Free first conversation · Reply within 1 business day
Facts behind this service
Contas Casal assistant (own product):
RAG with 10 retrieval tools running in parallel
Mirador, for The Hackett Group:
Prompt-injection defenses and an audit trail
In the same assistant:
Multi-provider, a semantic cache and cost tracked per model
20+years of software engineering; US enterprise clients since 2022
Facts behind this service
Contas Casal assistant (own product): RAG with 10 retrieval tools running in parallel
Mirador, for The Hackett Group: prompt-injection defenses and an audit trail
In the same assistant: multi-provider, a semantic cache and cost tracked per model
20+ years of software engineering; US enterprise clients since 2022
01
Recognize this situation?
The current chatbot only answers the FAQ: any question about an order, an invoice or an account goes to the human queue.
The platform charges per contact or per resolution, and the bill climbs together with support volume.
The bot makes up answers or promises what the company cannot deliver, and nobody measures how many are right or audits what it said.
The customer repeats the order number in every message and again to the human agent, because the bot keeps no history.
02
What we deliver
Agent integrated with your systems
Agent integrated with your systems
The agent reads and updates orders, accounts and tickets in your ERP, CRM and helpdesk, so the request is solved inside the conversation; every action has its own permission.
Answers grounded in your data (RAG)
Answers grounded in your data (RAG)
A knowledge base built from your policies, catalog and history: the agent answers from it, cites the source and says so when it does not know.
Short- and long-term memory
Short- and long-term memory
The agent remembers what the customer already said in the conversation and, across conversations, retrieves orders and preferences stored in your systems, outside the model, so nobody repeats data.
Context engineering
Context engineering
On every turn the agent receives only what the answer needs (rules, customer data, knowledge-base passages, a summarized history): more accurate answers at a lower cost.
Channels through official APIs
Channels through official APIs
The same answer on every channel: web widget, internal tools and messaging channels through official APIs, with one knowledge base and the same rules.
Handoff to humans with context
Handoff to humans with context
The customer repeats nothing: the human agent receives the conversation summary, the data already collected and the reason for the handoff, in the tool your team already uses.
Guardrails and audit trail
Guardrails and audit trail
Controls that keep the agent within your policies: input and output validation, defenses against prompt injection (attempts to manipulate the bot), limits per action and a record of every conversation.
Evals on real tickets
Evals on real tickets
An evaluation dataset of real tickets and edge cases runs on every change of prompt, knowledge base or model; if quality drops, the change does not ship.
Metrics, traces and caching
Metrics, traces and caching
You see resolution, handoff, latency and cost per conversation, with a record of every conversation (trace); prompt caching and a semantic cache cut time and cost on repeated questions.
From document to cited answerAt ingestion, documents become chunks and vectors in a vector store. At query time, hybrid search with reranking picks the right chunks, and the LLM answers with its source, inside the guardrails.
03
Results and experience
Contas Casal
Scope:
Own product · not a client case
No published client case for this service yet: we propose to start with a 2–4 week proof of concept on one support flow and real data. The closest thing we have built is the assistant in Contas Casal, our own product, which answers from the user's account data on structured RAG.
Intent routing and 10 SQL retrieval tools running in parallel
Embedding-based semantic cache and streaming answers
Multi-provider (Anthropic, OpenAI, Gemini, Groq, OpenRouter), with cost per model
~8.8Kautomated tests, with 95% coverage gates
The Hackett Group (NASDAQ: HCKT)
Scope:
Generative-AI security · another domain
Mirador: generative-AI application engine for enterprise clients, not a customer service agent. Its security layer is what an agent needs before it talks to your customers.
Free first conversation · Reply within 1 business day
04
Who it is for and who it is not
AI customer self-service agents are chatbots that answer customers from company documents and data, remember what was already said, read and update orders and accounts in your business and support systems, and hand over to a person when needed. A custom agent pays off when off-the-shelf platforms cannot reach those systems or charge too much per contact.
A good fit for
CX and operations teams with recurring support volume and questions that depend on data from internal systems.
