AI agents for business: how to choose the right model for clinics, real estate agencies, and e-commerce

AI agents for business: how to choose the right model for clinics, real estate agencies, and e-commerce

Choosing the right AI agent for your company comes down to three variables: the volume of leads you need to process, the system of record you already use (electronic health records, real estate CRM, or e-commerce ERP), and how much autonomy you are willing to delegate. The language model (GPT-4, Claude, Gemini) matters less than the integration architecture and the business logic built into the agent.

TL;DR: Clinics need agents that verify insurance coverage and fit appointments into the schedule; real estate agencies need qualification by budget and location; e-commerce needs order tracking and cart recovery. The technical choice comes after you define the flow.

Most articles about AI agents for business focus on which LLM to use. But in practice, anyone who builds custom agents knows the bottleneck lies in modeling the decision flow, integrating with the existing data, and governing autonomy (what the agent can solve on its own versus what gets escalated to a human).

We will break down the selection criteria by segment (clinics, real estate agencies, and e-commerce) and build a decision matrix that works for any operation with more than 200 leads per month.

Why your segment defines the type of agent, not the language model

An AI agent for clinics needs to verify insurance coverage and doctor availability and, in many cases, understand procedure acronyms (SADT, TISS forms, used in Brazilian health insurance). It talks to health record systems such as Simples Dental, MV, or Tasy.

An AI agent for real estate agencies qualifies leads by budget, location, and property type (sale, rental, new development). It integrates with CRMs such as Kenlo, Vista, or RD Station and, ideally, checks availability in the property database in real time.

AI agents for e-commerce, on the other hand, track orders, recover abandoned carts, suggest complementary products, and handle exchanges and returns. They talk to Shopify, VTEX, Nuvemshop, or ERPs such as Bling and Tiny.

Each agent's architecture is different because the decision context is different. There is no "configurable generic agent" that works well for all three; that promise usually delivers a chatbot with a beefed-up FAQ, not an autonomous agent.

Decision matrix: lead volume × legacy system × desired autonomy

Before choosing a vendor or a model, map where your operation fits:

If your monthly volume is below 200 leads, a well-configured chatbot with a decision tree is usually cheaper and more predictable than a generative agent. Above 500 leads per month, the time your team saves justifies investing in an autonomous agent.

The role of the legacy system

Integration with the existing system accounts for 70% of the implementation effort. If the clinic uses a health record system that does not expose an API (or charges a lot for integration), the agent is limited to sending data by email or spreadsheet, which cancels out part of the automation gain.

Real estate agencies that keep their property database in Google Sheets can integrate easily through the Sheets API. Portals with legacy proprietary systems, however, need an intermediate layer (webhook, Zapier, n8n, or custom middleware).

E-commerce stores on SaaS platforms (Shopify, VTEX, Nuvemshop) have mature, well-documented APIs. Stores on WooCommerce or custom platforms require a dedicated endpoint to be developed.

Rule of thumb: if your current system does not have a documented REST API or webhook, set aside 40% of the project budget for the integration layer alone.

How to choose the agent's level of autonomy

Autonomy is not binary. An agent can be configured in three layers:

1. Informational: answers questions, sends links, collects data, and hands off to a human. Low risk, moderate operational gain.
2. Transactional: books appointments, qualifies leads, sends proposals, updates status in the CRM. Medium risk (booking errors, wrong data in the system), high operational gain.
3. Decision-making: approves credit, issues refunds, cancels appointments, negotiates terms. High risk, strategic gain (frees the team for high-value relationships).

Clinics usually start at level 2 (booking + confirmation). Real estate agencies at level 2 (qualification + sending listings). E-commerce can go straight to level 2 or 3, depending on the average order value; stores with an average order above R$ 500 prefer human validation before a refund.

AI consulting helps map the acceptable level of autonomy by looking at the customer's real journey, not at what "would be ideal." An agent that escalates everything to a human for fear of making mistakes is not an agent; it is a form with a voice.

What to watch during the pilot

Run a 30-day pilot and monitor:

• Flow completion rate (how many leads reach the end without human intervention)
• Escalation rate (how many need a human partway through)
• Average resolution time (from the start of the conversation to the booking, proposal sent, or question answered)
• Stated satisfaction (NPS or a direct question at the end of the conversation)

If the completion rate is below 60%, the flow is poorly designed or the knowledge base is incomplete. If the escalation rate exceeds 30%, the agent is underpowered (it should have more autonomy) or the problem is too complex to automate.

Comparison: custom agent versus no-code platform

No-code platforms (Landbot, ManyChat, Typebot, Botpress) work well for linear flows: FAQs, lead capture, multiple-choice qualification. The limit shows up when you need complex conditional logic, queries to multiple APIs, or personalization based on customer history.

