Exotel vs Bolna AI in 2026: Telephony Platform or Voice Agent Layer? An Honest Comparison for Indian Teams

The engineering lead at a Pune-based consumer lender has two browser tabs open and a deadline on Friday. One tab is Exotel, the company his operations team has used for IVR and agent calling since 2019. The other is Bolna, the voice agent platform a developer on his team prototyped over a weekend, the one that held a passable Hindi conversation about a missed EMI on the first try. His CFO wants one line in the board deck: which vendor runs the AI calling programme next quarter.

He searches "Exotel vs Bolna AI" and finds the problem with the question. Half the pages call them competitors, the other half partners. Bolna's documentation explains how to connect an Exotel account, and Exotel's developer docs show how to stream a live call into any bot, Bolna included.

This post is the comparison he needed. We compete with both companies in parts of the market, so every vendor fact here is one the company publishes itself.

The short version: Exotel and Bolna are not two answers to the same question. Exotel is the telephony and contact-centre layer that has added voice AI on top. Bolna is a voice agent orchestration layer that needs telephony underneath it, and can use Exotel for that. You are not choosing between them so much as choosing which layer you want a vendor to own and which layer you want to own yourself. By the end of this post you will be able to map your team onto that choice, estimate the cost structure of each path, and know the specific questions that expose a weak fit before you sign anything.

Why this comparison is confusing in 2026

Two years ago the boundary was clean: telephony vendors moved calls, AI vendors built bots. Three things have blurred it.

Telephony vendors moved up the stack. Exotel acquired Cogno AI, a conversational AI company, in 2021, and merged with the contact-centre platform Ameyo the same year. In 2026 its AI product surface includes AgentStream (real-time call audio streamed to a bot over WebSocket), voicebot flows inside its contact-centre product, and, since August 2026, a beta Exotel MCP Server that lets an AI model such as Claude, GPT or Gemini place calls through Exotel from a plain-English instruction.

Agent platforms moved down the stack. Bolna started as an open-source orchestration framework on GitHub. It now sells phone numbers from its own dashboard (Indian numbers through Plivo or Vobiz at USD 5 per month each), documents how to get 140-series and 160-series numbers, and lets you plug in Twilio, Plivo, Exotel, Vobiz or your own SIP trunk.

Regulation raised the cost of getting telephony wrong. TRAI's designated number series (140 for promotional calls, 1600 for service and transactional calls by RBI, SEBI, IRDAI and PFRDA regulated entities) and the DLT registration regime mean the telephony layer is no longer a commodity pipe you can swap on a whim. The vendor that holds your numbers and your DLT registrations holds a lot of your compliance posture. We cover that regime in depth in our TRAI DLT compliance guide for AI outbound calling.

So both vendors now share shortlists and budget lines. Underneath, they remain different kinds of company.

What each company actually is

Exotel

Exotel was founded in Bengaluru in 2011. It describes itself as a customer engagement platform and says it serves more than 7,000 brands and processes over 25 billion interactions a year. Its core business is cloud telephony: virtual numbers, IVR, call routing, number masking, voice and SMS APIs, and, through the Ameyo merger, a full contact-centre suite for human agents.

Its voice AI offer comes in three shapes, and it matters which one a salesperson is describing:

1. AgentStream. A streaming layer. Exotel streams live call audio to your bot endpoint over WebSocket and plays your bot's audio back to the caller. It supports bidirectional streaming for conversational AI and unidirectional streaming for listen-only use cases such as agent assist. Exotel markets sub-20 ms media latency for this layer. Crucially, AgentStream does not include the brain: you bring the speech recognition, the language model, the voice synthesis and the conversation logic, or you bring a platform that does.
2. Voicebot and no-code flows inside the contact-centre product. Bot journeys built in a drag-and-drop builder alongside IVR and human-agent queues. This is the shape most existing Exotel contact-centre customers meet first.
3. Exotel MCP Server (beta). A Model Context Protocol server that exposes Exotel's voice, SMS and WhatsApp channels to any MCP-compatible AI client. The pitch is "from a sentence to a live call," with no SDK work.

