Custom sales and service automation: architecture, timelines and what stays with your team

Automation that works in sales and service combines AI agents and people, connected to the systems the company already uses. NorthStar designs that set custom: it defines what AI does, where a person steps in, how data flows and how the result is measured. The automation that hands the customer to the bot and forgets about them becomes abandonment.

What we defend

In our reading, the most expensive mistake is automating for what the technology does and not for what the customer needs. Gartner predicts that, by 2027, half of the companies that planned to heavily reduce human service will give up on that plan (statement from 10/06/2025). Klarna, which once reported an assistant doing the work of 700 agents, went back to prioritizing human support in 2025, according to tech.co. The two cases point to the same design: digital first, with people where they make a difference.

In sales, the gain has another shape. Salesforce’s State of Sales 2026 heard more than 4,000 professionals: reps spend 60% of their time off selling, and 88% of those who use agents say they help hit the target. Here AI gives hours back to the rep, and the rep keeps the conversation that closes the deal.

What we automate and what we don’t

We automate what is repetitive, has a clear rule and can be checked: first-contact replies, initial qualification, scheduling, order status, CRM updates, conversation summaries and follow-up.

We don’t automate alone what depends on judgment, relationship or risk: negotiation, serious complaint, dispute, strategic customer and any decision that affects the customer without review. In those cases AI prepares and a person decides.

The exact split changes by company, product and ticket. We start with few cases in AI’s column and expand as the customer score and recontact indicate it’s working. We detail the logic in hybrid service: human + AI.

Architecture: agents and people in the same flow

The design we use as a starting point has four layers:

1. Channels, where the customer already is, like WhatsApp, chat and email. The precautions for the most-used channel are in AI on WhatsApp: 10 precautions before automating.
2. AI agents, each with a specific function (service, qualification, scheduling), clear limits and access only to the data it needs.
3. Handoff rules, which define when the conversation goes to a person and carry history and summary along.
4. Data and systems, with CRM, ERP and customer base connected, so the agent consults and updates real information.

Integrations

The agent is only useful if it sees what the company knows about the customer. That’s why the work starts, in many cases, by connecting systems that already exist: WhatsApp, CRM and ERP. The idea is to connect what the company already has, not to swap the whole stack.

Governance and quality

An agent without monitoring gets worse over time, because product, price and policy change. We recommend three practices:

1. Weekly review of a sample of conversations, including the ones AI marked as resolved.
2. A quality metrics dashboard, with recontact, score by AI and by person and cost per resolved case. The detail is in how to measure an AI service agent.
3. Someone on your team accountable for the agent’s content and rules.

On personal data, we design the flow to collect the minimum and provide for human review on decisions that affect the customer, in line with art. 20 of the LGPD. The final legal validation is your team’s or your lawyer’s.

Timelines and format

We work in stages, with a pilot on one channel and one case type before expanding. The pilot exists to measure, not to demonstrate. Each stage’s timelines are defined in the diagnosis, according to volume, channels and integrations.

What stays with your team

• The conversations that require judgment, relationship or risk.
• The agent’s content and rules, with our help at the start.
• The decision to expand, adjust or turn off a flow, based on the metrics.

Whoever handles the hard cases takes on a more qualified role, with fewer repeated questions. That role calls for training and a metric of its own.

Want to design this with us?

Tell us which channel and case type consume the most of your team today. From there, we show what AI takes on, where a person steps in and how to measure the result. Talk to NorthStar on the contact page or get to know sales and service with AI.

Frequently asked

What is custom sales and service automation?

It’s the combination of AI agents and people designed for your process, your channels and your systems, instead of a standard tool applied the same to every company.

Will AI replace my service or sales team?

It’s not what we recommend. Gartner predicts that half of the companies that planned to heavily reduce human service will abandon that plan by 2027, and 95% of service leaders intend to keep human agents. The team’s role shifts to the harder cases.

How long does implementation take?

It depends on the channel, the integration and the volume. The timelines are defined in the diagnosis, and we always work with a pilot before expanding.

Which systems do you integrate?

We work with WhatsApp, CRM and ERP, connecting the systems the company already uses, instead of swapping the whole stack.

How do you know the agent is working?

By recontact, by the score separated for AI and person, by handoff quality and by cost per resolved case, not just by the volume of closed conversations.

Sources

• Gartner. 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI, 10/06/2025. gartner.com
• Salesforce. State of Sales 2026. salesforce.com
• Tech.co. Klarna reverses AI overhaul, 20/05/2025. tech.co
• Presidency of the Republic. Law No. 13,709/2018 (LGPD), art. 20. planalto.gov.br