7 Practical GPT 6 Astra Use Cases for Business, and What It Takes to Make Them Work

OpenAI’s new GPT 6 Astra can use computers, browse the web, work across software, write code, conduct research, create documents and complete increasingly complex workflows.

That is exciting. But for businesses, the more useful question is not what can Astra do?

It is:

What can we build now that AI models are capable of doing more of the work themselves?

OpenAI released GPT 6 Astra on 3 September 2026. One of its biggest advances is computer use. OpenAI says Astra can fill forms, update CRM records, organise calendars, conduct research, work in document editors, create websites, run frontend QA and troubleshoot software through the interfaces people already use (OpenAI).

That moves AI beyond generating an answer for someone to implement. The model can increasingly participate in the workflow itself.

But there is an important catch.

A powerful AI model is not the same thing as a useful business system.

For AI to become genuinely useful inside an organisation, it needs the right business context, connections to existing software, rules, permissions, interfaces, testing and decisions about when a person needs to remain involved.

The model raises the ceiling. The system you build around it determines what you can actually do.

Astra changes what AI can do inside a business

Most businesses first experienced generative AI through a chat box.

You ask a question. AI gives you something. A person takes it from there.

That model is already changing.

Astra is designed to carry work across code, browsers and professional software. OpenAI reports that on its OSWorld computer use evaluation, Astra completed tasks with higher accuracy than GPT 5.6 Sol while taking about 47 per cent less time in its simulation. These are OpenAI’s benchmark results, so they should not be treated as proof that every real business workflow will improve by the same amount.

OpenAI published a similar gap on the ARC-AGI-3 reasoning benchmark.

Two caveats are worth holding onto. These are OpenAI’s own published figures rather than independent evaluations, and a benchmark score is not evidence that a workflow inside your business will work. What a jump of that size does indicate is that the ceiling has moved, which is exactly why the useful question shifts from what the model can do to what you could now build.

The more interesting change is the loop that becomes possible:

Understand what is happening. Decide what to do. Use a tool. Observe what happened. Continue until the task is complete.

That opens up far more ambitious forms of automation.

Here are seven examples.

7 practical GPT 6 Astra use cases

1. Sales and account intelligence

Imagine opening your CRM in the morning and finding that the research has already been done.

An AI system could review your active opportunities, inspect recent email and meeting activity, research relevant company developments, identify deals that need attention and prepare suggested next actions.

For one opportunity it might notice that a prospect asked a technical question two days ago and nobody replied. For another, it might identify a new stakeholder who should be added to the account. For another, it could prepare a briefing before an upcoming meeting.

OpenAI specifically lists updating CRM customer records among Astra’s computer use capabilities. More interestingly, OpenAI recently described how Clay has experimented with persistent AI workspaces for accounts that bring together information from CRM records, email, Slack, calls, presentations and other conversations. A coordinating agent can then turn that context into daily recommended actions (OpenAI).

The model is only one part of that system.

To make it useful, you need to decide which systems it can read, how account information is structured, which sources it should trust, what it is allowed to change and which actions still require a salesperson.

That context is what turns a generic AI model into something that understands how your sales process actually works.

2. Customer service that can resolve the problem

A conventional chatbot can tell a customer where to find your returns policy.

An agentic system could potentially do much more.

It could identify the customer, inspect their account, find the relevant order, understand the applicable policy, determine the available options, prepare the correct action and update the appropriate business system.

For straightforward enquiries, much of that workflow could eventually happen automatically.

For an unusual complaint, a large refund or a decision with contractual consequences, the system could prepare everything and hand the decision to a person.

That distinction matters.

The goal should not necessarily be removing people from the process. It should be removing the repetitive work that prevents people from spending time on the situations where judgement is actually valuable.

3. Research and management reporting

Consider how much work goes into a recurring management report.

Someone collects numbers from several systems. Someone checks what happened during the month. Someone searches for relevant market developments. Someone finds the previous report to remember how it was formatted. Someone turns all of that into something other people can understand.

Astra is designed for research and complex professional work, including the creation of documents, spreadsheets and presentations. OpenAI also says it has trained the model to select relevant context rather than simply repeating everything it has access to (OpenAI).

That makes a more useful reporting system possible.

It could gather current internal data, retrieve previous reports, investigate material changes, find supporting evidence and prepare the next report using the organisation’s existing structure.

A person might still review important conclusions. But the repeated gathering, organising and formatting work can increasingly move into the system.

The difference between that and typing "write my monthly management report" into ChatGPT is context.

The system knows what a management report means for your business.

4. Marketing and content operations

There is a much more interesting opportunity here than asking AI to write more content.

