Coval

What it is

Coval is a simulation and evaluation platform for voice and chat AI agents. Before launch, it calls your agent with thousands of synthetic customers — different personas, accents, background noise, interruptions, off-script requests — and scores every conversation against metrics you define. After launch, it scores live calls the same way and routes the doubtful ones to human reviewers. The question it answers is the one every CX team asks the week before a voice agent goes live: how do we know it will not embarrass us on call number 4,000?

The approach comes from self-driving cars. Founder and CEO Brooke Hopkins led evaluation infrastructure at Waymo, and Coval treats an AI agent the way Waymo treats a car: you do not prove it safe by driving it around the block, you prove it by running it through a very large number of simulated scenarios and measuring the failure rate. The company went through Y Combinator’s Summer 2024 batch, raised a $3.3M seed led by MaC Venture Capital, and on 24 June 2026 announced a $28M Series A led by Norwest, with Base10 Partners, Twilio Ventures and Y Combinator participating, for $31M raised in total. Zoom and Deepgram are named customers in the funding release; the homepage adds Chime, Perplexity, ServiceNow, StubHub, Toast and Upstart. The vendor says it works with more than 60 organizations.

Why it shows up in CX and support stacks

• It tests the audio, not just the transcript. Evaluations cover timing, turn-taking, interruptions, audio quality, tool calls and caller sentiment. A voice agent that says the right words two seconds too late still fails, and a transcript-only eval will not catch it.
• It connects to how voice agents are actually built. Documented connections include inbound and outbound phone calls, voice over WebSocket, LiveKit, Pipecat Cloud, OpenAI Realtime, Gemini Live, Grok Realtime and Deepgram Voice Agent, plus chat over HTTP, WebSocket and A2A, and SMS. An agent built on Vapi or Retell AI is tested by calling its phone number.
• It grades actions, not only words. Agents can send OpenTelemetry traces, or Coval imports them from Langfuse, Arize Phoenix or LangSmith, so a metric can check that the refund tool was called with the right amount rather than that the agent said it issued a refund.
• Vendor bakeoffs. The same scenario set runs against several voice platforms, which makes it a procurement tool as much as a QA tool: a CX leader choosing between Vapi, Retell AI and PolyAI can compare them on the company’s own call types instead of on demos.
• Built for developer workflows. REST API, a coval CLI for CI pipelines, typed TypeScript and Python SDKs, and an MCP connector that lets Claude, ChatGPT and Codex launch runs and query results. All of it is included on every tier.

Pricing reality

Coval is one of the few vendors in this segment with a public rate card. Starter is $100 a month (100 simulation minutes, 1,000 monitored calls, 30-day trace retention, 5 concurrent simulations). Growth is $500 a month (1,000 simulation minutes, 10,000 monitored calls, 90-day retention, 25 concurrent simulations). Enterprise starts at $4,500 a month, or $54,000 a year, and adds SAML SSO, SCIM, data residency and private VPC deployment. Seats are unlimited on every tier; annual billing takes 20% off Starter and Growth; both carry a 7-day trial that needs a card. One inconsistency to clear before signing: the Growth plan is listed with unlimited projects in one place on the pricing page and 5 projects in the comparison table.

The included minutes are the floor, not the bill. Overage is $0.25 per simulation minute on Growth and $0.40 on Starter. A modest regression suite — 200 scenarios averaging 3 minutes — burns 600 minutes per run. Run it nightly and you use about 18,000 minutes a month: 17,000 minutes of overage at $0.25 is $4,250, plus the $500 base, or roughly $4,750 a month. That is above the Enterprise floor. The practical rule: Growth fits a team that runs its full suite on each release; a team that runs it on every commit or every night should price Enterprise from the start.

Best for

Voice AI product owners and CX operations leads at companies putting a voice or chat agent in front of real customers — contact-centre deflection, appointment booking, collections, order status — who need a pass/fail gate before each release and production scoring after it. It fits best when an engineering team owns the agent and will wire the CLI into CI, and when an operations or QA team will own the metrics and the human-review queue.

Skip it if you buy a fully managed agent such as Sierra or Decagon and have no access to change its prompts or tools: you can measure failures but not fix them, and a support-QA tool such as Solidroad or Intryc that grades bot and human conversations together is the better spend.

Versus the alternatives

Cyara, which owns Botium, is the enterprise incumbent in CX assurance and launched agentic testing in March 2026. It is sales-led with no self-serve tier and is the pick when you already test IVRs and contact-centre telephony with Cyara and want one vendor. Hamming is the other 2024-vintage, YC-backed specialist, with 50+ built-in metrics, a red-team module and a strong focus on audio signal, latency and telephony conditions; it publishes no prices and raised a $3.8M seed in December 2024. Pick it when audio-level failure analysis is the main problem. Cekura is the cheapest way in: pay-as-you-go at $0.25 per voice-testing minute with no base fee, or $500 a month for about 2,000 minutes — roughly twice the simulation minutes of Coval’s $500 Growth tier. Pick it for a small team that needs volume on a budget. Roark and Bluejay, both founded in 2025, are the newest entrants; Roark starts from production calls and replays them as tests. Braintrust is the pick when the agent is chat-first and you already run LLM evals there.

Coval wins on the combination of simulation depth, published pricing, and funding: at $31M raised it has the largest disclosed funding of the voice-testing startups named here. If none fit, script 30 of your highest-risk call types, have two people place them by phone against a staging agent before each release, and score them on a shared sheet. When that takes more than a day per release, you are ready to buy.

Watch-outs

• Overage turns a $500 plan into an Enterprise bill. Simulation minutes scale with scenarios × length × frequency, not seats. Guard: before the trial ends, count your scenarios, multiply by average call length and run frequency, and price both Growth-plus-overage and Enterprise on that number.
• Synthetic callers are LLMs, and the judges are too. A simulated caller is more articulate and more patient than the person who rings at 7am from a car. Guard: seed scenarios from uploaded production transcripts rather than invented personas, and double-score 100 calls by hand to measure how often each LLM-judged metric agrees with a human before letting it block a release.
• Compliance paperwork varies by page and tier. The pricing page lists SOC 2 Type II, HIPAA and GDPR on every tier, the homepage marks GDPR as in progress, and a third-party review in May 2026 placed the BAA and DPA on Enterprise. Guard: get a signed BAA and DPA, and check the trust centre, before any real call recording with health or payment data is uploaded.
• The headline outcomes are vendor-reported. The up-to-30× reduction in manual QA and 10× faster deployment come from Coval’s funding release. Guard: record the current reviewer hours per release and the number of production incidents per month before go-live, and compare at day 60.