GPT-6 is here: what changes for your business

On Tuesday someone on your team dropped the link in the group chat: OpenAI started rolling out GPT-6, codename Astra. On Wednesday your business partner asked if "we're migrating already". By Thursday nobody remembered the subject, because a quote fell through and the big client called to complain. This scene repeats with every launch. GPT-6 will repeat it again, with one difference: this time the model does more on its own, and that touches decisions that are not technical. They are yours.

Let me be direct about what changes, what does not, and how you decide without becoming a guinea pig.

What GPT-6 promises that is different

The previous launches sold better answers. This one sells execution. The core idea behind Astra is a model that handles long tasks without losing the thread: open a system, read what is there, compare it with another source, decide the next step and keep going. Less "give me some text", more "solve this and tell me when it's done".

The second thing that got attention was cybersecurity. The model got good at finding flaws in code. That cuts both ways, and OpenAI knows it, which is why the launch came tied to access controls for use in that area. Defenders get a strong tool. So do attackers.

For someone running a company, translated: AI moves from the role of the intern who writes into the role of the intern who touches your systems. The kind of mistake it can make changes. Bad text you throw away. A wrong action inside the ERP you have to undo.

Why switching models almost never solves your problem

Here is the unpopular opinion: switching models is the cheapest and most overrated part of the project.

In most of the companies I work with, the bottleneck is not model intelligence. It is context. The AI does not know your current price list, does not know client X has had special terms since 2023, does not know an order only goes through after finance approves it. You can throw the most expensive model in the world at that and it will be wrong with a lot of elegance.

I have a recent case that illustrates it well. A distributor wanted AI to answer quote requests over WhatsApp. The real problem was never the wording of the reply. It was that the price list lived in three places: a spreadsheet on Drive, the legacy system, and the head of a salesperson with eighteen years in the company. Organizing that took two weeks. Plugging the AI on top took two days. Average response time dropped from four hours to eleven minutes. Not because the model is brilliant, but because there is now a single source of truth for it to check.

A new model does not fix a messy process. It just makes the mess faster.

What really changes with more autonomous agents

That said, some things do change. When a model can handle long tasks, a category of automation opens up that used to not pay off.

Three concrete examples of what becomes viable:

• Reconciliation. Matching bank statements against issued invoices and flagging the gaps. Boring, repetitive work with clear rules. It used to require hand-built integration for every case. Now you can describe the rule and let it run.
• Document triage. Contracts, reports, supplier proposals. The model reads, pulls out what matters, flags anything outside the standard and calls a human only for those.
• Sales follow-up. Look at the CRM, find who stopped replying, check the history and suggest the next action with the right context. None of that generic "hi, how are you?" everyone ignores.

Notice the pattern. These are tasks with defined rules, high volume and a possible review step. That is where autonomy pays off. Not in strategic decisions, not in talking to a furious client, not in anything you cannot check afterwards.

The new risk that came along

The security part deserves your attention, even if you know nothing about technology.

If the model got good at finding flaws in software, scammers have access to a version of that too. The phishing email that used to arrive in broken Portuguese now arrives perfect, with the name of your real supplier and the context of a conversation that actually happened. I have seen a company almost pay a fake invoice of R$ 38,000 because the email copied the supplier's format with frightening precision. What saved them was a dumb rule: above ten thousand, someone calls to confirm by phone.

Do this today, it costs nothing:

• Any change to a supplier's bank details requires phone confirmation, using an old number, never the number that came in the email.
• Two people approve any payment above a threshold you define.
• Nobody installs an extension or AI tool on company systems without going through you.

And if you put an agent to work inside your systems, give it minimum access. Read before write. Test environment before production. Logs of everything it did. It is the same care you would take with a new employee in their first month, except the agent works faster and is not embarrassed to ask twice.

How to decide whether it is worth moving now

One approach that works, in four steps:

1. List the three tasks that eat the most of your team's time. Measured time, not felt time. Ask the people doing the work. You will be surprised by what comes up.

2. See which of them has clear rules. If two experienced people would do the same thing facing the same case, there is a rule. If it depends on instinct, leave it for later.

3. Check whether the information needed is somewhere accessible. A system, a spreadsheet, a folder. If it only lives in someone's head, the first project is getting it out of there.

4. Run a two-week pilot with human review on 100% of cases. You compare what the AI did with what your team would do. If accuracy lands above whatever you defined as acceptable, loosen the review to spot checks.

In that process, which model you use is almost a detail. GPT-6, the previous version, a competitor, it hardly matters at the start. You switch later, once the rest is standing, and it takes fifteen minutes.

What to do on Monday

Do not migrate anything because of a headline. Launches are always news, adoption is always slower, and the new version usually needs a month to settle.

What is worth doing now is the work no launch does for you: understanding where your company loses time, organizing the information that is scattered, and picking one task to test. Whoever did that homework will plug in GPT-6 and feel the difference the following week. Whoever did not will sign up for one more tool, use it for three weeks and forget it, exactly like the other five sitting on the corporate card statement.

The technology got stronger. The question is the same as ten years ago: which problem of yours does it solve, and how will you know if it worked.

If you want to talk about where it makes sense to start in your case, take a look at how I work.

GPT-6 is here and the question I heard most this week was "should we migrate already?".

Switching models is the cheapest and most overrated part of the project.

Last week a distributor wanted AI to answer quote requests on WhatsApp. The problem was never the text. It was the price list living in three places: a spreadsheet on Drive, a legacy system, and the head of a salesperson with eighteen years in the company.

Organizing that: two weeks. Plugging the AI on top: two days. Response time dropped from four hours to eleven minutes.

A new model does not fix a messy process. It just makes the mess faster.

Before signing up for one more tool, look at where your company loses time and where the information is scattered. Whoever does that homework plugs in GPT-6 and feels the difference the following week.

If you want to talk through where to start in your case, reach out.

#ArtificialIntelligence #GPT6 #Automation #DigitalTransformation #Management