PolyAI Raises $86 Million as Britain Buys a Voice Stake

PolyAI closed an $86 million Series D on December 15, 2025, taking total funding past $200 million. Georgian, Hedosophia, and Khosla Ventures co-led. The new name on the cap table is a UK state bank, sitting beside NVIDIA and Zendesk as they came back.

The company sells human-sounding voice agents. The proof it handed investors is labor: more than 1,000 full-time employees of work at several clients, and a Forrester study whose biggest line is agent hours taken out of the building.

Three Returning Firms and an $86 Million Check

London-based PolyAI said the round funds Agent Studio and a wider sales push. It listed more than 100 enterprise customers, over 2,000 live deployments, 45 languages, and more than 25 countries. Named users include Marriott, Caesars Entertainment, PG&E, UniCredit, and Foot Locker.

Emily Walsh, lead investor at Georgian and a backer through three straight rounds, said customer service is “a massive opportunity for value creation” at the largest brands. In Georgian’s own note on the deal, Walsh’s firm said the agents are built to handle over 95% of interactions from start to finish without a person.

Nikola Mrkšić, CEO and co-founder, put the customer-base effect at about $1 billion of value a year. Amadeus Capital, an existing investor, said the round values the company at $750 million. CB Insights had put the May 2024 Series C at $50 million, with a valuation of $450 million to $500 million.

THE FUNDING TRAIL

1. 2017: Mrkšić, Shawn Wen, and Pei-Hao Su found PolyAI in London after Cambridge dialogue-system research.
2. 2022: Georgian leads the Series B after Walsh’s team runs customer reference calls.
3. May 2024: Series C of $50 million, valuation $450 million to $500 million, per CB Insights, with NVentures and Zendesk Ventures already in.
4. June 2025: Forrester Consulting publishes a PolyAI-commissioned Total Economic Impact study.
5. December 15, 2025: $86 million Series D closes; total funding moves above $200 million.

Additional investors in the new round include NVentures, the British Business Bank, Citi Ventures, Squarepoint Ventures, Sands Capital, Zendesk Ventures, and Point72 Ventures. Several of those names were not first-time checks. They were doubling down on a voice stack that had already left the pilot stage.

The British Business Bank Put In £15 Million

The UK government-owned British Business Bank announced a £15 million investment into PolyAI on the same day as the company statement. That is public money sitting on a cap table with Khosla and NVIDIA, earmarked for Agent Studio and for keeping a fast-growing AI firm in the UK.

THREE CHECKS MOST COVERAGE SKIPPED

• The state bank: The British Business Bank’s £15 million is the clearest new public-policy stake in the round.
• The chip arm: NVentures, NVIDIA’s venture unit, returned after the Series C; Amadeus Capital said Jensen Huang flagged the PolyAI check on a London stage with Prime Minister Keir Starmer.
• The helpdesk vendor: Zendesk Ventures is back too, even though a contained voice call is a ticket that never opens.

Leandros Kalisperas, the bank’s chief investment officer, framed the cheque as industrial strategy, not contact-center software.

The UK has all the right components to be a world leader in AI, with strong technological talent and the venture market to back it up. For the UK to realise its AI ambitions, we must support fast growing AI companies like PolyAI by providing them with the capital they need to succeed and grow here in the UK.

Leandros Kalisperas, Chief Investment Officer, British Business Bank press release

That growth story collides with the product story. Amadeus Capital quoted Mrkšić saying that within five years, 90% of contact-centre work will be automated, with leftover human roles “more creative and strategic.” The bank is funding UK scale-up. The CEO is funding a machine that does the floor work.

The Forrester Study Is Mostly Labor Savings

PolyAI points to a June 2025 Total Economic Impact study by Forrester Consulting, which the company paid for. Forrester interviewed four customers and rolled them into one composite: a US firm with 4 million calls a year and 200 agents. That model produced a 391% return on investment over three years, an $11.3 million net present value, and payback in under six months.

The benefits chart is not a vague “efficiency” story. It is hours, hires, and abandoned calls.

FORRESTER COMPOSITE, THREE-YEAR PRESENT VALUE

The $2.9 million of extra profit and the $2.9 million of costs are different rows. One is recaptured revenue and a 20% upsell lift. The other is what the composite pays PolyAI, plus rollout and upkeep. Usage alone is about $2.6 million over three years in the study’s present-value math.

Call mix in the model rises from 25% handled in year one to 35% in year two and 40% in year three. Live-agent hours saved run 133,333, then 168,000, then 191,100. By year three the composite has avoided 92 new hires (25, then 31, then 36) by letting attrition stand. Abandonment falls 50%. Attrition among remaining agents falls 25%. Forrester counts more than 4.2 million calls resolved by the software by year three, and more than 211,000 missed revenue chances recaptured.

A hospitality contact-center director in the study put the human piece second.

We’re providing a much better customer experience because we can answer the call right away. We are then able to provide an experience and an opportunity for customers to take care of their business effectively without the need to speak to an agent; if they do need to speak to an agent, we have one available. So, the cost savings are obviously great, but it’s more about what we are doing to make it easier for our customers to do business with us.

Director, contact center operations, hospitality, Forrester TEI interviews

The four interviews behind that composite were not small shops. An energy senior director sat on 16 million calls and 850 agents. A healthcare co-founder ran 4.2 million calls. A hospitality director ran 960,000. A UK insurance customer-services director ran 480,000. Before PolyAI, they were on keypad IVR and hiring. After it, they stopped backfilling a chunk of the queue.