Exotel does not publish per-minute pricing for its AI products. Its business phone and contact-centre plans are listed on its pricing pages, and AI is a sales conversation. Check current pricing directly rather than trusting a number from a third-party blog, including this one.

Bolna

Bolna is a Bengaluru voice AI company that went through Y Combinator's Fall 2025 batch and raised a USD 6.3 million seed round led by General Catalyst, announced in January 2026. At that time the company said it had grown from about 1,500 calls a day at its first commercial rollout in May 2025 to more than 200,000 a day, across more than 1,050 paying customers. Those are the company's own figures.

The product is an orchestration layer for voice agents. It wires together a speech-to-text model, an LLM and a text-to-speech model, handles turn-taking over a live phone call, and exposes the whole thing through an API and a no-code playground. Three facts define it:

1. Open-source core. The framework is on GitHub under the MIT licence. The hosted APIs and the dashboard are closed-source and sit on top of it. New features, by the project's own description, land in the open-source package first.
2. Bring your own keys. You can connect your own accounts for the LLM, the speech recogniser and the voice synthesiser, in which case Bolna does not charge you for those components and you pay the providers directly plus Bolna's platform fee.
3. Telephony is pluggable. Twilio, Plivo, Exotel, Vobiz and SIP trunks are documented options. Bolna can also sell you the numbers.

Bolna publishes its pricing, which we cover below.

The layer model: the only framing that makes this comparison useful

Every AI phone call in India passes through the same seven layers, from the bottom up:

Layer	What it does	Exotel	Bolna
Numbers, carriers, DLT	Holds the number, routes to carriers, carries DLT and number-series compliance	Owns it, core business since 2011	Resells via Plivo and Vobiz, or you bring Twilio, Exotel, Vobiz, SIP
Media streaming	Moves live audio between the call and the AI	Owns it (AgentStream)	Relies on the telephony partner
Speech-to-text	Turns caller audio into text	Customer brings it on AgentStream	Orchestrates chosen provider, or BYOK
LLM	Decides what to say next	Customer brings it on AgentStream	Orchestrates chosen provider, or BYOK
Text-to-speech	Turns the reply into a voice	Customer brings it on AgentStream	Orchestrates chosen provider, or BYOK
Conversation logic	Prompts, tools, transfers, turn-taking	Partial: voicebot flows, MCP beta	Owns it, core product
Campaigns and CRM	Who to call, when, dispositions, handoff to humans	Owns it via contact-centre suite	Partial: batch calls and APIs

Read the table vertically and the comparison resolves itself. Exotel's centre of gravity is the bottom two rows and the top row: the parts of a calling programme that existed before AI. Bolna's centre of gravity is the middle four rows: the parts that are new.

That is also why the two products show up together in real deployments. Bolna documents an Exotel connector for both inbound and outbound calls on Exotel numbers. Exotel's AgentStream documentation is written for exactly the kind of bot Bolna provides. A setup with Exotel numbers and contact centre, Bolna agents in the middle and your own CRM on top is a perfectly coherent architecture, and for a team that already runs on Exotel it is often the lowest-friction one.

Where Exotel wins

You already run on Exotel. If your numbers, IVR, agent queues and call recordings already live in Exotel, every other choice adds a migration. Number porting in India is slow, DLT registrations are tied to entities and headers, and your operations team knows the Exotel dashboard. Adding AI on the platform you already have removes a category of risk that no feature list can outweigh.

You need the human contact centre and the AI in one place. The most common production pattern in Indian BFSI and D2C is not "AI replaces agents." It is "AI handles the first conversation and hands off to a human with context." Exotel owns the human side of that handoff natively. With a pure orchestration platform, the human queue lives somewhere else and the handoff is a SIP transfer you have to engineer and test.

Scale and carrier relationships. At 25 billion interactions a year, Exotel has carrier history, regional routing and failover a two-year-old platform has not had time to build. When an operator throttles a number range at 7pm on month-end, whose escalation path you are on matters.

Compliance surface. Exotel's security page lists ISO 27001, PCI-DSS and SOC 2. Regulated buyers (banks, insurers, large NBFCs) often need those attestations from the vendor carrying their call audio before procurement will even schedule a meeting.

