AI Consulting Firm Red Flags to Check Before You Sign a Contract

AI Consulting Firm Red Flags to Check Before You Sign a Contract

Guaranteed results, no baseline and no exit criteria are clear reasons to hold off signing an AI consulting contract. Here is what to check first.

•

Want to apply this in your company?

Book a call with Juan →

An AI consulting firm should raise doubts when it promises guaranteed results, when it cannot describe how your process works today, and when it does not define what would make the project stop. These three points show up before any technology choice and are visible in the first conversation. For the owner or operations leader, turning down a proposal like this is a management decision, not resistance to innovation.

Key takeaways

  • A promise of guaranteed results, with no conditions, data or time frame defined, is the first warning sign.
  • If nobody has measured the current process, there is no way to prove that AI improved anything.
  • A project with no exit criteria tends to grow by inertia, even when it delivers no value.
  • A scope that is too broad, such as "put AI in the entire customer service," often hides a lack of diagnosis.
  • A provider who does not talk about limits, human review and personal data has not yet thought through your operation.

Why a guaranteed result is the first red flag

Results depend on volume, data quality, business rules, and how customers and your team behave. An honest provider can estimate an expected gain and explain the assumptions behind it, but cannot guarantee that gain before measuring the real process. When a proposal sounds like a guarantee, one of three things is usually happening: the provider does not know your process, it is selling a ready-made tool as if that tool solves everything, or it is shifting to you a risk that both sides should share.

Some phrases deserve special attention: "the AI will solve your customer service," "just connect it and start," "in a few weeks your operation will be automated." None of them says what the inputs are, what outputs are acceptable, or who answers when the agent makes a mistake. Ask the provider to write, in plain language, what it considers success and what it considers failure. If the answer is vague, the contract will be vague too.

No baseline: how to know there is a before to compare against

A baseline is a snapshot of the process before AI: how long a step takes, how many messages arrive per day, how many are answered without human intervention, how many typing or registration errors occur, and how many team hours each customer interaction consumes. Without that reference, any result presented later is an opinion. The problem is that the opinion usually favors whoever sold the project.

The practical question is: "How will we know, two months from now, whether the project worked?" If the consultant cannot answer with a metric, a data source and an owner, the baseline does not exist yet. That does not stop you from starting, but measuring must be the first paid piece of work, not something assumed later.

A good sign is a consultant asking for real samples: examples of conversations, order spreadsheets, support tickets, response deadlines. A bad sign is starting with the choice of model or software before looking at those records.

No exit criteria: a contract that does not know how to end

Exit criteria are the written conditions that define what happens at the end of a pilot: expand, adjust or stop. They protect both parties. Without them, the project continues because nobody wants to admit it did not work, and the operation starts depending on a tool that nobody knows how to evaluate.

Before signing, check whether the proposal answers these questions: which metric decides whether the project continues, what minimum value of that metric is acceptable, who reviews the logs, how long the pilot runs, what happens to the data at the end, and how the agent is switched off without affecting the rest of customer service. If any of these answers is missing, the risk of being stuck with a poorly evaluated solution goes up.

The criteria also need to be realistic. A WhatsApp customer service pilot might use as its criterion the share of conversations closed without an improper handoff to a human, combined with a weekly review of wrong answers. Vanity metrics, such as the number of messages processed, do not show whether the business actually improved.

Other signs that often appear together

Beyond the three core points, watch for signs that the provider lacks maturity. One is the absence of any discussion about personal data. If the proposal involves customers, orders, histories or conversations, the consultant must address LGPD, which data will be sent to third parties, and where it will be stored. Anyone who treats this as a technical detail to handle later has not understood the risk.

Another sign is a demonstration built on fictional data and a perfect flow. Demonstrations are useful, but they do not replace a test with the real exceptions of your business. Ask what happens when a customer writes two messages in a row, changes the subject, asks for a refund outside the policy, or requests something the agent should not do.

Also be wary of anyone who does not talk about handing off to a human. An agent without escalation rules is not a complete product. It is a risk with a nice interface. And when the internal team is not involved in the design, the project tends to be abandoned after the first week.

Checklist before you sign

Use this list as a filter in conversations with any provider. If three or more items have no clear answer, postpone the signature.

  • The proposal describes the current process, with its steps, volumes and owners.
  • A measured baseline exists, or the first step of the work is to measure it.
  • There is a success metric and a reference value, agreed in writing.
  • The pilot covers a real, limited flow, not the entire operation at once.
  • The data the agent can access and the actions it cannot take on its own are defined.
  • Decisions and logs are recorded in a way your team can review.
  • There is an exit criterion: expand, adjust or stop, with a date and an owner.
  • The handoff to human service is designed, with clear triggers.
  • LGPD, data retention and use of third parties have been discussed.
  • No part of the proposal depends on a guaranteed result.

Illustrative scenario: a shop that signed without criteria

Illustrative scenario: an accessories store receives a proposal to "automate the entire WhatsApp channel" within a month. The provider promises to reduce the team's workload and does not ask for access to the order spreadsheets or the conversation history. The store signs. After a few weeks, the agent answers simple questions well but gets delivery deadlines wrong and promises exchanges that the policy does not allow. Nobody can tell whether the problem lies in the model, the data or the rules, because no record of earlier service exists to compare against. The contract has no exit route, so the store keeps paying while the team corrects errors by hand.

The same store, with a different approach, would start with a week of measurement: message volume by topic, response time and reasons for complaints. The pilot would be limited to order status questions, with automatic handoff to a human for any exchange request. The criterion for continuing would be set in the first meeting, with a weekly review of the logs.

How the working method responds to these signs

In Juan Carlo's work, the sequence is usually map, limit, pilot and transfer. Mapping means surveying the current process and the baseline before any agent design. Limiting means restricting data and actions to what is necessary, with permissions and human approval wherever the decision is sensitive. Piloting means testing a real flow with logs and an agreed metric. Transferring means handing over the operation with documentation and deciding, based on the numbers, whether the project should grow, be adjusted or stop. This method exists so that the sales conversation does not depend on promises.

If you want to see how this process runs in practice, the page on how Juan Carlo works with companies details the steps. For leadership decisions about priorities, the article on what a founder should decide personally and what not to helps separate what belongs to leadership from what belongs to the technical team.

Before putting an agent in contact with customers, it is also worth testing its behavior with simulated scenarios. The article on mock tools for testing AI agents in a safe sandbox shows how to do this without exposing the live operation.

Frequently asked questions

Is an AI consultant that does not guarantee results a sign of honesty? Not necessarily, but it is a good start. A serious provider talks about hypotheses, metrics and conditions, and accepts that the result will only be known after the process is measured. The absence of a guarantee alone is not enough: what matters is clarity about how the result will be verified.

Do I need a baseline before hiring an AI consultant? You do not need everything ready, but you do need to know what to measure and start measuring. Many companies use the first weeks of the engagement to collect this data, and the proposal should make that explicit. What does not make sense is contracting a pilot with no reference point for the current process.

How do I define exit criteria for an AI pilot? Choose a metric tied to the business, such as the share of requests resolved without an improper handoff or the time to first response, and set a minimum acceptable value in writing. Add a review date, an owner, and the three possible decisions: expand, adjust or stop. If the pilot does not reach the minimum, stopping should also be treated as a valid result.

Should LGPD be discussed before signing with an AI consultant? Yes, whenever the project involves customer data, employee data or conversations. Before contracting, you need to know which data will be accessed, where it will be stored, which third parties are part of the flow and how long it is kept. To go deeper on this, read the article on international data transfer under LGPD when using AI providers.

Sources