Skip to content

Converse

Scheduled Sep 25, 2026

In-product messaging on your own sockets.

Replaces

  • Intercom · Intercom, Inc. $85 per seat per month (Advanced), about $408,000/yr at 400 seats

Rented, per year

$408,000

Intercom list price at 400 seats.

Owned, per year

$7,440

AWS infrastructure only. Azure is $8,520.

Build

10 weeks

2 engineers. To parity on the core features above, including migration of conversation history, saved replies and active tours from Intercom.

Difficulty

4/5

Substantial. Compliance or realtime depth involved.

Reference organisation: 100 agent seats, 300,000 visitor sessions a month, 45,000 conversations a month, 18 months of transcript retention, single region, prod + non-prod. API Gateway WebSockets at $1.00 per million messages and $0.25 per million connection-minutes on AWS; four Web PubSub Standard units on Azure at the rate that applies in the customer's region. 3-year Compute Savings Plan on Fargate / 3-year reservation on Container Apps. Aurora minimum 0.5 ACU. Excludes LLM gateway token cost for AI answering, which at this volume typically runs $200 to $600 a month depending on model. Figures are infrastructure only and exclude the build.

What it is

Converse is in-product messaging: a messenger embedded in your application, live chat with routing and presence, outbound banners, tooltips and product tours, and an AI agent that answers from your own knowledge base before a human is involved. It is aimed at product and support teams past roughly 60 seats, where the messenger has become both a licence line and a resolution-fee line at the same time. Converse charges you for sockets and tokens, which are measurable, rather than for seats and resolutions, which are commercial constructs.

Why this one stops making sense

  • Intercom Advanced lists at $85 per seat per month. A hundred seats is $102,000 a year before a single AI resolution is billed.
  • Fin is charged from $0.99 per resolution. Twenty thousand deflected conversations a year is another $19,800, and the better your knowledge base gets, the more you pay.
  • Per-resolution pricing puts the vendor's revenue and your cost-reduction goal on the same axis, pointing in opposite directions.
  • The messenger runs inside your product, on your customers' sessions, with a third-party script on every page. That is a performance budget and a security review you did not choose.
  • Conversation transcripts and tour completion data describe how your product is actually used. They are product research, and they are stored somewhere you cannot join them to anything else you own.

What ships

  • Embeddable messenger with your design system, loaded as a small first-party bundle
  • Live chat with agent presence, typing indicators, routing rules and office hours
  • AI answering from your Atlas knowledge base with citations and a confidence threshold for handover
  • Conversation handover to a human that carries the full transcript and the AI's reasoning trail
  • Outbound messages: banners, posts, tooltips, checklists and surveys with audience targeting
  • Product tours with step conditions, branching and completion tracking
  • Visitor identification and session context: plan, role, feature flags, last error
  • Conversation history unified with Frontdesk tickets against the same contact
  • Mobile SDK hooks for in-application messaging on iOS and Android
  • Full transcript export and event stream into your own warehouse

Data model

VisitorContactConversationMessageAssignmentAgentRoutingRuleOutboundMessageTourTourStepTourCompletionAiAnswerCitationAuditEvent

Stack

frontend
Next.js 15 admin plus a standalone messenger bundle under 40 KB gzipped
backend
NestJS on Node 22 — REST for history, a socket layer for live conversation
database
PostgreSQL 16 with conversations partitioned by month
cache
Redis 7 for presence, routing state and socket-to-agent mapping
queue
BullMQ for outbound targeting, email fallback and transcript export
auth
OIDC for agents; signed visitor identity hashes for authenticated end users
search
PostgreSQL tsvector over transcripts; vector search over Atlas content for retrieval
ai
Retrieval-augmented answering with citation enforcement through a pluggable LLM gateway

Parity, honestly

What we match, and what we do not.

The rows marked No are the important ones. If one of them is the reason you bought Intercom, keep buying it — and we will tell you that on the first call rather than the fifth month.

CapabilityIntercomOursNote
Embedded messenger and live chatYesYesFirst-party bundle, so it goes through your own performance budget and CSP.
Routing, office hours and agent presenceYesYes
Product tours, tooltips and checklistsYesYes
Outbound banners and targeted messagesYesYes
AI deflection from a knowledge baseYesPartialConverse answers from your Atlas content with citations and a handover threshold. It is a retrieval-augmented answerer using your model of choice, not Fin.
Per-resolution AI pricingYesNoYou pay tokens instead. At current gateway rates a cited answer over your own documentation costs cents, not $0.99 — but you now own answer quality, and that is a real job.
Fin's answer quality out of the boxYesNoFin is tuned on an enormous corpus of resolved support conversations across thousands of companies, and on day one it will outperform what we ship. Ours improves as your knowledge base does, which takes months of deliberate work by a real person. If nobody owns your documentation, do not do this.
App store of prebuilt messenger integrationsYesNoWe build the integrations you use. Intercom's app store is broad and we do not reproduce it.
Vendor-run SOC 2, HIPAA BAA and ISO 27001YesNoConverse inherits your cloud posture. If you need a signed BAA from the messenger vendor, that is a reason to keep buying.
Unlimited free seats for read-only observersNoYesProduct managers watching conversations cost nothing.
Native iOS and Android SDKsYesPartialWe ship a thin SDK over the same API. Intercom's mobile SDKs are mature and ours will not match them in the first release.
Conversation transcripts joined to your own product analyticsPartialYesSame database, so a query joining tour completion to churn is a query, not an integration project.
Real-time typing indicators and read receiptsYesYes
Email fallback when a visitor leaves the pageYesYesHandled through Dispatch or your existing transactional sender.

