Artificial intelligence: what moved in September and what to watch in October

Every month brings a new round of announcements about artificial intelligence, and September has been no exception. Models that get updated, tools that change their pricing, features that move from the demo to the work interface, and regulations that advance at different speeds depending on the country. The problem is not the lack of news: it is that most teams do not have the time to separate what changes their day-to-day from what only sounds good in a headline.

This article is not a summary of headlines. It is a reading of the fronts that are shifting the ground and, above all, of what is worth reviewing within your own organization going into October. If something in what follows does not fit your case, that is a good sign: it means there is a pending decision that has not yet been made.

More capable models, and an uncomfortable question about cost

The trend of recent months holds: models are improving in reasoning, in handling long contexts and in their ability to work with extensive documents. Each new version expands what can be asked without breaking the task into ten intermediate steps. For a development team, that translates into fewer back-and-forths; for a business team, into the possibility of loading a complete contract or an annual report and asking specific questions without preparing the text beforehand.

The counterweight is cost. The conversation no longer revolves only around "which model is the best," but around which is the most suitable for each task. A large model to classify emails is money down the drain; a small one to analyze a set of specifications may fall short. The practice that is consolidating is to assign models by use case and measure.

• Define the level of demand for each task: not all of them need the most powerful model.
• Measure consumption by flow, not by user: spending spikes in automated processes, not in chat.
• Set a testing budget: without room to experiment, no one discovers the use case that actually delivers.

Agents: from enthusiasm to oversight

The dominant topic of the month continues to be AI that not only responds but executes: queries systems, drafts, sends, updates records. The difference compared to last year is that the conversation has become more sober. Organizations that have already tried agents have learned that the bottleneck is not the model's capability, but permissions, data and traceability.

An agent that can write to the CRM is an agent that can make mistakes in the CRM. That is why implementations that work share a pattern: limited scope, reversible actions, a record of every step and a person who reviews before anything goes out to a customer. Full autonomy remains, for now, for environments where an error has no external consequences.

If your organization is evaluating an agent, it is worth answering three questions before starting: what actions it can execute without approval, how an error is undone and who is accountable when something fails. If any of the three has no clear answer, the project is not ready for production.

AI in the workplace: real adoption lags behind deployment

Licenses are bought quickly; habits change slowly. It is one of the most repeated patterns of the year: tools available to the entire workforce, with use concentrated in a small group of people who have incorporated them into the way they work. The rest open them, try them twice and go back to business as usual.

What is making the difference is not training in lecture format, but incorporating the tool into a specific process with an identified owner. A typical example: moving from "we have AI for drafting" to "commercial proposals are prepared with this draft and this checklist before being sent." The second case has an owner, a quality criterion and a way to measure whether it works.

• Choose a process, not a department: it is easier to demonstrate value and correct course.
• Document the flow: what is requested, what is reviewed and what is never delegated.
• Share what works: internal cases convince more than any vendor demonstration.

Regulation and governance: the work that is not seen

The regulatory framework advances at different speeds and with different levels of demand depending on the jurisdiction, and that forces companies operating in several countries to ask which rule applies to each flow. The practical consequence is that AI governance has ceased to be a document that is signed once and filed away.

What is being seen in orderly organizations is an inventory of use cases with minimal information: what data goes in, which provider is involved, what decisions rely on the system's output and who is responsible. A sophisticated platform is not needed to start; a well-maintained spreadsheet solves eighty percent of the problem and avoids surprises when an audit arrives or a customer asks.

A point that is often overlooked: AI that is used without anyone knowing. Free tools opened in the browser with customer information inside are a risk that does not appear in any record. Detecting that use and offering an approved alternative is more effective than prohibiting it.

Security: the same attacks, with more surface

AI expands the attack surface in two ways. The first is the classic one: more connected tools, more credentials, more integrations that someone configured on a Friday. The second is specific to these systems: malicious instructions hidden in a document or on a page that the model reads and obeys, and that can lead to leaking information or executing unforeseen actions.

The measures that are working are not very spectacular: limiting what each integration can read and write, not giving access to data the task does not need, reviewing the permissions of connections with the same discipline as those of any other system and maintaining the activity log. The usual rule still applies: AI does not replace controls, it puts them to the test.

What is worth having resolved going into October

October usually brings planning for the last stretch of the year, and with it the temptation to pile up projects. Against that, a short checklist delivers more than an ambitious plan:

• Updated inventory of AI tools in use, including those opened without going through IT.
• One use case with metrics and an owner, instead of five scattered tests.
• Review of permissions for active integrations and the data they access.
• Human review criteria written for everything that reaches a customer or affects a sensitive decision.
• Experimentation budget capped, to be able to test without compromising the quarter's spending.

The conclusion of the month is less exciting than the headlines suggest: competitive advantage is no longer in accessing the latest novelty first, but in having processes, permissions and criteria that allow adopting it without breaking anything. Organizations that move forward are those that treat AI like any other critical system: with an owner, with controls and with a clear idea of what problem it solves.

If you are reviewing where to start in your company, at Estratos we can help you organize that starting point and define what is worth testing first.