Superintelligence for Contract and Subscription Renewals

A superintelligent agent for contract and subscription renewals, in practice, is an AI agent with a narrow scope. It reads the contracts and subscriptions you have registered, warns you early about expiry dates and price increases, prepares a draft message and hands everything to the owner to decide. It does not renew, cancel or negotiate on its own. The value is that nobody discovers a renewal after the notice deadline has already passed.

Key takeaways

• The agent monitors dates, price increase clauses and notice deadlines, and alerts the owner with enough time to negotiate.
• It drafts messages and summaries, but the decision to renew, renegotiate or end a contract always stays with a person.
• Quality depends more on how contracts are registered and on the alert rules than on which AI model you choose.
• Limiting data and actions from the start reduces risk around privacy law (LGPD), misread clauses and messages sent without review.
• The safe path is a pilot with one contract type, logs for every alert and a success metric agreed in advance.

What does a renewals agent actually do?

The word superintelligence tends to create the wrong expectations. What companies actually run today are AI agents that combine fixed rules with text reading and draft generation. If you want to understand that difference calmly, read the article on the difference between AGI, ASI and AI agents. We use the term here because it is how many people search for the topic, but the design is a follow-up assistant, not a system that decides for the company.

In the renewals routine, the agent performs four functions:

• Read each contract and extract the information that matters: parties, term, renewal mechanism, notice period, price increase rule and internal owner.
• Calculate and track dates, triggering alerts at defined milestones.
• Gather the context for the decision: usage history, open complaints, current price versus adjusted price, known alternatives.
• Prepare a draft message to the supplier or customer, for human review.

What it does not do: sign, accept a price increase, pay, cancel or send communication without approval. These limits are part of the project design, not a wish list for later.

What data does the agent need, and what should stay out?

Most renewal problems start with scattered information: the contract as a PDF in a folder, the date in someone's head, the price adjustment in an outdated spreadsheet. Before any AI, you need a minimal and reliable register. For each contract or subscription, the agent needs:

• Identification: supplier or customer, subject matter, internal number.
• Dates: start, end of term, and the deadline for giving notice of non-renewal.
• Renewal condition: automatic, upon acceptance or subject to negotiation.
• Price increase rule: the index, how often it applies and in which month, as written in the contract.
• Owner: who decides and who is the backup.
• Notes: open issues, history of problems, alternatives under evaluation.

What stays out matters too. Full banking details, personal documents, passwords and sensitive personal information do not need to pass through the agent for it to warn you about an expiry date. When a contract involves an individual, there is personal data and LGPD applies: a clear purpose, minimal access and a record of who consulted what. ANPD, Brazil's data protection authority, is the official source for guidance on the subject. If your AI provider processes data outside Brazil, the article on international data transfer under LGPD when using AI providers helps you structure that conversation.

One technical point: the text of contracts and of emails from third parties is untrusted input. A document can contain instructions that try to divert the agent's behavior. For that reason, an agent that reads contracts should not have permission to act on its own. OWASP's guidance on preventing prompt injection describes this risk and the basic defenses.

How to design early expiry alerts

A good alert is one that arrives while you can still act. The common mistake is warning on the eve of expiry, when the notice deadline for non-renewal has already passed. The right design starts from the decision deadline, not from the contract end date.

An alert scheme built on milestones, in plain terms:

• Planning milestone: a long-lead notice, to review whether the contract still makes sense and to collect usage data.
• Negotiation milestone: an intermediate notice, with a summary of terms, the expected price increase and a draft message.
• Decision milestone: a short notice, ahead of the contract's notice deadline, asking for a recorded decision.
• Escalation milestone: if nobody has responded, the alert goes up to the backup or to management.

The interval for each milestone varies with the type of contract and with how long your company takes to negotiate. Set them with the people who run the process, not with a generic number.

The alert itself should be short and actionable: which contract, what the deadline is, what is at stake, what the agent has already prepared and what decision is expected. For the channel, email and an internal tool are usually enough. If your team lives on WhatsApp, the internal notice can go there, as long as the number and the flow are properly set up. To choose the channel, the article on SMS vs RCS vs WhatsApp for customer notifications lays out the criteria.

How to handle price increases without letting the agent decide

Price increases are where money leaks out quietly. The contract provides for an index, a frequency and sometimes a cap or a negotiation. The supplier sends a letter or an invoice with the new amount, and nobody checks it.

The agent helps in three places, always with human verification:

• Anticipate: warn that the adjustment month is approaching, citing the clause and the original contract passage.
• Check: compare the amount charged with the amount expected under the rule, flagging any gap. The calculation should come from a deterministic rule (a formula and an index supplied by a person or an official source), not from a language model's estimate.
• Contextualize: gather what supports the negotiation, such as real use of the service, incidents and alternatives already identified.

