Every week, a proposal to build an AI agent lands on an executive's desk. It comes with a polished slide deck, cost-reduction promises, and a number that sounds too good to be true. Before signing off, ten questions separate investment from loss. This text is the checklist that is missing from most approval meetings.
This is not a technical guide. It is a management roadmap for owners, directors, managers, controllers, or anyone who will approve (or reject) the budget. Each question has an expected answer and a warning sign. If the proposal on your desk does not pass at least 8 out of 10, send it back to the vendor and ask them to redo it. Losing a week now is cheaper than discovering the problem in the fourth month of the project.
1. Is the pain real or is it hype? (questions 1 and 2)
Question 1: what business decision changes when the agent is running? Good answer: "we will go from 18 min to 2 min for first response time in customer support, and that frees up 3 agents for active prospecting." Bad answer: "we will modernize customer service." Modernizing is not a decision, it is a label. If the proposal does not have a clear decision, it is selling a concept, not a solution.
Question 2: how is the problem solved today, without AI? There needs to be a detailed answer ("agent A does X, escalates to supervisor B in case C, and uses spreadsheet D to track"). If the vendor cannot describe the current process clearly, they will build an agent to solve a problem they do not understand. Guaranteed result: the agent will solve something else entirely.
2. Does the scope fit or is it bloated? (questions 3 and 4)
Question 3: how many different processes will this agent cover in the first cycle? Healthy answer: one, or at most two that are very similar. Dangerous answer: four, five, six different processes (support plus collections plus sales plus HR). Bloated scope drowns projects. If the proposed agent does everything for everyone, it will do nothing well for anyone.
Question 4: what is out of scope, and when does it come in? A good proposal lists what it does NOT do and provides a roadmap for future phases. A bad proposal promises the world at a bargain price. Pay special attention to "and the agent can also do X" said verbally, without making it into the contract. That becomes a dispute later.
3. Do you have the right data to feed this? (questions 5 and 6)
Question 5: which systems does the agent need to access, and who on our side guarantees those accesses? Expected answer: a named list of systems (CRM, ERP, BI, e-commerce) with the name of the internal person responsible for each one. If the answer is "we will figure it out later," the project will stall in phase 3, guaranteed, when you discover that the ERP only responds via a manually exported weekly spreadsheet.
Question 6: is the knowledge the agent will use already documented somewhere? Manual, FAQ, video, internal base, old prompt. If nothing is documented, the first month of the project will be spent interviewing your team to extract what is in people's heads. That is real work, it costs time, and it needs to be built into the budget. If the proposal ignores this, they will either charge extra later or deliver something weak.
4. How do you measure the return? (questions 7 and 8)
Question 7: what 3 indicators will we look at to say the project worked? Good examples: resolution rate without human intervention, average response time, post-service NPS. Red flag: "we will measure overall satisfaction." Overall does not translate into decisions. And without decisions, nobody defends the investment in the next budget meeting.
Question 8: what is the current baseline, and what is the target? You need to know where you start and where you are going. If the vendor did not ask for your current data to size the target, they are guessing. Every target without a baseline is wishful thinking. Reject it.
5. Who operates, who approves, who audits? (questions 9 and 10)
Question 9: once delivered, who on our team operates the agent day to day? Name, name, name. If no one in the company will own the product, the agent becomes an orphan after delivery and the consultancy has to stay on forever. If you want a sustainable project, knowledge transfer needs to be planned, with a deadline and a deliverable.
Question 10: how will we know if the agent made a mistake before the customer complains? There needs to be a panel, an alert, and a sample of conversations reviewed by a human on a regular basis. Without that, you only discover the error when it becomes a viral post. If the vendor does not include an audit layer from day one, the risk sits on your company's registration.
6. The budget red flags
Three proposal patterns that should trigger an alarm before you sign.
Pattern 1: fixed price for open scope. "R$ 80 thousand to build your company's AI agent." Nobody can price an AI agent without having defined the process, integrations, and volume. A fixed price on a vague proposal is an invitation to an amendment in month three, a poor delivery, or both.
Pattern 2: ROI promised in month 1. A well-built agent starts delivering measurable results between the second and third month after go-live. Whoever promises ROI in 30 days is selling a dream, or will deliver a cosmetic agent to fake it.
Pattern 3: no exit clause. Who gets what if the project is cancelled halfway through? Is your data yours? Does the code (even when the agent uses a third-party API) have parts that stay with the vendor? Without a clear clause, you are held hostage. More than once, a company has changed vendors only to discover they had to start from scratch.
For deeper context on what to decide before the budget, also see our checklist on when a private AI agent makes sense and the 4 doors AI cannot walk through in operational governance.
Frequently asked questions about building an AI agent for a company
How much does it cost to build an AI agent for a company in 2026?
Typical implementation range: R$ 40 thousand to R$ 200 thousand for a 12 to 16-week cycle covering one process. Monthly operation between R$ 3 thousand and R$ 30 thousand, depending on volume, integrations, and AI model used. Cosmetic projects ("chatbot in disguise") may cost less and deliver less.
What ROI should I expect from building an AI agent?
For well-defined use cases (customer service, qualification, support), ROI in the first year is between 2x and 5x the investment. For poorly scoped projects, ROI is negative. The difference lies in the quality of the approval checklist, not in the vendor.
Is it worth building an AI agent for a small business?
It is, if the monthly volume of repetitive service interactions exceeds a few hundred and the processes have clear rules. For volume below that, the investment does not pay off. It is better to use SaaS AI tools with light customization.
Can I build an AI agent with an internal team?
You can, if you have at least one senior engineer with experience in generative AI, someone who deeply understands the process to be automated, and someone who handles integrations. Without those three, either hire outside help or accept a 6 to 12-month learning cycle before reaching production.
How do I know if an AI agent proposal is serious or just hype?
A serious proposal asks for real data before quoting, describes the current process in detail, lists success indicators with a baseline, has clear exit clauses, and promises measurable results starting from the second or third month. A hype proposal promises vague agility, immediate ROI, and an "agent that learns on its own."