Companies that already use a chatbot platform and hit its integration limits or its per-contact cost.
Operations with their own rules (policies, contracts, deadlines) that a generic bot does not know.
US companies that want a custom support agent built with senior engineering in an overlapping time zone.
Not a fit for
Operations with few conversations a month and simple questions: an off-the-shelf SaaS chatbot does that for less.
Anyone who wants to eliminate the support team: the agent covers the repetitive work and hands the rest to people.
Companies with no knowledge base and no owner to validate the answers the agent gives.
Anyone who needs a full contact center (telephony, dialer, queue management): that is a platform product.
05
How it works and timelines
01
Support discovery
Support discovery
We analyze real conversations, the systems involved and the support rules. You receive a map of the requests the agent can resolve and the ones that need a person, with the plan and the estimate.
02
Knowledge base and evaluation dataset
Knowledge base and evaluation dataset
We organize the sources (policies, catalog, history) and build the evaluation dataset from real tickets, edge cases and expected answers. You approve what the agent may say, do and refuse.
03
Proof of concept
Proof of concept
One end-to-end support flow, on real data and with one integration. You receive the agent in a test environment, the evaluation results and a recommendation to proceed, adjust or stop.
04
Production
Production
Remaining integrations, memory, handoff, guardrails, traces and the metrics dashboard. The rollout is gradual: first a share of the conversations, with human review. You receive the agent in production on the agreed channels.
05
Operate and improve
Operate and improve
We review the traces of handed-off conversations and poorly rated answers, add new cases to the evaluation dataset, tune the knowledge base and prompts and track cost. Operation can continue under a monthly retainer or with your team.
Typical ways to engage this service
Scope
Typical timeline
Engagement model
Support discoveryAnalysis of conversations, systems and rules; map of automatable requests, plan and estimate
1–2 weeks
Fixed-scope project
Proof of conceptOne support flow on real data, one integration and evals on real tickets
2–4 weeks
Fixed-scope project
ProductionAgent integrated with the ERP, CRM or helpdesk, with memory, handoff, guardrails, traces and a metrics dashboard
6–12 weeks
Fixed scope or time & materials with a cap
Operation and continuous improvementConversation and trace reviews, new evaluation cases, knowledge-base and cost tuning
Ongoing
Monthly fractional retainer
Timelines are typical ranges. Discovery defines the scope and the estimate of each project; the contract can be in USD or BRL.
Free first conversation · Reply within 1 business day
06
Technical details
How we use AI with governance, the decision comparison and the stack
AI with governance
In this service the AI talks to your customer. So the rules are explicit: the agent answers only from approved sources, acts within limits, passes evals and every conversation is recorded.
Where AI comes inThe model interprets the message, retrieves the information from your sources and writes the answer; actions on systems go through tools with explicit permission, in a workflow built with LangGraph or Mastra.
Context and memoryOn every turn the agent receives rules, customer data and a summarized history within a token budget; long-term memory lives in your systems, and the same summary goes to the human agent on handoff.
EvalsA dataset of real tickets and edge cases, approved by your team, runs on every change of prompt, knowledge base or model, with automated checks, LLM-as-a-judge and human review of a sample; a regression blocks the release.
Traces and metricsEvery conversation produces a trace (OpenTelemetry, LangSmith or Langfuse) with model calls, tools, tokens and time; the dashboard shows resolution, handoff, latency and cost per conversation.
GuardrailsInput and output validation, sanitized LLM responses, prompt-injection defenses and limits per action; low confidence or a sensitive topic moves the conversation to a person.
Your dataClient data never trains models, and we work under NDA.