A bespoke marketing system could understand your:

• brand guidelines
• products and services
• audience research
• approved claims
• previous campaigns
• existing website
• commercial priorities
• publishing process

When a new campaign begins, it could research the subject, find relevant internal information, prepare campaign material, populate templates, create CMS drafts and send the appropriate work to the appropriate reviewer.

The same underlying model is available to thousands of other businesses.

The valuable part is the layer of business specific context and workflow around it.

That also makes AI easier to control. Instead of hoping someone remembers to paste the latest brand guidelines into every prompt, those rules can become part of the system.

5. Website and digital operations

This is where models capable of operating browsers become particularly interesting.

Imagine a website operations system connected to analytics, Search Console, support requests, your CMS and the live website.

A significant drop appears in a conversion journey.

Rather than merely alerting someone, the system could investigate.

It could inspect the affected pages, compare analytics, look for recent changes, test the journey in a browser and identify likely causes.

If the problem were a broken piece of content, it might prepare a CMS change.

If it were a development issue, it might create a technical task with evidence showing how to reproduce the problem.

Eventually, for sufficiently constrained situations, it could prepare and test the fix itself.

Astra can create websites and perform frontend QA through computer use, which makes workflows that combine diagnosis, implementation and verification increasingly realistic (OpenAI).

The key word is constrained.

A website agent should not automatically receive unrestricted production access because the model happens to be capable of using it. The system should determine exactly which actions are available and where approval is required. It is the same discipline we apply to any website design and build engagement.

6. Internal operations that understand the business

Consider this instruction:

Prepare me for tomorrow’s client meeting.

A generic AI model does not know what that means.

A business specific system could.

It could know which client you mean, which project is active, what has happened since the last meeting, what was promised, which tasks remain open, which issues have been raised, who is attending and which decisions need to be made.

The result could be a concise briefing containing:

• important changes since the previous meeting
• outstanding commitments
• recent correspondence
• unresolved problems
• relevant project information
• decisions required during the meeting

This is one of the clearest examples of why context architecture matters.

The intelligence of the underlying model might be identical.

What changes is what the system knows.

7. Software development and digital QA

Software development is already one of the clearest examples of the move from generation to execution.

Instead of simply generating code, an agent can increasingly receive a problem, investigate it, modify a codebase, run the software, inspect what happened and make another change.

OpenAI demonstrated this with Playco, where Astra was connected directly to Unity and Godot game development environments. The system could modify a game, run it, inspect the result and iterate. Playco reported 50 per cent fewer manual fixes than with the previous model across its three prototype evaluation, although that result comes from a customer example published by OpenAI and should not be generalised to other development workflows.

This highlights a bigger idea.

Generation is useful. Generation with access to the real environment and the ability to verify the result is much more useful.

The same principle applies outside software development.

An AI that drafts an answer is helpful.

An AI that can investigate the problem, complete the work, check what happened and escalate when something looks wrong is a different category of tool.

The model is only one part of the system

It is tempting to look at everything Astra can do and think the next step is simply giving it access to more software.

That would miss the most important part.

A useful agentic system usually needs several layers working together:

Context

What does the AI need to know?

This might include company documentation, customers, projects, policies, products, previous decisions, brand rules, analytics or current operational data.

Connections

Where does that information live?

The system may need to interact with your CRM, email, calendar, CMS, project management platform, databases, APIs, internal documents or other software.

Actions

What is the AI actually allowed to do?

Reading information is different from updating a record.

Updating a record is different from sending an email.

Sending an email is different from approving a payment.

Those actions should not automatically receive the same level of autonomy.

Controls

When should the system stop?

When should it request approval?

How does someone inspect what it did?

What happens when information conflicts?

What happens when an integration fails?

How do you reverse an incorrect action?

This is why building useful AI is increasingly an architecture problem rather than a prompting problem.

Anthropic describes a related concept as context engineering: instead of concentrating only on the wording of prompts, developers need to determine which instructions, tools, external information, message history and state should be available to a model as it works.

We think there is a useful broader business idea here: context architecture.

Context architecture determines what the AI knows, where it can look for more information, which tools it can use, what it remembers, which rules apply and when it needs to involve a person.

That architecture is what turns generic intelligence into a system that understands a particular organisation. It is the substance behind most of our AI and website services.

Why useful business AI will become increasingly bespoke

For decades, businesses have largely adapted themselves to software.

You buy a CRM and adopt its workflow.

You choose a project management platform and structure projects around it.

You implement a CMS and learn the way it expects content to be organised.

Over time, the gaps between those systems get filled with spreadsheets, manual processes, duplicated information and people whose job includes remembering how everything fits together.

AI creates an interesting opportunity to reverse some of that relationship.

Instead of beginning with:

How do we change our process to fit this software?

We can increasingly begin with:

What is the process we actually want, and what should the system do around it?

The software does not disappear. Your CRM, accounting platform, CMS, database and other systems still perform important jobs.

