THE ADVISORY · ISSUE 02 · 28 SEPTEMBER 2026

AI at Work: Adoption, Cost & Control

A practical briefing on SME adoption, model costs, task-running assistants and connected-app controls—and what to do next.

Illustration for four AI developments affecting small businesses

AI use is spreading among UK businesses. Deep adoption is still rare.

A new survey suggests many firms have started experimenting with AI, while relatively few have changed how the business works around it.

NatWest reports that its survey of 1,400 UK small, medium and mid-market businesses found 44% currently use AI. A further 41% expect to adopt it within five years. Yet only 6% said AI had reached the most advanced stage in the survey: changing operations, decisions and customer experience. Among businesses already using AI, 29% remained at early stages such as awareness building or pilot projects.

The figures describe the businesses surveyed; they are not a guarantee that every company will see the same results. They do, however, point to a useful distinction. Trying a chatbot is not the same as improving a real process. A business may save time on a few drafts while still relying on manual copying, repeated checks and people who must fix inconsistent outputs.

Start with a task that repeats

More developed use is likely to involve a clear job, reliable information and a way to check whether the result helped. That could mean sorting incoming enquiries, preparing a first draft from approved company information or finding patterns in routine reports. The process should be measured against the current way of doing the work, including review time and errors—not just how quickly an AI tool produces text.

NatWest also reports that more advanced adopters in its research were more likely to report business benefits. That is an association in this survey, not proof that adopting AI alone caused the improvement. Skills, management attention, data quality and the choice of task may all matter.

How it affects your business

Pick one repetitive task and record how long it takes today, how often it needs correction and what information it uses. Trial a small change with a named owner and approved data. Keep it only if the total work improves and the people affected can check the result.

Primary source · NatWest →

Anthropic says its new Opus model costs 40% less to run on typical workloads.

A lower model price could make capable AI more affordable. The useful question is whether it lowers the cost of completing your whole task.

Anthropic announced Claude Opus 5.5 on 22 September. The company says it performs at the level of its Fable 5.1 model on most work and costs 40% less to run than Opus 5 on typical workloads. Anthropic also lists lower per-token prices for input and output, and says cached input is cheaper. These are the company’s comparisons and workload estimates.

Token prices are only one part of a business cost. A task may also involve long prompts, repeated model calls, waiting time, staff review, corrections and software fees. A cheaper model can still be poor value if it needs more retries or produces work that takes longer to check. Conversely, a suitable lower-cost model may make a routine task practical where the previous option was too expensive.

Compare the finished job

For a real workflow, compare the total spend and staff time needed to reach an acceptable result. Use the same task, source material and quality checks for both models. Include the rate of errors and how often a person must rewrite the output. Keep sensitive information out of a test unless the service and data settings are approved for it.

How it affects your business

If you already pay for model use, choose one recurring task and compare its full cost from start to approved finish. Record model fees, staff review time and corrections. Do not switch based on a headline percentage alone; keep the cheaper option only if quality and total effort remain acceptable.

Primary source · Anthropic →

Microsoft is bringing chat, task delegation and app building together in Copilot.

The announcement points towards assistants that can do more than answer questions. Availability is still limited, so treat it as a direction to watch rather than a tool every SME can use today.

Microsoft’s 25 September announcement describes three parts of its updated Copilot experience. Home brings chat and Cowork together; Code is designed to help people create applications and automations using natural language; and Autopilot is an agent intended to continue tasks independently. Microsoft also says Word, Excel and PowerPoint will be integrated into Copilot.

The rollout stages matter. Microsoft says Home and Code will begin rolling out to its Frontier program over the coming weeks, while Autopilot is expanding to private preview at the end of September. A private preview is a limited test, not general availability. Features, eligibility and controls may change as they are tested.

More capable means clearer boundaries

Chatting, building a small application and delegating a task are different levels of responsibility. Before using an assistant that can work across files or take actions, decide what it may access, which actions need a person’s approval and how to stop or review the work. Start with a contained, low-risk task and use business-approved accounts and data.

How it affects your business

You do not need to redesign your work around an announcement. List one admin task that repeats, its source files and the checks a person would make. When a suitable feature becomes available to your account, test it on a low-risk version of that task before granting broader access.

Primary source · Microsoft →

Connected apps make AI more useful—and make permissions matter.

When an AI assistant can use business apps, the information it can reach depends on account permissions and workspace settings.

ChatGPT can work with connected apps such as file storage and business services. The exact apps, features and controls available depend on the plan and workspace. OpenAI’s documentation says managed workspaces can have administrator controls over app access; connected apps also use the permissions attached to the authorised account. Some actions may require approval.

This is a practical control point, not a promise that connecting an app makes every use safe. An assistant should only have access to information required for its task. Admins should understand which apps are enabled, what roles can use them and whether an action can change or send information. Staff should connect only the accounts and files they are authorised to use.

Review access before connecting more

Start with the settings for the actual workspace and plan your business uses. Check whether an app is enabled by an administrator, which account is connected, what that account can access and whether the assistant can take actions or only retrieve information. Keep approval steps on consequential actions and remove connections that are no longer needed.

How it affects your business

Ask your workspace administrator to review connected apps and account permissions before rolling them out to a team. Use a limited account where possible, test with non-sensitive material and require approval for changes or external sends. Do not assume a control described for Enterprise or Edu applies to every plan.

Primary source · OpenAI →

Use the tool that fits the work.

AI products are changing quickly. A useful decision still starts with a real business task, approved information and a clear way to check the result.

Browse every Advisory edition