The separation matters: text and summaries are AI work; arithmetic is fixed-rule work. This follows a principle from Anthropic's guide to building effective agents: use simple, predictable workflows where they are enough, and keep the model's judgment for where it adds value.

When the contract is with an end consumer and the agent could, for example, classify the customer to decide on a condition, take extra care. Decisions that affect people and are made only by automated processing carry a right of review under LGPD. The article on automated decisions and Article 20 of LGPD explains what to watch.

Draft messages: what the agent writes and what the owner approves

The draft is the most visible day-to-day gain. Instead of starting from scratch, the owner receives a ready message for three situations: asking for a renewal proposal, disputing a price increase that does not match the contract, and giving notice of non-renewal within the deadline.

Good practices for these drafts:

• The agent cites the contract, the clause and the date, without inventing conditions that are not in the register.
• The tone follows a short company guide (formal, direct, no promises).
• The draft never states a decision as made: it uses phrases such as "we would like to evaluate" until the owner approves.
• Sending is done by the owner or released by them, with a record of who approved and when.

If the message goes to a customer and is sent over WhatsApp, channel rules apply: contact permission, approved templates and limits on outbound messages. The article on WhatsApp outbound messages, templates and opt-in explains what the company needs in order before automating this kind of notice.

Illustrative scenario

Illustrative scenario: a services company has a management software subscription, an equipment maintenance contract and a rented office space. Each one is looked after by a different person.

The agent reads the three registered contracts and identifies that the software subscription renews automatically and requires advance notice to cancel. At the planning milestone, it alerts the finance owner and attaches a summary: term, price increase clause and notice deadline. At the negotiation milestone, it delivers a draft asking for the renewal terms and questioning a module the team barely uses.

The supplier sends the new invoice. The agent compares it with the contract rule and flags a difference. The owner reviews the cited passage, adjusts the dispute draft and sends it personally. The final decision to renew with an adjustment is recorded with a date and a name. No step was taken by the agent without approval.

A step-by-step plan to put it into operation

The working method we use in this kind of project follows four moves: map, limit, pilot and hand over. Applied to renewals, it looks like this:

• Map: find out where the contracts are today, who decides each type, how notices work and which renewals have already caused problems. This is the baseline for measuring later.
• Limit: define which fields the agent reads, which sources it consults, which actions are prohibited (send, accept, cancel, pay) and who approves each message. Apply the principle of least privilege, as described in OWASP's authorization guidance.
• Pilot: choose a single contract type (for example, software subscriptions), run it with logs of every alert, every draft and every decision, and agree in advance on a success metric, such as renewals handled within the notice deadline.
• Hand over: deliver the operation with documentation, clear owners and a criterion for expanding, adjusting or stopping. If the company prefers, ongoing operation can be run with monitoring and maintenance.

In parallel, review performance with real examples: cases where the agent got the clause right, got the deadline wrong or proposed an unsuitable tone. The NIST AI Risk Management Framework offers a useful vocabulary for organizing risks, responsibilities and follow-up during use. To see how this follow-up fits into a larger architecture, the article on the company superintelligence layer in five building blocks gives the overall picture.

When an agent is not worth it

Not every company needs this. If you have few contracts, all with fixed dates and an organized owner, a calendar with well-configured reminders solves it. The agent starts to make sense when there is volume, contracts with different price increase rules, several owners and a history of missed deadlines. An honest question before you start: is the problem a lack of organized information or a lack of a decision routine? AI helps in the first case and does little in the second.

Frequently asked questions

Can the agent renew or cancel contracts on its own? It is not recommended. Renewing, accepting a price increase or canceling creates legal and financial obligations, and misreading a clause has a real cost. The safe design keeps those actions with the owner, and the agent only prepares information and drafts.

Does the agent need to read the whole contract? Generally, no. It needs the fields that support the decision: term, renewal, notice period, price increase and owner. Extracting those data once, checking them with a person and keeping a reference to the original passage reduces errors and limits the amount of data processed.

How do you stop the agent from inventing a clause? Separate rules from text. Dates and calculations come from registered fields and fixed formulas, and each summary should cite the contract passage it came from. If the agent cannot find the information, the instruction should be to flag the gap for review instead of filling it in.

Where should you start if the company has not organized its contracts? Start with the register: pick one contract type, gather the documents and fill in the minimal fields with an owner. That survey already delivers value on its own and is the base for any pilot. Only then does it make sense to connect alerts and drafts.

Sources

• Anthropic: Building effective agents
• NIST AI Risk Management Framework
• OWASP: prompt injection prevention
• ANPD: Autoridade Nacional de Proteção de Dados