AI in October 2026: five developments worth paying attention to

DEVELOPMENT 01

AI agents are starting to do the work

For the past few years, most of us have used AI by asking a question and doing something with the answer. That is changing.

Google introduced its Gemini agent for work on 8 October, with the ability to work across applications such as Gmail, Drive, Docs and Sheets. It is part of a wider move towards systems that can complete multi-step work, rather than simply suggest what a person should do next.

Imagine an agent checking a customer enquiry against your records, preparing a response and updating your CRM. Or comparing supplier invoices with purchase orders and flagging the exceptions for approval. Those tasks still need clear boundaries, but they are becoming much more realistic.

DEVELOPMENT 02

More capable AI is getting cheaper

An encouraging change this month has been the improvement in what businesses can get for their money.

Anthropic released Claude Haiku 5.5 on 7 October and said it costs about 75% less to run than Haiku 4.5 on average. OpenAI introduced GPT-6.1 Terra on 26 October, with standard API pricing 50% below GPT-5.6 Terra. Those figures are vendor-reported and actual costs depend on workload.

The practical lesson is that you probably do not need a premium reasoning model to classify incoming documents, route enquiries or extract invoice numbers. More complex analysis might justify a different model. Matching each task with an appropriate model can have a substantial impact when processing thousands of requests.

DEVELOPMENT 03

Security and accountability matter more when AI can act

There is a meaningful difference between a chatbot producing a questionable answer and an agent making a change in a business system.

On 8 October, the UK Information Commissioner’s Office opened a call for evidence on agentic AI, covering data security, transparency, accountability and automated decision-making. On 12 October, the UK government announced further work on risks involving autonomous AI systems.

If an agent can access sensitive records, update accounts or trigger a payment workflow, organisations need to know precisely what it can see and do. Permissions, approvals, logging, testing and a reliable way to interrupt a process are essential design decisions, not optional extras.

DEVELOPMENT 04

Enterprise adoption is moving beyond experiments

Big organisations are beginning to deploy AI as part of everyday work, not just as small technology trials.

On 22 October, Uber announced plans to make ChatGPT Work available to more than 30,000 employees, expanding from a 1,000-person pilot. It is a significant rollout, although an announcement alone does not tell us the eventual productivity gains or return on investment.

Smaller businesses should not be intimidated by the scale. They often have the advantage of simpler processes, fewer layers of approval and the ability to change quickly. Automating a repetitive administrative task can be more valuable than launching a large programme with no clear outcome.

DEVELOPMENT 05

You do not have to build around one AI provider

One announcement that deserves more attention came on 21 October, when OpenAI and Google Cloud said they were partnering to bring OpenAI models and Codex workflows to Google Cloud’s Model Garden.

That does not mean every model is interchangeable. It does reinforce a useful principle: businesses should be able to choose the right technology for each problem and reconsider that choice as the market changes.

For a custom application, that might mean using one model for document extraction, another for demanding reasoning and a smaller model for everyday classification. The trade-off is additional testing and governance, but a modular architecture can help avoid unnecessary dependence on a single provider.

PRACTICAL NEXT STEPS

What should your business do next?

The pace of AI development can make it tempting to wait for the next breakthrough. Yet the technology is already capable of handling useful work in document processing, customer service, reporting, administration and software development.

Choose one process that takes more time than it should. Map how it works today, identify where people need to exercise judgement, and decide which parts can be automated safely. Measure the result before expanding.

It is also worth asking whether another software subscription is the best answer. Sometimes an existing SaaS product is exactly right. In other cases, a focused, purpose-built application can fit the way your business operates better and avoid unnecessary ongoing costs.

The businesses that benefit most will not necessarily be the ones using the newest model. They will be the ones choosing useful problems and building reliable solutions around them.