Superintelligence for In-House Legal Teams: Contract Triage and Deadlines

Superintelligence in an in-house legal team, in practice, is not a system that decides contracts on its own. What works is a set of AI agents with narrow functions: read documents, extract fields, compare them with an approved template, flag deviations, and feed a deadline tracker. Legal analysis, risk interpretation, and responsibility for the opinion stay with the lawyers.

Key takeaways

• The agent is for triage and organization, not for issuing a final opinion on clauses or risks.
• Before any automation, map how contracts arrive, who reads them, and where the information is recorded.
• Deadlines need a single source of truth, a defined owner for each item, and human review before any external action.
• Contracts contain personal data and confidential information, so limited access, usage logs, and a privacy assessment under LGPD are mandatory.
• The pilot should cover one contract type, one success metric agreed in advance, and a clear decision at the end: expand, adjust, or stop.

What the agent does and what it should not do

In an in-house legal team, the agent works as a routine operator. It receives the draft, identifies the contract type, extracts the key fields (parties, purpose, term, renewal, notice period, jurisdiction, penalties, confidentiality), compares the text with the approved template, and produces a summary that references the clause behind each point. This work is repetitive, easy to forget when done by hand, and simple to verify against the document itself.

What the agent should not do is decide whether a clause is acceptable, recommend signing, draft a reply to the counterparty without review, or label a risk as low. This division matters because the lawyer answers for the opinion. A wrongly extracted field is easy to correct. A recommendation to accept unlimited liability, issued without analysis, is not. For that reason, the workflow design should make explicit what is extraction (verifiable against the text) and what is judgment (always human).

Teams that want to understand the professional limits of AI use in legal practice can start with what lawyers should keep in mind.

How to set up contract triage

Triage depends on two documents that the legal team must approve before any test: the field sheet and the set of acceptable positions for each contract type. Without these references, the agent only summarizes, and a summary does not tell you what is urgent or what falls outside the standard.

The field sheet usually includes:

• Contract type and requesting area
• Parties, corporate documents, and the representatives who sign
• Purpose, term, automatic renewal, and notice period for exit
• Price adjustment method and payment terms, without projections or calculations
• Liability, penalties, confidentiality, and data protection clauses
• Venue and governing law
• Deviations from the approved template, with the quoted passage

Exceptions should go straight to a lawyer, bypassing the standard queue: new contract types, custom negotiations with a key customer, operations involving data transfers outside the country, and documents tied to ongoing litigation. In these cases, the agent can only record that the document was received and routed.

Deadline tracking without relying on anyone's memory

Term dates, renewals, notice periods, recurring obligations, and power of attorney validity often live in emails, old spreadsheets, and conversations. When a contract renews without warning, the problem is rarely technical. It is the lack of a single source of truth and a named owner for each date.

The agent can read new contracts and propose entries for the calendar or the tracking spreadsheet. An assistant or lawyer confirms each entry. Internal alerts should reach two people, so that one person's absence does not break the control. No notice to a supplier or customer should leave without human approval.

There is one point the agent should not resolve alone: counting procedural and administrative deadlines depends on the rules that apply to each case. The system can store the date provided and the counting criterion approved by the legal team, but the counting rule must be validated by whoever has the authority to do so.

A minimum deadline control meets a few conditions:

• Each deadline has an owner, a date, an advance alert period, and a status
• The agent proposes the entry, and a person confirms it
• Internal alerts reach two responsible people
• Date changes are recorded with who changed them and when
• Overdue or unowned deadlines appear in a weekly review

Step by step for a pilot

The pilot should be small enough to review completely and real enough to show where the workflow breaks. The sequence below works as a roadmap:

• Map the current flow: intake channels, owners, time to first reading, and where each version of the contract is stored.
• Choose a frequent contract type with controlled risk, such as standard confidentiality agreements or low-impact suppliers.
• With a responsible lawyer, define the approved template and the field sheet.
• Limit access: the agent reads only the pilot folder, with no access to litigation files or to personal data not needed for triage.
• Build a set of test contracts, including clauses outside the standard, and compare the agent's extraction with human review. For this step, it is worth following how to test agents with mocks and sandboxes.
• Log every run: document, version, extracted fields, deviations flagged, and reviewer.
• Agree on a success metric (for example, how many fields needed correction) and a threshold that forces the pilot to stop, both before starting.
• At the end, decide based on the logs: expand to other contract types, adjust the field sheet, or stop.

