Confluence, Contentful and Kustomer, Model Context Protocol connectivity and API integrations. That is a broad menu, and the choice among the three call routes is the first architectural decision on the project, not a detail to settle after signature.

Each route moves a different piece of the call away from you. Work through the table below with your telephony lead before the first demo, because the answers change the contract, the recording policy and who you ring when a call drops. If you want a second pair of eyes on the route, talk to our team.

The transfer path deserves its own test. Decagon says escalations pass a concise summary to the human agent, and its Decagon Assist product gives human agents summaries, suggested replies and live transcription. Check that the summary lands in the desktop your team actually uses, at the moment the call connects, using the checks in our warm transfer guide. If the summary arrives after the agent has said hello, the caller repeats themselves and the benefit is gone.

How should a Decagon voice agent check who is calling?

A Decagon voice agent can work with callers who arrive already verified, through an app or web session, a callback after a one-time code or identity data passed with the call, or it can verify them mid-call with app prompts, several identifiers or one-time passwords. Decagon itself warns that caller ID can be spoofed, so a phone number alone is not proof of identity.

Decagon's post on voice authentication, dated 4 December 2025, is one of the more useful vendor documents on the subject. It names three reasons voice is harder than chat: long alphanumeric identifiers are hard to transcribe, callers are often driving or cooking and are less willing to open an app, and phone numbers are unreliable identifiers. It then separates pre-authenticated calls from mid-conversation checks and ranks the mid-call methods by identification confidence, from app push notifications at the top to a single basic identifier at the bottom, and rates the friction of each.

The practical lesson is to decide what the agent may do at each level of confidence before you decide how it verifies. A caller who has given only a name might hear opening hours and order status; a caller who has passed a two-part check might change an appointment; only a caller verified through an app or a one-time code should reach account changes or payments. Write that ladder down, give it to every vendor on the shortlist, and test it with calls that fail each step. The Decagon security page lists voice authentication alongside short-lived API tokens scoped for minimal privilege and discarded after each session, which is the right shape, but the ladder is yours to define. The UK-specific side of this, including how Ofcom's caller line identification rules work, is covered in our caller ID authentication guide.

Who owns a Decagon agent once it is live?

On Decagon, ownership can sit with the operations team, because agents are built and changed through Agent Operating Procedures written in natural language rather than code. Checkatrade, in Decagon's case study, placed ownership inside customer operations, trained every builder through Decagon University and validated performance in quality assurance before gradually increasing call traffic.

This is the most transferable part of the Checkatrade story. The study says Checkatrade launched an inbound voice agent called AI Charlie, made resolution its primary measure, and tested resolution, routing and identity verification before raising traffic. Decagon's case study reports that Charlie handles 65% of Checkatrade's inbound calls each week, and that direct resolution rose from 5% to 12% as the team moved from knowledge answers to targeted AOPs. Those are vendor-published figures about one customer, and they measure different things, handling and direct resolution, without a shared denominator in the study, so read them as separate signals rather than a funnel. The instructive part is that the study ties the resolution gain to the move from knowledge answers to targeted procedures. Measuring that kind of shift needs a baseline taken before launch, as our pre-pilot baseline guide explains.

That points to the real question for your organisation. If operations staff can change agent behaviour in plain language, someone has to decide which changes ship, what gets tested before they do, and who signs off a change that touches identity checks or complaints. Decagon offers testing, live experiments and Watchtower for review, and its newer Agent Workbench is described as an assistant that explains agent decisions from logs and suggests fixes. Those tools help, but they do not assign accountability. The same diagnostic logic underpins our AI operating model work, which sets the roles, sign-offs and change control around an agent before it takes live calls.

Ask Decagon to show you a change being made, tested and released, including how a regression would be caught before it reached callers. Then ask who in your team would do each step. The same test set should sit behind every release, whichever platform you choose.

What should a UK buyer confirm about Decagon's data terms?

A UK buyer should confirm Decagon's data terms in the contract, because Decagon's website privacy policy says it does not apply to information processed on behalf of business customers. Ask for the data processing terms, the sub-processor list, where recordings, transcripts and model calls are processed for a UK deployment, and current reports behind the security badges.