Reference architecture

Built twice, on purpose.

Every application in Techtons ships with a production architecture for AWS and one for Azure, using the same diagram grammar so you can read them side by side. Your cloud, your account, your bill.

Amazon Web Services

Compute
ECS Fargate for the API and workers; API Gateway WebSocket APIs for visitor sockets
Data
Aurora Serverless v2 PostgreSQL with pgvector, ElastiCache Redis
Storage and edge
S3 for attachments and transcript archives, CloudFront for the messenger bundle
Identity
Cognito for agents federated to the customer IdP; HMAC-signed identity for end users
Observability
CloudWatch and OpenTelemetry; socket connect failures and answer-citation rate as SLOs
Infrastructure as code
Terraform, one module per environment

Services

API Gateway WebSocket APIECS FargateALBAurora Serverless v2ElastiCache RedisS3CloudFrontCognitoSQSEventBridgeLambdaSESSecrets ManagerWAFCloudWatch

The decision worth arguing about

This is the one place in the catalogue where we run WebSockets on API Gateway rather than on a container we own, and the reasoning is the exact inverse of the Cadence deployment. Cadence's sockets belong to named employees who connect at 09:00 and idle all day, which makes API Gateway's $0.25 per million connection-minutes a tax on doing nothing; Converse's sockets belong to anonymous visitors whose median session is under four minutes and whose arrival rate is set by your marketing calendar rather than your headcount. At $1.00 per million messages and $0.25 per million connection-minutes, a busy month of 300,000 visitor sessions costs a few tens of dollars, and no Fargate service sized for the launch-day spike gets anywhere near that. The trade-offs are real and shape the client: API Gateway closes a connection after two hours regardless, and after ten minutes of idle, so the messenger must treat reconnection as normal rather than exceptional and resume from a server-held cursor instead of socket state. The crossover point is worth knowing before you copy this pattern — around 25 to 30 million messages a month the per-message charge overtakes a pair of always-on Fargate tasks holding the same sockets, and at that volume we move the socket layer onto containers and keep the client unchanged, which is why the reconnection logic exists on day one.

Microsoft Azure

Compute
Container Apps for the API and workers; Azure Web PubSub for visitor sockets
Data
PostgreSQL Flexible Server with pgvector, Azure Cache for Redis
Storage and edge
Blob Storage for attachments and archives, Front Door Standard for the messenger bundle
Identity
Entra ID for agents; HMAC-signed identity for end users
Observability
Azure Monitor and Application Insights
Infrastructure as code
Bicep, azd-compatible layout

Services

Web PubSubContainer AppsFront Door StandardPostgreSQL Flexible ServerCache for RedisBlob StorageEntra IDKey VaultService BusEvent GridCommunication ServicesLog Analytics

The decision worth arguing about

Azure Web PubSub bills on a completely different axis and it changes which workloads are cheap. A Standard unit covers 1,000 concurrent connections and includes a daily message allowance of a million messages per unit, so you are buying concurrency rather than traffic — the opposite of API Gateway, where concurrency is nearly free and messages are the meter. For a chatty product this is better: a support conversation with 60 messages costs the same as one with three. For a spiky one it is worse, because units are provisioned in blocks of a thousand connections and scaling is a deployment operation rather than something that happens on its own, so a product launch that puts 4,000 visitors on the page at once needs the units in place beforehand or connections are refused. We provision to the 95th percentile of concurrent visitors plus one unit of headroom, alert when sustained concurrency crosses 80 percent of capacity, and accept paying for idle units overnight — the same money API Gateway would have charged us in connection-minutes, just committed in advance instead of metered. Azure does not publish the per-unit rate on the pricing page itself; it renders through the region selector in the calculator, so we size it per customer rather than quoting a headline number here.

Sources

Every price on this page, with the page we read it from and the date we read it.

ProductPlanList priceCheckedSource
IntercomAdvanced$85 / seat/month2026-09-05www.intercom.com/pricing

Essential lists at $29 and Expert at $132 per seat per month. The Fin AI Agent is billed separately from $0.99 per resolution on every plan.

Should you still be paying for Intercom?

Two weeks, fixed price. We audit what you actually use, map it against Converse, cost the replacement on your own AWS or Azure account, and give you a delivery plan. If the honest answer is to keep the licence, that is what the report will say.