
AI for Companies Without an IT Team: What to Hire and What to Own
A company without an IT team needs five capabilities to keep AI agents under control: scope, access, logs, testing and upkeep. Here is what to hire.
Want to apply this in your company?
Book a call with Juan →A small company without an IT team does not need to build an engineering department to use AI agents. It does need, either inside the company or through a trusted provider, the capabilities that keep the agent under control once it goes live. Those capabilities are scope definition, access control, decision logs, testing before launch and ongoing maintenance. Companies that hire only the build and leave these parts without an owner usually find the gap when a customer gets a wrong answer and nobody can explain why.
Key takeaways
- Every agent needs a business owner with the authority to say what it may and may not do.
- Access to data and tools should follow the minimum necessary, with read and write permissions kept separate.
- Each conversation and each relevant action needs a record that a person can review.
- Tests with real cases and failure cases should run before any customer talks to the agent.
- Maintenance is a recurring job, because instructions, integrations and rules change and someone has to keep up.
What having no IT team really means once an agent is live
Having few technical people does not mean the company can ignore technical questions. It means those questions have to be handled by someone outside the company or by an operations person who has been given a clear role. A customer service agent on WhatsApp, for example, touches customer data, scheduling, inventory and conversation history. Each of these is a point of risk, even when the setup looks simple.
It also helps to separate the language model from the design of the operation. The model is one component. What makes the agent useful and safe is everything around it: what information it receives, which answers it may give, when it hands the conversation to a person and what gets recorded. That part is largely a business decision, and it is where the owner or operations leader needs to be present.
Which capabilities need an owner
Before deciding what to hire, list the capabilities that support any agent in production. If one of them has no named person responsible, the agent is out of control even if it was delivered working.
- Scope: a written description of what the agent handles, what it declines and when it passes the conversation to a person.
- Access: the agent's own credentials, with minimum permissions and no access to whole spreadsheets or full systems.
- Logs: a history of conversations, tool calls and handoffs, kept for a defined retention period.
- Evaluation: a set of real cases to test before launch and whenever a rule changes.
- Maintenance: someone who updates instructions, responds to incidents and decides when to expand, adjust or stop.
When to hire out and when to keep it in-house
No single split works for every company. A practical test is to ask who makes the decision and who feels the impact of it. Business rules, tone of voice, commercial terms and return policies should stay with someone in the company. Technical integration, monitoring and component updates are usually safer with a specialized provider, as long as the contract is clear about who is responsible for each part.
- Keep in-house: defining the scope, approving sensitive answers, handling human handoffs and reviewing logs on a regular schedule.
- Hire out: architecture design, integration with the official WhatsApp API and with company systems, building the tests, monitoring and responding to failures.
- Handle with care: personal data. The company remains responsible under Brazil's LGPD, so any provider contract must cover the purpose, the retention period and where the data is stored.
When maintenance is contracted as a service, it should include a written review routine and an incident channel with an agreed response time. This is the kind of work Be Human delivers as a managed operation.
How to build control without a large technical team
Control does not require a large team. It requires simple rules applied every time. The first is to limit actions: the agent can check availability but cannot cancel an order without human confirmation. The second is to log with judgment: keep what was asked, which tool was used, what answer went out and whether there was a handoff. The third is to have an escalation path that works outside business hours, with a named person responsible for follow-up.
The risk of manipulation belongs in this calculation too. Customer messages can try to make the agent ignore its instructions, which OWASP classifies as prompt injection. That threat is not solved by a warning in the instruction text. It is reduced when the agent lacks the power to cause harm beyond what the company accepts, which is why the minimum access rule is the most concrete defense a manager can require. For the WhatsApp service window, it is also worth reading about the 24 hour WhatsApp window and how an AI agent should act, which explains when the agent can reply freely and when it depends on a template.
Checklist before the agent talks to customers
- The scope is written on one page and approved by the person accountable for the results.
- Each connected tool has its own credential and the minimum permission for its task.
- There is a list of real cases, including out-of-scope questions and manipulation attempts.
- The handoff to a person works end to end and has a named owner.
- Logs are accessible for review, have a defined retention period and do not expose unnecessary data.
- There is an agreed process for when the agent gets something wrong with a customer, including who corrects it and who informs the customer.
- Someone has a scheduled date to review the logs and decide the next adjustments.
Before launch, it is also useful to test the flow in an isolated environment. This is covered in our guide on testing AI agents with mocks and a sandbox.
Illustrative scenario: a building materials store
Illustrative scenario: a building materials store wants an agent on WhatsApp to check stock and state the standard delivery lead time. The owner decides that the agent may report availability and lead time, but may not negotiate discounts or promise delivery on a specific date. When a customer asks for a quote for a construction job, the conversation goes to the sales team with a summary of what was already asked. During the first weeks, the responsible person reviews the conversations that ended in a handoff and adjusts the instructions where the agent hesitated. The decision to extend the agent to orders waits until those logs show it is ready, rather than relying on the impression of whoever configured it.
In this example, the company did not need a programmer on payroll. It needed an owner for the scope, a person responsible for handoffs and a provider to handle integration and testing. That arrangement is what most small companies can realistically sustain.
How we apply the method in practice
Our method has four steps: map the current process and its baseline, limit data and actions, pilot a real flow with logs and a success metric agreed in advance, and then transfer the operation and decide whether to expand, adjust or stop. The capabilities described in this article show up mostly in the limiting and transfer steps. We do not promise fixed results, because the pilot exists so that the decision is made with data from the company's own operation. For a more grounded view of what an agent does today, see what an agent does in practice, without the hype.
Frequently asked questions
Do I need a developer on the team to maintain an AI agent? Not necessarily. An operations person can manage the scope, review the logs and handle handoffs, as long as they have the authority to request changes. The technical side, such as integrations and monitoring, can stay with a provider whose responsibilities are written into the contract.
Can we use the team's ChatGPT account as a WhatsApp customer service agent? A general-purpose account was not designed to serve customers with limits, logs and handoffs. For customer service you need a dedicated flow, with controlled access to data and clear rules about what the agent may answer. An internal policy on how the team uses AI tools is a separate and complementary step.
Who is responsible if the agent gives a customer wrong information? The company is, which is why the process has to be agreed before any incident. The process should state who corrects the information, who contacts the customer and who analyzes the cause in the logs. The LGPD also covers automated decisions and the right to review them, a topic explained in our article on LGPD automated decisions and Article 20.
When is it worth expanding the agent to new tasks? It is worth it when the logs from the current flow show that answers are correct, that handoffs arrive well and that the responsible person can review the work without excessive effort. Before that, expansion adds risk without evidence that the operation is ready. If the logs show recurring problems, the right path is to adjust or even stop, and that is also a valid decision.