Start with scope. The Decagon privacy policy, effective 6 May 2025, says Decagon is headquartered in the United States and may transfer the personal information it collects to the United States or elsewhere, and it also says the policy does not apply to information processed on behalf of business customers, which is governed by those customers' agreements. Your callers' data falls in that second category, so the website tells you nothing about it. If any of it leaves the UK, your team needs to work through the ICO's brief guide to international transfers, published on 15 January 2026, which includes checklists for identifying a restricted transfer. The questions match those in our voice AI security questionnaire guide.

Next, read the security page precisely. It describes encryption with AES-256 at rest and TLS 1.2 or higher in transit, single sign-on with identity providers such as Okta and Microsoft Entra, zero-data retention with AI providers such as OpenAI and Anthropic so conversation data is not stored or used for training, and automatic redaction of personal data using Google's DLP service just after a conversation ends. It shows a row of badges, including SOC 2 Type II, GDPR, CCPA, HIPAA and the EU AI Act, plus ISO and PCI entries, without saying which are certifications held and which are frameworks addressed; the EU AI Act is a law, not a certificate. Request the current reports through Decagon's Trust Center rather than relying on the badges. Ask, too, what happens to personal data during the call and in recordings, since the redaction described runs after the conversation ends.

Finally, ask about memory. Decagon's Spring 2026 Proactive Agents release describes User Memory, which carries customer context across sessions with granular control over what is stored and how it is used. Context that follows a caller from chat to phone is useful, and it is also personal data held in a second place alongside your CRM. Agree what is stored, for how long, and how a caller's erasure request reaches it. Our guide to voice AI memory across calls sets out the design choices in detail.

Which calling rules apply if Decagon makes outbound calls?

If a Decagon agent makes outbound calls for a UK business, the calling rules bind that business, not Decagon. Under PECR, the ICO says you generally must not make marketing calls to numbers on the Telephone Preference Service or Corporate TPS without specific consent, and Ofcom's persistent misuse policy sets expectations for abandoned and silent calls. Telling callers they are speaking to an AI also needs agreeing with the vendor.

Outbound is a growing part of Decagon's offer. The Proactive Agents release describes Outbound Voice, with AOP-driven logic and call handling that navigates callbacks, voicemails and robocall screeners, and the Checkatrade study says the company expanded into outbound voice six weeks after launching its inbound agent, with three outbound agents live. The ICO's telephone marketing guidance is clear that call lists need screening against the TPS and CTPS for marketing calls, and that duty sits with the organisation making the calls. Calls such as helping a tradesperson finish an ID check may fall outside marketing, so classify each campaign before it starts. Ofcom's persistent misuse statement lists the steps it takes into account on abandoned calls, including an abandoned-call rate not exceeding three per cent of live calls per campaign over 24 hours and presenting a caller line identity, and those steps fall to the caller too. Our UK outbound calling guide covers the full set.

Disclosure is the other question, and Chord makes it sharper. Where the EU AI Act applies, for example because calls reach people in the EU, Article 50(1) of the EU AI Act states that "Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect." The text is on the Article 50 page, and the duty is placed on providers, which for a bought platform usually means the vendor rather than you, so ask Decagon how its agents disclose and check the opening script yourself. In our view, a voice that Decagon's own test says most listeners could not tell from a human is exactly the case where the "obvious" exception cannot be relied on, so the disclosure belongs in the opening of every call. Our Article 50 disclosure guide covers how to record that it happened.

When is Decagon the right choice for a UK contact centre?

Decagon is the right choice when you want one platform and one set of AOPs behind chat, email and voice, when callers speak many languages, when an operations team is ready to own and change agents itself, and when a speech model tuned for customer calls matters. It is a weaker fit when you want to test on your own calls before a sales conversation, or when the phone line is the only job you need covered.

Be fair to the platform. Decagon covers ground a call-only product does not: chat and email from the same procedures, 70+ languages with automatic switching, its own speech model, outbound voice, agent assist for human staff, live experiments and an always-on quality reviewer. A consumer brand whose customers move between chat and phone, and which already runs its help desk on Salesforce or Zendesk, gains most from that breadth, the case our omnichannel voice AI strategy guide examines. Decagon's Checkatrade and Hunter Douglas stories describe deployments serving UK customers.