Decision comparison
Scroll sideways to see more
SaaS chatbot platform or custom integrated agent: what changes in the decision
Criterion
SaaS chatbot platform
Custom integrated agent
System integration
Ready-made connectors for popular tools; in-house systems depend on what the platform offers
Direct connection to the ERP, CRM, helpdesk and internal APIs, with your business rules
Cost model
Subscription per contact, conversation or resolution; grows with volume
Development project, model usage (tokens) and maintenance; no per-contact fee
Time to first use
Days, with ready-made flows and templates
Weeks: proof of concept in 2 to 4, production in 6 to 12
Control and auditing
Model, prompts and logs defined by the platform
Model, prompts, evals and traces defined and inspected by the company
When it makes sense
Low or medium volume and questions that do not depend on internal systems
High volume, specific rules and answers that depend on data from your systems
Stack and integrations
Claude
OpenAI
Gemini
RAG
MCP
LangGraph
Mastra
pgvector
PostgreSQL
Redis
Langfuse
LangSmith
OpenTelemetry
TypeScript
Node.js
Python (FastAPI)
07
Frequently asked questions about AI Customer Service Agents
How much does a custom AI customer service agent cost?
The cost depends on four factors: how many systems the agent must query, the state of the knowledge base, the number of channels and what it may do on its own. We do not publish a price list. A 1–2 week discovery defines the scope and the estimate. After delivery, the recurring cost is model usage (tokens), which we forecast from conversation volume, reduce with prompt caching and a semantic cache and monitor per model, plus maintenance.
Should we build a custom support agent or buy Intercom or Zendesk AI?
Start with a platform if volume is low and the questions do not depend on your systems: it goes live in days. A custom agent pays off when answers need data from the ERP, CRM or helpdesk, when the operation has its own rules or when per-contact pricing weighs on the budget. In the first conversation, which is free, we tell you which of the two your case calls for, even when the answer is the platform.
Will an AI customer service agent replace our human agents?
No. The agent takes the repetitive, rule-based requests and hands to a person whatever involves an exception, a negotiation or a sensitive topic. The expected effect is a team that spends less time on repeated questions and more on the cases that need judgment. How much of the volume the agent resolves depends on your operation: we measure it in the proof of concept and promise no percentage beforehand.
How does the handoff from the AI support agent to a human agent work?
The agent transfers the conversation when confidence is low, when the customer asks for a person or when the topic is on the list of sensitive subjects. The human agent receives the summary, the data already collected and the reason, in the tool they already use, so the customer repeats nothing. That summary comes from the conversation memory, built with the same context engineering that feeds the agent. Handoff rules are defined with your team and revised from real conversations.
Is AI customer service compatible with data-privacy laws such as LGPD?
It can be, provided the project handles personal data with a legal basis, a clear purpose and minimal access; the legal assessment belongs to your privacy officer or counsel. On our side, we design the agent so the model receives only what it needs, personal data is masked where the use case allows and every access is logged. Customers are told they are talking to an AI assistant and can ask for a person. Client data never trains models, and we work under NDA.
How long until a first version of the AI support agent is ready?
The first testable version comes out of the proof of concept, in 2–4 weeks, after a 1–2 week discovery. It covers one support flow on real data with one integration. The production version, with the remaining integrations, the handoff and the dashboard, takes 6–12 weeks. What weighs most on the timeline is access to the APIs of your systems and the state of the knowledge base.
How do you measure whether the support agent answers well (evals and metrics)?
With evals before every release and metrics in production. The evaluation dataset gathers real tickets and edge cases, including questions the knowledge base cannot answer: there the agent must say it does not know and hand off, and the test checks that. Automated checks, LLM-as-a-judge and human review of a sample score the answers; a regression blocks the release. In production every conversation produces a trace, and the dashboard shows resolution, handoff, latency and cost per conversation.
Which channels can the AI customer service agent run on?
On a web widget, on internal team tools and on messaging channels, always through official APIs. The agent has a single logic: knowledge base, rules and guardrails live in one place, and each channel is an adapter. We start with one channel in the proof of concept and add the others in the production phase, according to the volume of each.
Search, recommendations and agents wired to your catalog.
View service
Let's review your support operation
In the first conversation we look at the volume, the types of request and the systems involved, and tell you plainly whether the case calls for a custom agent or an off-the-shelf platform will do.