Where it is weaker. On AgentStream, the AI itself is your job. Your team chooses and integrates the speech recogniser, the LLM and the voice, then owns latency tuning, barge-in handling and prompt quality. Exotel publishes good guidance on this (its February 2026 latency guide targets an average response under 700 ms when every layer is tuned), but guidance is not a managed agent. The no-code voicebot flows are easier but more constrained than a developer-first agent platform, and the MCP Server is labelled beta.

Where Bolna wins

Speed from idea to a talking agent. A developer can have a working Hindi or English agent on a real phone number in an afternoon, and the playground lets a non-developer iterate on prompts. For a team that has not yet proven voice AI works for its use case, that speed is worth a lot.

Model choice and cost control. Because Bolna orchestrates rather than builds models, you can swap the speech recogniser, LLM or voice without changing vendor, and with BYOK you pay model providers directly at your own negotiated rates. For high-volume buyers who already have an LLM contract, this is a real lever.

Transparent, published pricing. Bolna lists a standard rate of 6 US cents per minute, falling to about 4.51 cents at the top of its volume curve, a USD 0.02 per minute platform fee when you bring your own model keys, and pilot packs of USD 300 for 6,500 minutes or USD 500 for 12,000 minutes. You can model your cost before you talk to anyone.

Open-source escape hatch. The MIT-licensed core means you are not fully locked in. If the hosted product stops fitting, you can in principle run the framework yourself. In practice, self-hosting a voice stack at production quality is a serious engineering commitment, but the option changes your negotiating position.

Where it is weaker. Concurrency defaults are modest: by Bolna's own FAQ, paid accounts get 10 concurrent calls by default and the pilot plan caps at 25, with higher limits on enterprise plans. That is fine for a pilot and a constraint for an EMI reminder run on the 5th of the month. The telephony layer is someone else's, so carrier-level issues route through two vendors. Campaign management, dispositions and the human contact centre are thinner than in a contact-centre suite, so you will build or buy those around it. And the bill is a composite: platform fee plus model providers plus telephony plus numbers, which finance teams sometimes find harder to forecast than one invoice.

The scorecard above is an editorial judgement, not a benchmark, and the same scores in a table:

Dimension	Exotel	Bolna	Why
Telephony depth and carriers	5	2	Exotel's core business; Bolna resells or connects partners
Human contact centre handoff	5	2	Exotel owns the agent queue; Bolna transfers out
Time to first working agent	3	5	Bolna's playground and defaults versus building on AgentStream
Model choice (ASR, LLM, TTS)	4	5	AgentStream is model-agnostic but you integrate; Bolna orchestrates with BYOK
Pricing transparency	2	5	Bolna publishes per-minute rates; Exotel AI is quote-based
Default scale without a contract	5	2	Bolna defaults to 10 concurrent paid calls, 25 on pilot

Note what is missing from that table: call quality. Neither vendor's model choices are fixed, so the speech recogniser, LLM and voice you select matter more than the platform. A Bolna agent on a weak Hindi recogniser will lose to an AgentStream bot on a strong one, and the reverse. Test on your own audio. Our voice AI latency benchmarks for India explain how to measure it.

The cost structure, honestly

The search results for this comparison are full of confident per-minute figures for both vendors. Most of them compare different things. Here is what can actually be verified, and what you have to ask for.

Bolna, published as of October 2026:

Item	Published figure
Standard all-in rate	6.00 US cents per minute
Best volume rate shown	about 4.51 US cents per minute
Platform fee when you bring your own model keys	USD 0.02 per minute, plus your model providers' charges
Pilot pack	USD 300 for 6,500 minutes, or USD 500 for 12,000 minutes, concurrency cap 25
Indian phone number	USD 5 per month per number
Free trial credit	USD 5

Bolna's pricing page shows the standard 6-cent rate as ₹5.52 per minute. The flat rate covers Bolna's preferred speech, LLM and voice models; other models are billed separately. Check whether telephony minutes are included for your specific configuration, especially if you bring your own Exotel or Plivo account, because then the carrier bills you directly.