The weaker fit is narrower. Decagon's pages route buyers to "Get a demo", and we found no self-serve trial on the pages we read, so expect to talk to sales before testing it on your own calls. And if your brief is a high-volume UK outbound programme with calling-window and screening controls built in from day one, compare that requirement line by line, because those controls are yours to configure on any platform; our inbound vs outbound guide explains why the two jobs differ.

How does Dilr Voice differ from Decagon for UK phone work?

Dilr Voice is an enterprise voice AI platform from DILR.AI built for phone calls. It chains specialised agents into one call, ships per-country calling rules such as recording consent, DNC checks and permitted hours by default, and runs telephony on Twilio or Vobiz. Decagon covers chat, email and more languages; Dilr Voice is narrower and call-specific.

The difference is scope, and the table keeps to what each vendor publishes. The Decagon column comes from the Decagon pages linked above; the Dilr Voice column comes from DILR's Dilr Voice product page and FAQ.

Inside a call, each Dilr Voice agent passes context to the next on handover, so a greeter can hand to a qualifier and then to a customer-care or action agent; our enterprise voice AI agents guide sets out how agents like these fit a contact centre. Inbound work covers business-hours configuration, after-hours routing and warm transfer to a human with full context, and responses arrive in under 500 milliseconds on DILR's own on-platform measurement. Outbound campaigns run from an uploaded contact list with scheduling windows, retry logic and auto-pause at the configured daily end time. After the call, CRM records, email, SMS and Slack notifications and lead updates run without code, with HubSpot, Salesforce and Google Calendar integrations, and knowledge bases built from your own documents are searched during live calls.

On data, Dilr Voice is hosted on Google Cloud Platform, encrypted at rest and in transit, with dedicated tenancy and regional data residency options for enterprise customers. The honest limit is scope: the product page lists no chat or email channel, and Decagon's language count is larger. If those matter as much as the phone line, Decagon covers ground Dilr Voice does not. If the phone line is the job, you can start Dilr Voice on a free trial with no credit card required, and our DATS team can help design a like-for-like test.

What is the best Decagon alternative for UK teams in 2026?

The best Decagon alternative for a UK team depends on what Decagon was shortlisted to do. If the brief is one agent across chat, email and phone, compare it with other multi-channel suites such as Sierra or Parloa. If the brief is the phone line, compare platforms built for calls, such as Dilr Voice or PolyAI. Test every option on your own recorded calls.

There is no single winner, and a vendor that claims to be one has skipped the criteria. Sort the list by the job first. For multi-channel suites, our Sierra guide is the closest comparison, and Parloa belongs on the same list. For call platforms, our PolyAI alternative guide covers another option.

Then run the same four steps for every vendor, Decagon included, in the order shown below. Decide which team owns the phone line and signs off changes. Test the agent on your own recordings, with interruptions, accents and long identifiers. Place each duty, from TPS screening to AI disclosure to transfer data terms, with a named party in writing. Then pilot on a small share of real calls with a pass mark agreed in advance, as set out in our pilot-to-scale guide. A vendor that clears all four is worth the next stage. If you want help writing the test, talk to our team.

Frequently asked questions

Can Decagon voice be bought without its chat product?

Decagon's Checkatrade case study lists Voice as the only Decagon product the company used, which indicates a customer running Decagon on the phone line. Decagon's own pitch is a single platform across channels, so ask Decagon directly how a voice-only deployment is scoped and supported before assuming the full platform is required.

That matters for the comparison. If you would buy Decagon for voice only, the like-for-like shortlist is other call platforms, and the questions on telephony routes, verification and calling rules above carry most of the weight. You can see the category view in the voice AI blog.

Is Decagon's voice trained on our customers' calls?

Decagon says Chord, the speech model behind Voice 3, is trained on licensed data and consented voice talent and never on customer-owned data, and its security page says it enforces zero-data retention with AI providers such as OpenAI and Anthropic. Confirm both statements in the contract, alongside where your recordings and transcripts are stored.

The contract is where these statements become enforceable. Ask for the training and retention commitments to appear in the data processing terms, and for the sub-processor list to name every model provider that can receive UK call content.

Want to see this on your own calls? Try Dilr Voice live, compare vendors in the best AI voice agent guide, book an AI placement diagnostic, or read how our approach places AI inside enterprise systems.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.