Text agents that complete transactions across web and WhatsApp, not just answer questions about them. 70% deflection in production today.
Conversational AI development is the building of text interfaces that handle customer conversations end to end: understanding intent, retrieving the right answer from your content, calling your systems to complete the transaction, and handing over to a human cleanly when it should. AMT builds these for web, WhatsApp and in-product use, grounded in your real documentation and integrated with your live systems.
// the_problem
Every vendor quotes a deflection rate. Almost none of them tell you the query mix it was measured on, which is the only thing that makes the number mean anything. A bot deflecting 80% of "what are your opening hours" is not comparable to one deflecting 45% of billing disputes, and the second is worth considerably more.
Here is the thing. Your deflection rate is mostly decided before anyone builds anything, by what your customers actually contact you about and by how much of that can be resolved without a human touching a system.
// the_honest_numbers
Sort your top twenty query types into these four groups and you will have a realistic estimate before you talk to anyone.
| Query type | Realistic band | Why |
|---|---|---|
| InformationalHours, policy, status, how-to | High | The answer exists in a document. This is retrieval, and retrieval done properly is reliable. |
| Account lookupBalance, order status, usage | High | One read-only API call. Reliable once the integration exists. |
| TransactionalPayments, upgrades, changes, cancellations | Moderate | Needs write access, error handling and rollback. This is where the engineering cost and the real value both sit. |
| Exception handlingComplaints, disputes, edge cases | Low | Should route to a human quickly and well. Chasing deflection here damages the customer relationship for a marginal saving. |
The blended number that appears in a case study is a weighted average of these four. Which is why the honest way to estimate yours is to weight them by your own volume, not to borrow somebody else's headline.
Estimate your own deflection rate
Put in your top query types and volumes and this returns a realistic band, weighted to your mix. It is the figure you will have to defend in a business case, so it should be yours rather than a vendor's.
// what_we_build
// method
// proof
70% of customer queries resolved without a human agent, under 3 seconds average response. One agent handles FAQ and live transactional flows including bill payments, plan upgrades and SIM management, with RAG over policy documents and real-time calls into billing, CRM and provisioning. Deployed across web, WhatsApp and IVR.
Analyses live conversation logs, identifies drop-off nodes and ambiguous intents, and recommends specific flow changes with the reasoning attached. 35% higher flow completion.
// cost
The effort depends on the type of conversations, enterprise integrations, channels, transaction complexity and production requirements.
| Scope | Typical complexity | What drives the effort |
|---|---|---|
| FAQ / knowledge assistant | Low | Content preparation, knowledge ingestion, retrieval quality, conversation design and evaluation. |
| Transactional AI agent | Medium | Typically integrates with 2-3 enterprise systems. Complexity is driven by APIs, authentication, business workflows, error handling, security and transaction recovery. |
| Voice-enabled AI agent | High | Adds telephony or voice-channel integration, real-time STT/TTS, latency management, interruption handling and speech optimisation. See Voice AI Development. |
// voice
If your volume is phone calls rather than chat, the engineering problem is different and it has its own page. See voice AI development.
// the_honest_section
If your query volume is low, the business case will not carry the integration work. Improve your help content first and revisit when volume justifies it.
If most of your contacts are complaints and exceptions, a bot will frustrate people and save very little. Route those to humans faster instead, which is a workflow fix rather than an AI project.
If your systems have no APIs, you will get an FAQ bot rather than a transactional one, and you should price the value accordingly.
// faqs
Software that handles a customer conversation end to end: understanding what they want, retrieving the right answer from your content, calling your systems to complete the task, and escalating to a human with context when it should.
It depends entirely on your query mix. Informational and account-lookup queries deflect well. Transactional queries deflect moderately and cost more to build. Complaints should not be deflected at all. Weight the four groups above by your own volumes.
A chatbot answers questions. An agent completes tasks in your systems. Most useful deployments are somewhere between the two, and the transactional part is where both the cost and the value sit.
Yes, with write access scoped narrowly, idempotent operations so a retry cannot double-charge, rollback paths and full audit logging. Our telecom deployment handles bill payments, plan upgrades and SIM management in production.
Yes. WhatsApp Business is one of the channels our telecom agent runs on, alongside web and IVR.
A transactional agent with two or three integrations typically takes three to six months including discovery. An FAQ-only bot is considerably faster and worth considerably less.
It hands over to a human with the full conversation context attached, so the customer does not repeat themselves. Designing this well matters more to customer satisfaction than the deflection rate does.
Thirty minutes with an engineer. Bring your top ten query types and we will tell you which ones an agent can genuinely close and which ones still need a human.
Not ready to talk? Read how the telecom agent reached 70% deflection on live transactions.