What changes is the intelligence layer connecting them.

OpenAI’s own enterprise guidance now recommends investing in proprietary company context, trusted connectors, curated knowledge, evaluations, observability, permissions and workflow logic when those elements create differentiated value. It also argues that organisations should measure completed work and business outcomes rather than treating raw model usage as the measure of success.

That points towards a future where simply having access to the newest AI model becomes less interesting.

Your competitors can access powerful models too.

What they do not automatically have is your:

data + context + processes + rules + integrations + interfaces + organisational knowledge

That is where more of the practical value can live.

Astra might provide the intelligence.

The bespoke system around it makes that intelligence useful to your organisation.

More capable agents need better boundaries

There is another side to this capability.

The more an AI system can do, the more carefully access needs to be designed.

Astra is the first OpenAI model that OpenAI has classified at its Critical cybersecurity capability level. OpenAI says that with suitable tools and access, the model can identify previously unknown vulnerabilities and develop ways of exploiting well protected systems without requiring a person to direct each individual step. OpenAI has consequently introduced stronger isolation, monitoring and deployment controls around the model.

For most businesses, the practical lesson is not to be frightened of Astra.

It is much simpler:

Capability and permission are different things.

An agent may be capable of accessing your entire database.

That does not mean it should.

It may be capable of sending an email.

That does not mean every email should be sent without approval.

It may be capable of modifying a website.

That does not mean it needs unrestricted production credentials.

This is also becoming an explicit concern for Australian organisations.

The Australian Signals Directorate updated its Information Security Manual guidance in September 2026 to include AI agent registers. For organisations using that framework, ASD recommends recording each agent’s owner, business purpose, identities, credentials, tools, permissions and accessible data repositories.

Separate guidance coauthored by ASD and international cyber security agencies recommends introducing agentic AI progressively, limiting privileges, maintaining human oversight, monitoring agent behaviour and restricting autonomous systems to clearly defined scopes.

A useful principle is:

Give an AI the minimum capability it needs to complete the job. Expand autonomy when the workflow proves itself.

For example:

Assist: the AI gathers information and prepares the action.

Approve: the AI completes most of the workflow but a person authorises consequential actions.

Automate: predictable actions can happen independently, while exceptions are escalated.

That gives organisations a much safer path than jumping directly from a chatbot to unrestricted autonomy.

How to find a useful AI opportunity inside your business

The best starting point is probably not asking:

What can Astra automate?

Start with the friction already inside the business.

Look for work where people repeatedly:

• collect information from several places
• copy data between systems
• search for context before doing the real work
• create similar outputs again and again
• follow reasonably predictable processes
• coordinate several pieces of software manually
• make repeatable decisions using information that already exists
• spend substantial time preparing work that someone else then reviews

Then choose one workflow and map it.

Ask four questions.

1. What outcome are we trying to achieve?

Be specific.

"Use AI for customer service" is vague.

"Reduce the manual work involved in triaging new support requests while keeping complex complaints with people" is much more useful.

2. What context is required?

Identify what a knowledgeable employee would need to understand before they could complete the task properly.

That tells you what the AI system may need access to.

3. Which actions are required?

List what actually happens during the workflow.

• Read a record.
• Search a document.
• Create a draft.
• Update the CRM.
• Send a message.
• Modify a file.
• Schedule a meeting.

Each action can then receive an appropriate permission level.

4. How will we know it worked?

Do not measure success in prompts, tokens or how impressive the demo looks.

Measure the actual outcome.

• Did the case get resolved correctly?
• Did the change pass review?
• Did the report contain the required information?
• How much human correction was required?
• What did the successful outcome cost?

OpenAI’s current enterprise guidance similarly recommends evaluating AI by useful work per dollar, including completed tasks, quality, human review and business outcomes, rather than token price alone.

That last point will become increasingly important as models become more capable and their pricing structures change.

The cheapest model per token is not necessarily the cheapest system to operate.

The useful measure is closer to cost per accepted outcome.

Start with the business, not Astra

GPT 6 Astra is exciting because it expands the range of work AI can realistically attempt.

It is unlikely to be the last model that does.

The bigger opportunity is learning how to turn increasingly capable models into systems that fit the way a specific organisation actually works.

Start with the business problem.

Design the workflow you want.

Decide what information the system needs, what it can access, which actions it can take and where people should remain involved.

Then choose the model capable of doing the job.

Better models raise the ceiling. Context architecture, integrations and thoughtful system design determine what becomes useful.

For businesses that get this right, AI becomes less like another piece of software people need to work around and more like a bespoke layer built around the organisation itself.

That is where things start to get genuinely interesting.

If there is a repetitive or fragmented workflow in your business that feels like it should already be handled by a system, we are happy to think it through with you.