PG&E Already Answers the Wildfire Spike

Pacific Gas and Electric is the deployment Georgian and Mrkšić both reach for when they need a peak-load proof. The utility takes about 16 million calls a year. Weather turns that into a crush. Mrkšić said the system still answers when wildfires set off over a million calls in a single day. Wait times on the old tree ran past an hour.

PolyAI’s case study records 35,000 labor hours saved at PG&E, a 22% CSAT lift on outage calls, and 67% containment, 6 percentage points above the legacy IVR. The internal ROI gate was 3.5% containment. The agent hit 15% on day one. Georgian adds that the outage line runs in English and Spanish.

UniCredit’s Croatian unit, Zagrebačka banka, is the regulated-language proof. The bank has over 1.3 million retail customers and 60,000 corporate clients. The old IVR left 25% of calls abandoned. Mia, a Croatian-speaking assistant, went live in 3 months against a one-year forecast. It now automates 27% of calls, routes 83% faster, cuts abandonment 10%, and raised NPS 14 points in the first 6 months. Vedran Mrvica, head of the contact center, has already pushed the same stack elsewhere in the UniCredit group.

TWO LIVE DEPLOYMENTS

Mrkšić’s December 17 letter stacked other loads next to those two. Tens of millions of medical-transport calls. Las Vegas Strip bookings through Golden Nugget. Thousands of restaurants, with revenue up as much as 10%. The through-line is the same: the phone still rings first when something breaks, and the new round is for the voice that picks up.

Why Zendesk Is in a Voice-Agent Round

Zendesk Ventures does not need a press quote for the conflict to be visible. A voice agent that finishes authentication, a booking, or a payment is a ticket that never lands in a helpdesk queue. Writing a second cheque is a hedge: own a piece of the channel that can starve the old one.

NVIDIA’s repeat cheque is easier to read. Spoken agents run recognition, turn-taking, and a live model on every call. That is inference on a clock a chat widget never sees. NVentures was already in at Series C. Coming back with Khosla and a state bank is a claim that voice, not another chatbot wrapper, is where the GPUs get paid.

The market that formed around those cheques is not asking whether an agent can chit-chat. It is asking whether the 40 common call types close without a person. Promo posts around the round treated 2,000 live deployments as the end of the pilot era. The sharper read is simpler. Capital is now priced on replacement, not on a copilot sitting next to a headset.

That is also why a $750 million London company can look small next to US agent startups that raised on chat-first stories. PolyAI’s own figures are production loads, Forrester hours, and a state bank that wants the firm to stay put. The multiple is a later argument. The floor is already in the case studies.

Mrkšić Tied the Money to Agent Studio

PolyAI said the $86 million goes into the proprietary stack behind Agent Studio and into finding more global buyers. Georgian describes that stack as frontier models plus PolyAI’s own speech models, aimed at multi-turn calls in regulated industries, with forward-deployed engineers sitting inside the client’s systems rather than dropping a bot on the IVR and leaving.

Mrkšić’s public wager on the back of the round was blunt. If PolyAI cannot automate 70% of a prospect’s calls, it will donate $100,000 to a charity the prospect picks. He posted that offer the day after the raise, with the Series D video.

When something goes wrong — a power outage, a cancelled vacation, even a wildfire — the first thing people do is pick up the phone. At PolyAI, we’re the voice that answers those calls.

We just closed our Series D, surpassing $200M in total funding to make sure your critical… pic.twitter.com/4iqVNJx7qR

— Nikola Mrkšić (@nikola_mrksic) December 16, 2025

The same week, Amadeus Capital carried his five-year line: 90% of contact-centre work automated, humans kept for the creative remainder. He has been selling that future since the VocalIQ days, before Apple bought that dialogue shop into Siri and before he left to rebuild the stack for enterprises. The $86 million is the latest fuel. The overlooked signatures on it are a UK state bank, a chipmaker, and a helpdesk vendor, all paying for a voice that already does a thousand people’s days.

Frequently Asked Questions

Who Co-Led PolyAI’s Series D Round?

Georgian, Hedosophia, and Khosla Ventures co-led the $86 million round, and Walsh at Georgian had already led the 2022 Series B. NVentures, Zendesk Ventures, Sands Capital, and Point72 Ventures were returning names from the May 2024 Series C that CB Insights logged at $50 million. Citi Ventures, Squarepoint Ventures, and the British Business Bank appear on the Series D list as the fresh institutional layer around that core.

What Did the Forrester PolyAI Study Measure?

Forrester Consulting interviewed four PolyAI customers in energy, healthcare, hospitality, and insurance, then built a single US composite with 4 million calls a year and 200 agents. Implementation in that model is four weeks and about 200 internal labor hours, with several later professional-service bursts as new use cases go live. The 391% ROI is risk-adjusted present value over three years, not a single-site cash figure.

Who Founded PolyAI and When?

Nikola Mrkšić, Tsung-Hsien Wen (Shawn), and Pei-Hao (Eddy) Su founded the company in 2017 as a London spin-out from Cambridge’s dialogue-systems group, working under speech researcher Steven Young. Mrkšić was the first engineer at VocalIQ, the dialogue firm Apple bought and folded into Siri, before the three rebuilt an enterprise voice stack from scratch instead of wrapping a general chatbot around a phone tree.

How Much Did the British Business Bank Invest in PolyAI?

The bank put in £15 million as part of the $86 million Series D and said the cash is for Agent Studio, so more of the company’s growth stays in the UK. Its own release named FedEx, Marriott, UniCredit, and Foot Locker as reference customers, a slightly different mix from PolyAI’s Marriott-Caesars-PG&E list, and tied the cheque to a wider UK AI scale-up brief rather than to a product feature.