Exotel: business phone and contact-centre plans are on its public pricing pages. AI products (AgentStream, voicebot, MCP) are priced through sales. If you run AgentStream, your per-minute cost is Exotel's telephony and streaming charge plus whatever your chosen speech, LLM and voice providers charge, plus the engineering time to integrate and maintain them. That last line is the one most comparisons ignore and the one that usually dominates in the first year.

Costs that sit outside both vendors: DLT principal-entity registration and template approval (Bolna's documentation cites a ₹5,900 payment link for 140-series principal-entity registration through its partner), number-series procurement, call recording storage under your retention policy, and the CRM integration work.

The practical conclusion: for a pilot under about 10,000 minutes a month, Bolna's published pricing is easy to model and hard to beat on setup cost. At scale, the comparison becomes "Bolna's per-minute rate or platform fee plus model bills" against "Exotel's quoted rate plus your model bills plus your engineering team," and the engineering line decides it. We go deeper on why per-minute comparisons mislead in per-minute vs per-outcome voice AI pricing.

What goes wrong on each path

These are the failure modes we see most often when Indian teams pick one of these paths without mapping the layers first.

1. Picking Bolna and discovering the concurrency ceiling on campaign day. EMI bounces cluster between the 3rd and 7th of the month. A reminder run that needs 300 simultaneous calls on a platform defaulting to 10 is a 30x gap. Negotiate concurrency in writing before the pilot ends, and load-test it. Our concurrency benchmark shows what breaks as channels scale on any platform.

2. Picking AgentStream and underestimating the bot. Streaming audio is the easy 20%. Barge-in, end-of-turn detection on code-switched Hinglish, retries on silence, tool calls into your loan-management system and transfer to a human are the other 80%. Teams that budget two sprints routinely spend two quarters.

3. Two vendors, no owner for a dropped call. In a Bolna-on-Exotel setup, a call that connects and goes silent could be the carrier, the stream, the speech recogniser or the LLM. Agree in advance which vendor owns first-line triage and what logs each will share. Ask both for a sample incident report.

4. DLT treated as a later problem. Neither vendor can make an unregistered promotional campaign legal. Header registration, template approval, 140-series versus 1600-series selection and DND scrubbing at dial time are your obligations. Start them in week one, because approvals take longer than the bot.

5. Comparing demo voices instead of production audio. Demos use clean audio and a script. Real Indian calls are noisy, code-switched and regionally accented. Run your own recordings through each configuration and measure word error rate and latency yourself.

6. Ignoring the human handoff. If 15 to 25% of calls need a human, the quality of the handoff determines customer experience more than the AI does. On Exotel the agent queue is native. On Bolna you are transferring into something, and that something needs to receive context, not just audio.

Which one should you pick?

In words, the tree resolves to four answers:

1. Already on Exotel, with a strong engineering team: use Exotel AgentStream and build your own agent. You keep one telephony vendor and full control of the AI, at the cost of owning the hard 80% of the bot.
2. Already on Exotel, without that team: put Bolna on top of your Exotel numbers. You keep your numbers, DLT setup and agent queues and get a working agent layer without building one.
3. Not on Exotel, running a pilot under about 25 concurrent calls: start with Bolna hosted. Published pricing, fast setup and an open-source exit if you outgrow it.
4. Not on Exotel, needing production scale from day one: you need a contract either way. Negotiate Bolna enterprise concurrency, take an Exotel enterprise quote, or look at a managed platform that sells the outcome rather than the stack.

Compliance: what neither vendor can do for you

Indian buyers sometimes assume the telephony vendor "handles TRAI." It handles the infrastructure. The obligations are yours.

DLT registration. As the principal entity, you register on DLT, register headers and, for promotional voice, register and get approval for templates. Bolna's documentation is explicit that DLT registration is mandatory before it can procure 140-series or 160-series numbers for you. Exotel customers go through the same regime.

Number series. Promotional calls go on 140-series numbers. Service and transactional calls by entities regulated by RBI, SEBI, IRDAI or PFRDA go on the 1600 series, and the phased rollout has been extending to smaller regulated entities through 2026. Using the wrong series is a compliance failure regardless of which vendor connected the call.

