Scenario · AI & automation

Building an AI voice interviewer for an HR startup: scope, weeks, and what it really costs

What AI voice agent development costs for a recruiting startup: an 8-week voice interviewer build, the cost table, what to cut, and the compliance traps.

By Akash Bansal, founding engineerUpdated 4 November 20268 min read

Illustrative scenario. The founder in this story is a composite, not a real client. The process, costs and decisions are drawn from builds I have done; names and business details are invented to show how a build like this typically goes.

Who is the founder, and where did she start?

Meera is a composite: an invented founder built from the conversations recruiting founders tend to have. She spent eight years as an agency recruiter in Toronto filling high-volume hourly roles: warehouse, retail, customer support. She has a sharp, specific pain. Her clients get 400 applicants per opening, and the first-round phone screen is the same ten questions every time. Her recruiters spend their days asking "Can you work weekends?" and "Tell me about a time you handled an angry customer."

Her idea: an AI voice interviewer that calls or meets every applicant within minutes of applying, runs a structured 10-minute screen, scores it against the role's rubric, and hands the recruiter a ranked shortlist with the evidence attached. She had a Figma prototype, two staffing firms willing to pilot, and roughly $40,000 of her own money set aside for product.

She is non-technical. She can't read code, and she was quoted wildly different numbers for what she thought was the same thing.

Agency, voice platform, freelancer or founding engineer: which did she pick?

She picked a founding engineer, because the interview logic and the scoring were the product, and she needed to own both. Here is how her four options compared.

The honest version of my advice to a founder like Meera: if you haven't yet tested whether candidates will finish a call with an AI, spend two weeks on a hosted platform first. You don't need me for that. She already had pilot partners asking for custom rubrics and a recruiter dashboard, so that question was answered.

The evidence behind my side of the call: I built NextMantra solo. It is an AI recruiter in production that runs live voice interviews, scores résumés against roles, and produces ranked shortlists. The interview agent is Python on FastAPI with Pipecat, Daily for real-time audio, and Groq for fast inference. Scoring uses several LLMs against the role's rubric. I've already hit the walls a product like Meera's runs into.

What got cut from v1?

About half of her wish list got cut. The test for every feature was simple: does a recruiter at the pilot firm need this to trust a shortlist in week one?

• Cut: video and avatars. Audio-only is cheaper to run, works on bad phone connections, and avoids a whole class of appearance-bias questions.
• Cut: ATS integrations. CSV export and a shareable shortlist link cover the pilot. Each ATS sync is its own project.
• Cut: multi-language. English first. French came next for Canadian clients, after the English rubric scoring was measured.
• Cut: phone dial-out. A browser link sent by SMS and email came first. Telephony adds per-minute carrier cost and its own failure modes.
• Kept: compliance features. AI disclosure text, candidate consent, transcript retention settings and a human-review step. These are not extras in hiring. More on why below.

Week by week: how the 8-week build went

Weeks 1–2: scoping, the design doc, and the text pipeline

Week one produced a written design doc: the role setup screen, the rubric format (each criterion, its weight, and what a strong or weak answer sounds like), the data retention policy, and the fixed quote. Week two built the whole interview as text first. A recruiter creates a role, the system generates a question plan, and a typed "candidate" answers in a chat box. If the questions and scoring don't work in text, voice won't save them.

Weeks 3–5: the real-time voice loop

This is the part people underestimate. A voice interviewer is a pipeline: the candidate's audio streams in, speech-to-text transcribes it, the LLM decides what to say, text-to-speech speaks, and all of it has to feel like a conversation. Pipecat handles the pipeline and Daily handles the audio rooms, so the engineering time goes into behaviour. When has the candidate finished talking? What happens when they pause to think, talk over the bot, or say "sorry, can you repeat that"? What if their connection drops halfway through?

A deeper technical write-up of this loop, including the latency budget, is in how I built a real-time AI voice interviewer solo.

Weeks 6–7: scoring and the recruiter view

Each transcript is scored against the rubric by more than one model, and each score comes with quoted evidence from the transcript. Where the models disagree, the candidate is flagged for a human rather than averaged quietly. NextMantra also scores with several LLMs against the role's rubric, and I explain why one model shouldn't grade a person alone in multi-LLM scoring. The recruiter sees a ranked list, can play any answer, and can override any score. The overrides are logged.

Week 8: compliance, load test, pilot

The last week covered the disclosure and consent screens, a load test simulating a burst of applicants after a job post goes live, and the first real interviews with the pilot firm's applicants.

What does an AI voice interviewer cost to build and run?

The build is a fixed quote, and running cost is per interview minute. This is the shape of the budget for a v1 like Meera's. For a rough build range on your own scope, the app cost calculator shows every assumption it uses.

The structural point matters more than any single price. The build is a one-time cost. The per-minute cost scales with every interview, forever. So the cuts that lower the running cost (audio-only, short structured interviews, scoring once after the call) are worth more than the ones that save a week of build time.

What went wrong, and what surprised her?

• Candidates pause, and the bot hears silence as "done". People thinking about a behavioural question go quiet for a few seconds. The first version jumped in too early. Tuning end-of-turn detection, and having the bot say "take your time" instead of moving on, fixed most of it.
• Recruiters didn't trust a number. They trusted a quote. A score of 7.2 meant nothing to them. "Said she's covered weekend shifts for two years at a grocery chain" next to the score did. Evidence-linked scoring became the feature that sold the pilot.
• Compliance came up in the first sales call, not the tenth. One pilot client hires in New York City. Its counsel asked about the bias audit before asking about the product.

Which hiring laws does an AI interviewer have to respect?

At least two apply to a Toronto founder selling across North America, and both change the product.

• Ontario. Since January 1, 2026, Ontario employers with 25 or more employees must say in publicly advertised job postings whether AI is used to screen, assess or select applicants (Osler summary). Meera's product generates that disclosure line for her clients.
• New York City. Local Law 144 bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within the past year. They must publish a summary of the results and notify candidates 10 business days before the tool is used (NYC DCWP). That shaped the data model: demographic fields are collected separately, optionally, and never shown to the scoring models.

I'm an engineer, not a lawyer. The design choice that matters is to build the hooks (disclosure, consent, audit exports, human override) into v1, so a lawyer's answer becomes a configuration change rather than a rebuild.

What did launch look like?

Launch was a quiet pilot with one staffing firm on two roles. Applicants got an interview link minutes after applying. Recruiters reviewed the ranked list each morning and overrode scores they disagreed with. Meera read every override. Each one was either a rubric fix or a bug, and the rubric fixes cost nothing to ship. She went into her second pilot with a demo built from real anonymised interviews instead of a Figma file.

Checklist: building an AI voice interviewer for your HR startup

1. Prove the interview works in text before you pay for voice.
2. Write the rubric before the prompt. Each criterion needs a weight and an example of a strong answer.
3. Ship audio-only first. Video, avatars and phone dial-out each add cost and risk.
4. Show the evidence, not just a score. Recruiters trust quotes.
5. Route disagreements between models to a human, and log every override.
6. Build disclosure, consent, retention and audit exports into v1.
7. Price your product per interview, because that is how your costs behave.
8. Own the repo, the cloud account and the model API keys from day one (who owns the code).

If you're a recruiting founder at this point, with a pilot partner and a rubric in your head, this is the kind of build I do: one engineer, a fixed quote, a working demo every week, and you own everything. Here's how I work. If you're still unsure whether candidates will talk to an AI at all, start with a hosted platform and come back when you need the scoring to be yours. To compare the ways of getting this built, read hiring an AI agent developer.