A custom agent, built in Python, Node.js, or a framework such as LangChain/LlamaIndex, enables:

• Persistent context: the agent remembers previous interactions (e.g., "ma'am, you called last week about the property in Morumbi")
• Orchestration of multiple sources: queries the property database + financing API + CRM in parallel
• Smart fallback: if the health record API goes down, the agent lets the lead know and offers the front desk's direct contact
• Fine-tuning of tone and vocabulary: oncology clinics need empathetic language, luxury real estate needs an aspirational tone, fashion e-commerce needs calculated informality

A custom agent costs between 3x and 10x as much as a no-code platform, but the conversion rate (lead → booking/proposal) is usually 40–80% higher, according to internal data from projects Agência Rollin followed between 2025 and 2026.

When to go no-code: volume below 500 leads per month, a simple flow, an implementation budget below R$ 8 thousand.

When to go custom: volume above 1,000 leads per month, a need for deep integration, a requirement for full control over data (LGPD, HIPAA for healthcare in the US).

Real cases by segment

Physical therapy clinic (450 appointments/month)

Before: the receptionist handled WhatsApp 10 hours a day, the no-show rate was 28%, and rescheduling was manual.

After the agent: automatic confirmation 48 hours in advance via WhatsApp, autonomous rescheduling if the patient gives notice at least 24 hours ahead, and the no-show rate dropped to 11%. The receptionist now focuses on first-time visits and complex cases (denied coverage, special authorizations).

Technical choice: transactional agent (level 2), integrated with the clinic's scheduling system via REST API, running on a dedicated instance. LLM used: GPT-4o-mini for cost and latency, with a human fallback configured for times outside the available schedule.

Regional real estate agency (280 leads/month, 12 agents)

Before: leads from Facebook Ads landed in a spreadsheet, agents called 48–72 hours later, and the lead → visit conversion rate was 8%.

After the agent: qualification within 5 minutes via WhatsApp (budget, neighborhood, type), automatic delivery of 3 matching properties with photos and a 360° tour, and visit booking with an available agent. The lead → visit conversion rate rose to 19%.

Technical choice: transactional agent (level 2), integrated with the Kenlo CRM and a property database in Sheets (migration to a native CRM planned for the second half of the year). LLM: Claude Sonnet 3.5 for its ability to follow message formatting instructions (each property always sent as a card with image + link).

Electronics e-commerce store (8,200 orders/month)

Before: 34% of support tickets were "where is my order," the average response time was 6 hours (business hours), and support NPS was 42.

After the agent: automatic tracking via WhatsApp (the customer sends the order number, the agent queries the Correios/carrier API and returns the status + estimated delivery), abandoned cart recovery with progressive discounts (5% after 2 hours, 10% after 24 hours), and accessory recommendations after the sale. 71% of tracking tickets resolved without a human, and NPS rose to 68.

Technical choice: transactional + partial decision-making agent (level 2.5: it can offer discounts of up to 15% on its own and escalates anything above that). Integrated with VTEX via the GraphQL API. LLM: GPT-4o (full version) for its better accuracy in extracting order numbers from poorly formatted messages ("hi my package disappeared code 5RT890009BR").

Key takeaways

• Choosing an agent starts with mapping the real decision flow, not with choosing the LLM.
• Monthly volume above 500 leads justifies a custom agent; below that, a chatbot or no-code platform is usually enough.
• Integration with the legacy system takes up 40–60% of the implementation effort; confirm that an API exists before promising deadlines.
• Autonomy should be configured in layers (informational → transactional → decision-making) and adjusted after a 30-day pilot.
• Customer satisfaction with an agent usually beats that of an overloaded human team, as long as the flow completion rate stays above 65%.

What to consider before hiring

Before choosing a vendor, be clear about:

1. How many leads your operation handles per month (WhatsApp + chat + phone).
2. Which system you already use and whether it exposes an API (ask your current vendor for the documentation).
3. How much of the operation you are willing to automate: list 5 tasks that, if the agent got them wrong, would cause unacceptable reputational or financial damage. Everything else can be automated.
4. How much human time you want to free up: if the goal is to "reduce the team's workload," calculate the current cost per hour and compare it with the investment in the agent.

Agência Rollin offers a free feasibility analysis: you share your current flow (it can be a spreadsheet screenshot, a process description, or a CRM export), and we map where the agent fits and how much time and money it saves. No generic slide deck, just a recommendation based on your real scenario.

If the volume justifies it, the next step is to prototype the critical flow (booking, qualification, or tracking) in 7–10 days and run a controlled pilot with 10% of the traffic. It is the fastest way to confirm that the technical choice matches the real operation before scaling to 100% of leads.

Want to know whether your operation justifies an AI agent? Send a message to Agência Rollin on WhatsApp. We will run a quick analysis of your current flow and show you where the agent would fit, with no strings attached. We do not sell off-the-shelf solutions; we design the right model for your volume and system before writing the first line of code.