Calling hours and consent. Commercial calls are restricted to 9am to 9pm, enforced at the platform, and consent under the DPDP Act has to be purpose-specific. Both vendors can enforce windows if you configure them. Neither can decide on your behalf whether a particular borrower consented to an AI call.

Sector conduct rules. RBI's Fair Practices Code governs how a recovery call may be conducted. IRDAI expects disclosed recording on insurance sales calls. These rules live in your prompts and call flows, so they are your engineering responsibility on either platform.

Data residency and recordings. Ask both vendors, in writing, where call audio and transcripts are processed and stored, and which subprocessors (speech, LLM, voice providers) touch the audio. On a BYOK setup, the model providers you bring are your subprocessors too.

Where Caller Digital fits

We compete with both companies in parts of this market, so take this section with that in mind.

Caller Digital sits in a different spot from either vendor. We do not sell raw telephony like Exotel, and we do not sell a stack you assemble yourself like Bolna. We run the whole call, from numbers and DLT setup through Indian-language agents to CRM write-back, and we price at scale on outcomes rather than minutes. That suits teams that want a working collections, verification or lead-qualification programme without staffing a voice-AI engineering team. It suits a team that wants to own its model choices and prompt engineering less well, and for that team Bolna is a reasonable choice.

If you are evaluating us alongside either vendor, the direct comparisons are on Caller Digital vs Exotel and Caller Digital vs Bolna, and our longer write-ups are Bolna vs Caller Digital and voice AI vs Exotel, Knowlarity and Ozonetel.

A two-week evaluation plan you can run on both

Run the same test on each path before committing.

1. Days 1 to 2: pick one workflow and 200 real recordings. One workflow, such as EMI reminders or COD confirmation, not three. Pull 200 anonymised recordings from your own calls across regions and languages.
2. Days 1 to 5: start DLT and numbers in parallel. Register or confirm your principal entity, headers and templates, and request the right number series. This is the step that slips.
3. Days 3 to 7: build the same agent twice. One on Bolna (hosted or on your Exotel numbers), one on Exotel AgentStream or voicebot with the same speech, LLM and voice models where possible, so you compare platforms rather than models.
4. Days 8 to 10: measure, do not listen. Word error rate on your recordings, time to first audio at p50 and p95, barge-in handling, transfer success to a human, and correct write-back to your CRM.
5. Days 10 to 12: load-test at your real peak. Your 5th-of-the-month number, not a round 50. Record drops and latency at that concurrency.
6. Days 12 to 14: price the full year. Per-minute or platform fees, model bills, numbers, DLT, and engineering hours for build and maintenance. Put the engineering line in rupees.

The lowest total at your real concurrency, with acceptable error rates on your audio, wins. Often that answer is "both, on different layers."

What changes in the next 12 months

Expect the layers to keep converging. Exotel's MCP Server, once out of beta, moves it further into agent territory, with the customer bringing the model and Exotel supplying execution and telephony. Bolna's numbers, regulated-series guidance and telephony connectors move it further down the stack. Both will claim to be end-to-end by late 2027.

Two things will not converge: fifteen years of carrier relationships and contact-centre depth on one side, and model-agnostic, developer-first speed on the other. Model-layer prices will keep falling as Indian-language speech and voice models compete, which makes BYOK more valuable, and TRAI's push on designated number series keeps the telephony layer a compliance asset rather than a commodity.

The bottom line

"Exotel vs Bolna AI" is the wrong question asked for a good reason. Exotel is India's telephony and contact-centre layer with AI built on top. Bolna is a voice agent orchestration layer that needs telephony underneath and can use Exotel's. If your numbers and agents already live on Exotel, the real choice is whether to build your own bot on AgentStream or put Bolna on top. If you are starting fresh with a small pilot, Bolna's published pricing and setup speed make it the easier start, provided you negotiate concurrency before campaign day. In every case, DLT, number series and conduct rules are your job, and the only comparison that counts is the one you run on your own recordings at your real peak.