AI-Native Isn't About Spending More Tokens — It's About Redesigning Workflows

Recently, OpenAI shared the practices of three AI-native companies: Basis, Clay and Exa.

These cases are not about simply letting employees use a few more AI tools — they are about putting agents inside the company's actual workflows:

Basis turned its employee onboarding process into reusable agent skills; Clay builds a continuously updated workspace and a dedicated agent for every customer; and Exa lets agents proactively discover integration opportunities, gather background information, write code, run tests, and hand the results to humans for review.

News link: OpenAI: How AI-native companies turn workflows into operating capability

This article made me rethink what it actually means to be an AI-native company.

Many companies use AI today. Employees use it to write emails, summarise meetings, generate copy or answer questions. The company burns through a huge number of tokens every month, and it all looks very "AI-ified".

But if the way the company works hasn't changed, AI is still just a more convenient assistant.

Using AI a lot does not mean a company has become AI-native.

The real change happens in the workflow itself.

Traditional software usually waits for people to push the process along: an employee logs in, finds the feature, organises the materials, fills in the form, and passes the task to the next person.

In an AI-native way of working, an agent can continuously observe state, start work proactively when trigger conditions appear, call the tools it needs, preserve execution evidence — and hand decisions back to humans at critical points.

The human role changes accordingly: from repeatedly executing every step, to defining goals, handling exceptions and reviewing results.

This is also the direction we are exploring in designing SnapLedger and Snappy.

For example, traditional accounting software might show a dashboard telling the user how many transactions remain unprocessed. But it still requires the user to discover the problem, understand it, and then hunt down the corresponding feature item by item.

The idea behind Snappy of the Day is different.

Snappy should combine the user's bank transactions, documents, ledger state, tool-usage traces and compliance dates to judge what the user most needs to get done right now:

Is there a large transaction waiting for confirmation today? Is there an invoice with no matching payment? Is the VAT filing date approaching? Does an unfinished task from last time need to be continued? Could a new tool help the user solve the problem in front of them?

It doesn't just tell the user "here is a feature" — it connects context, tools and the next action.

Likewise, Snappy Front Desk should not merely generate a marketing email. It needs to start from discovering a potential customer, understand their background, choose the right communication channel, record the progress of the conversation — and keep pushing registration and adoption once the customer shows interest.

And when a customer raises a request or reports a bug, it shouldn't stay as a line in a chat log. The relevant information can be organised into a ticket, enter the development and testing pipeline, and I periodically review which changes can actually be merged into the product.

In all these examples, the value of AI is not in how much text it generates, but in whether it pushes a piece of work from start to result.

So measuring an AI-native company shouldn't be about model call volume or token consumption either. It should be about more real questions:

• How many human hand-offs does a complete task require?
• How much context can an agent retain over time?
• How much work moves from reminder to completion?
• Are exceptions handed back to humans in time?
• Can successful ways of working be saved and reused?
• Do efficiency gains ultimately translate into lower cost, faster delivery or better service?

I've come to believe more and more that the AI gap between companies in the future will not simply come from who uses the stronger model.

The real gap lies in this: whether a company can clearly decompose its most important work, provide agents with the right context, tools, permissions and completion criteria — and then let successful processes repeat and improve continuously.

Models provide intelligence. Workflows turn intelligence into capability.

AI-native is not about everyone using AI a little more — it's about the whole organisation getting work done in a new way.