Anthropic, in its essay on building effective agents, argues for starting with simple, composable patterns before moving to highly autonomous agents. That advice fits a triage pilot well, because the flow is predictable and human review sits at the right point.

Illustrative scenario

Illustrative scenario: a legal team receives requests for supplier contracts through a form from the procurement area. When a draft arrives, the agent fills in the triage sheet, notices that the renewal is automatic with a shorter notice period than the approved standard, and creates a deadline reminder for two owners. A lawyer reviews the passage, decides to ask the supplier for an adjustment, and records the decision in the log. At no point did the agent contact the supplier. If, after a few weeks, the team finds that the field sheet misses a recurring type of clause, the next step is to adjust the sheet, not to widen the agent's access.

Security, confidentiality and LGPD

Contracts contain data about partners, representatives, witnesses, and sometimes individuals on the counterparty side. Triage must therefore follow basic LGPD principles: a defined purpose, data collection limited to what is necessary, and access restricted to those who need it. Before using a model hosted outside the country, the team should understand where the data is processed. On this point, it is worth reading what LGPD requires for international data transfers when using AI.

There is also a specific risk: the text of a contract can contain written instructions meant to mislead AI systems. Good practice is to treat document content as data, never as a command, and to keep the agent without permission to act outside the defined workflow. OWASP documents this kind of attack and ways to reduce it. The NIST AI Risk Management Framework organizes AI risk management into the functions of govern, map, measure, and manage, which helps legal and technology teams speak a common language.

If triage starts to influence decisions about people, for example in hiring or in evaluating individual suppliers, the issue takes on another dimension. In those situations, review the rules on automated decisions, such as Article 20 of the LGPD.

How to apply the Juan Carlo method

In AI projects for companies, we follow four steps: map, limit, pilot, and transfer. In legal work, that means starting with the current flow and a baseline of time and errors, then defining which data and actions the agent can touch, running one real contract type with logs and an agreed metric, and finally deciding whether to expand, adjust, or stop. The transfer step includes training the team to review what the agent produces, because review is what keeps the analysis with the lawyers. This sequence does not promise a specific result, but it prevents expansion before there is evidence.

If the board still uses the term superintelligence loosely, align the vocabulary before approving any pilot. The superintelligence glossary for executives helps with that conversation.

Frequently asked questions

Can the agent draft or review contract clauses? It can suggest alternative wording for a passage already identified as a deviation, provided a lawyer reviews the suggestion and decides whether to use it. The agent should not be responsible for the final version or negotiate terms with the counterparty. Every suggestion must be logged along with the version of the approved reference template.

Which contract type should I start with? Choose a frequent type with controlled risk and a standardized template, such as confidentiality agreements or low-impact suppliers. Avoid starting with litigation contracts, strategic negotiations, or operations involving international data transfers. The first pilot is meant to test the workflow, not to solve the hardest case in the house.

Can the agent send deadline alerts directly to a customer or supplier? We do not recommend it. Internal alerts to the responsible people in legal and the requesting area are the starting point. Any external communication should go through human approval, partly because it may carry legal effect and must reflect the company's position.

Do I need a new legal system or an expensive tool to start? Not necessarily. A pilot can run on a control spreadsheet, a restricted-access folder, and a workflow in an automation tool, as long as access, execution logs, and human review are defined from the start. Tool choice comes after mapping and setting limits, not before.

Sources

• Anthropic: Building effective agents
• OWASP: LLM Prompt Injection Prevention Cheat Sheet
• NIST AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados