Superintelligence in Agencies: What Changes in Briefs and Delivery

Superintelligence in Agencies: What Changes in Briefs and Delivery

Superintelligence in agencies is still, in practice, a planning scenario: what changes if AI agents can run most of the research, drafting and analysis on their own. What you can do today is prepare the ground with more structured briefs, review with written criteria and traceable reports, while keeping creative judgment and client data decisions with people.

Key takeaways

• The brief stops being a document that humans interpret and becomes a specification an agent must also be able to read, with an objective, limits and acceptance criteria.
• The more capable the agent, the cheaper variations become and the more expensive it is to choose well. The bottleneck moves from production to review.
• Creative judgment needs to be described as an agency decision with a named owner, not as a step the agent can skip.
• Client data calls for per-account permissions, usage logs and attention to LGPD (Brazil's data protection law), including when the AI tool processes information outside Brazil.
• The safe path is to pilot one real workflow with logs and an agreed metric, and only then decide whether to expand, adjust or stop.

What does superintelligence mean for an agency today?

The term describes a system that would outperform human experts at nearly any intellectual task. That is a research discussion, not a product an agency can buy. What is available are AI agents that use tools, read long documents, follow multi-step instructions and produce material of uneven quality. If you want to keep the concepts straight, read the difference between AGI, ASI and AI agents and what a company can actually run.

Why talk about it at all? Because an agency's planning should not depend on a date. If agent capability keeps growing, the questions that matter are the same ones you face today, only more urgent: who defines what is good, who answers for what was delivered, and what the agent is allowed to touch in client data. An agency that has answered those questions with today's agents is ready to absorb better ones. An agency that has not will find the gaps in public, in the middle of a delivery.

This article separates three things: what vendor documentation says, what is common implementation experience in agent projects, and what is hypothesis. The part about superintelligence itself is hypothesis. The practices around briefs, review and reporting are process engineering, and they work now.

How does the brief change when an agent reads everything?

The traditional agency brief mixes context, desire, tone and constraints in a text that relies on the experience of the reader. A senior professional fills the gaps with repertoire. An agent fills the gaps with the most likely guess, and that is not always what the client wanted.

The central change is to split the brief into blocks that serve different functions:

• Business objective: what the client wants to happen after the piece or campaign, in one sentence.
• Audience and context: who receives it, in which channel, at what moment.
• Tone and creative territory: what is allowed, what is off limits and which references count as examples.
• Approved facts: information the agent may state, with the source. Anything outside this list is a suggestion, not a claim.
• Legal and brand limits: promises that cannot be made, terms that need legal approval, industry rules.
• Acceptance criteria: how the agency and the client will decide that the delivery is good.

This format helps people too. The classic agency friction, the client who says that is not what I meant, almost always comes from acceptance criteria that were never written down. A capable agent simply makes the cost of that gap more visible, because it produces a lot of material quickly in the wrong direction.

One practical point: what goes into the agent's context is a project decision. The context window is limited, and stuffing everything into it lowers quality. Anthropic describes context engineering as carefully choosing what the agent sees at each step, and that applies to briefs: a short approved block is better than a client's entire folder.

Where does creative judgment sit when the agent produces everything?

With more capable agents, production stops being scarce. Ten script variations, twenty headline options and several formats come out in minutes. What stays scarce is knowing which one fits that client, at that moment, without sounding generic.

This changes the role of the creative. Instead of starting from zero, the person does three things: defines the territory before generation, selects and combines after it, and rejects with a stated reason. That reason is valuable, because it becomes reusable criteria. If the creative director rejects a line because it sounds like every ad in the sector, that rule goes into the next brief.

To protect judgment, make explicit what the agent does not decide:

• The core campaign concept and brand positioning.
• Final approval of any piece that goes public.
• Any factual claim about the product, the price or competitors.
• Use of the image, voice or name of real people.

The agent can propose, draft, adapt and check consistency. The decision is recorded under a named owner. This is not distrust of the technology. It is the same split that already exists between whoever writes and whoever approves.

Another risk is flattening. If every agency uses the same models with the same generic requests, deliveries converge. The defense is to feed the process with your own material: house repertoire, examples of approved and rejected pieces, tone rules per client. That is what sets the agency apart, and it is also what needs to be stored carefully.

How do you organize review without it becoming the bottleneck?

When generation gets cheap, the review queue grows. If review stays informal, it becomes the point of failure. A few principles help.

First, layered review. Fact, brand and format checks can be largely automated: the agent compares the piece against the approved facts list, the brand guide and the channel specs, and returns findings. Human review concentrates on what requires judgment: concept, tone, reputation risk.

Second, evaluations of the process itself. Anthropic and OpenAI publish material on evaluating agents with sets of known cases. For an agency this can be simple: collect examples of past briefs with an approved result and rerun the workflow to see whether the output still meets the criteria when something changes, such as the model, the prompt or a client's rules.

Third, a decision log. Every approved piece should have, at minimum, the brief used, the generated version, who reviewed it and what was changed. If a client complains, this lets you reconstruct what happened without relying on memory.

Fourth, be careful with the metric. Measuring how many pieces each person approves per day encourages rushed approval. Read about how to measure AI productivity without turning it into surveillance before you define indicators for the team.

Reports and client data: what to protect and what to record

Reporting is where agents help most and where they can err most quietly. Pulling together data from ad platforms, analytics and CRM and writing a business-language reading is a good task for an agent. The problem appears when the reading sounds convincing but does not show where each number came from.

A trustworthy report has three traits: every number points to its source and period, the interpretation is separate from the data, and there is a field for what the agent could not verify. That last part is often the most useful one for the client.

On client data, a few decisions need to be made before the first use:

• Per-account access: one client's agent does not see another client's data. Scoped authorization with the least privilege possible is recommended practice in guides such as OWASP's guidance on authorization.
• Personal data: lead lists, emails and support conversations are personal data. Define what may go into the AI tool, on what legal basis and under what contract with the client. The ANPD, Brazil's data protection authority, is the official reference for guidance.
• Where processing happens: if the vendor processes data outside Brazil, understand what LGPD requires about international data transfer when using AI.
• Internal policy: the team needs to know what may and may not be pasted into an AI tool. A company policy for ChatGPT use is a good starting point and applies to any assistant.
• Prompt injection: if the agent reads third-party pages, emails or comments, that content may contain malicious instructions. OWASP maintains specific material on preventing prompt injection, and it should guide which tools the agent is allowed to trigger.

The NIST AI Risk Management Framework helps organize this conversation around four verbs: govern, map, measure and manage. It is not a legal requirement for Brazilian agencies, but it works as a neutral outline for talking with more demanding clients.

Illustrative scenario: a regional retail campaign

Imagine a small agency serving a regional store chain. This is an illustrative scenario, with no real client and no results.

Before, the brief arrived through WhatsApp messages and a meeting. The creative team interpreted it and produced pieces, account management sent them to the client, and at the end of the month someone built the report in a spreadsheet.

After the redesign, account management fills in a structured brief with the objective, approved facts (payment terms, dates, promotion restrictions) and acceptance criteria. The agent generates text variations for three channels and checks each claim against the facts list. The creative director selects, adjusts and approves. The reporting agent builds the weekly summary with a source and period on every number and a section for unverified points. A person reads it before it goes out.

The expected gain is not producing more. It is making fewer mistakes in claims and losing less time to rework caused by interpretation. If the pilot does not show this in the agreed metric, the agency stops or adjusts instead of insisting.

A step-by-step plan to redesign the workflow

Juan Carlo's approach for companies follows four moves: map, limit, pilot and hand over. Applied to an agency, it looks like this.

• Map: draw the real path of one type of deliverable, from the client request to approval and reporting. Note where rework happens, where decisions live only in someone's head and what data passes through each step. Record a simple baseline, such as the typical number of review rounds, without inventing precision.
• Limit: choose one workflow, one client willing to take part and a short list of permitted data. Define what the agent can do, what requires human approval and what is out of bounds. Review contracts and privacy notices with the client.
• Pilot: run the real workflow for an agreed period, with logs of input, output, reviewer and changes. Agree the success metric beforehand, for example fewer review rounds with no increase in factual errors. Anthropic's guide to building agents recommends starting with the simplest solution and adding complexity only when needed, which works well as a rule for the pilot. To test without risk, use mock tools and a sandbox before connecting real accounts.
• Hand over: document the standard brief, the review criteria and the test cases, train the team and decide together whether to expand, adjust or end. If the workflow depends on a single person to function, it has not been handed over yet.

If the agency prefers not to run the agent on its own, managed operation can be provided by Be Human (be-human.ai), with monitoring and maintenance. The decision about who operates should come after the pilot, not before.

What should you watch if agent capability keeps rising?

Without betting on dates, you can define signals that the design needs to be revisited:

• The agent starts running long tasks without supervision. Increase the frequency of sampled audits and reinforce the approval points.
• The team starts approving without reading. This indicates volume has exceeded review capacity, and the fix is to reduce generation, not to pressure the reviewer.
• A client asks for direct access to the agency's agents. Define in advance which data and which actions are exposed, and with what logging.
• New regulatory requirements appear. Follow the debate on Brazil's AI regulation and ANPD guidance.

The common thread is that governance must grow together with autonomy. Giving the agent more freedom without expanding logging, review and limits is the shortest path to a client incident.

Frequently asked questions

Does superintelligence already exist for agencies? It does not exist as a product you can buy. What exists are AI agents that are increasingly capable at specific tasks, and sensible planning treats superintelligence as a future scenario, not a current tool.

Can the agent replace the creative director? It can produce and organize material, but choosing the concept, judging tone and taking responsibility for approval remain with the agency. Writing this split down prevents the decision from getting lost as generation volume grows.

Can I put client data into an AI tool? It depends on the contract, the legal basis, the type of data and where the tool processes the information. Start with the minimum necessary, define access per client, log usage and check ANPD guidance and your legal counsel before including personal data.

Where do I start without overhauling the whole agency? Choose one recurring workflow, such as a client's monthly report, map the steps, limit the data and run a pilot with logging and an agreed metric. At the end, decide whether to expand, adjust or end, based on what the logs show.

Sources

• NIST: AI Risk Management Framework
• Anthropic: Building effective agents
• Anthropic: Effective context engineering for AI agents
• OWASP: prompt injection prevention