Before connecting AI to a business account, define the job, limit the information and actions available to it, and decide how a person can stop it. An automation that saves admin time still needs a clear owner. Your main login should not become a shortcut around those decisions.
Research released by Dashlane on 19 November puts the issue into focus. In its survey, 21% of knowledge workers said they had shared a password, API key or login credential with an AI tool or agent. The study covered 3,000 workers and 800 IT/security executives across the UK, US and France, with fieldwork conducted between 18 August and 5 September 2026.
These are self-reported findings from vendor-commissioned research. They do not measure the breach rate among small businesses, and they do not establish a UAE statistic. The practical question for owners in either market is the same: what access has your business actually granted?
Start with the business job
“Use AI” is too vague to judge commercially. “Prepare a draft response to a new quote enquiry” gives you something to test.
Write down the current workload before choosing software: how many enquiries arrive, how long a first response takes and how often someone has to chase missing information. Those figures give you a baseline. If the proposed tool creates more checking and correction than it removes, the workflow needs changing.
Also ask whether a simple form, template or rule would solve the problem. An automatic acknowledgement does not necessarily need an AI agent making decisions.
Separate reading, drafting and acting
Consider an illustrative enquiry workflow for a small installation business. A website form collects the service required, a broad location and the customer’s preferred contact method. A defined process checks whether the essential fields are complete. An assistant uses approved service information to prepare a response for a person to review.
That job does not automatically require access to the owner’s entire mailbox, every customer record or payment settings. Sending a reply, changing a quote and issuing a refund are separate actions with different consequences. Decide which are genuinely needed before connecting systems.
This is workflow design, not a claim that one configuration fits every business. The right setup depends on the information involved and what a mistake could cost.
Make the limits real
The UK NCSC’s established guidance warns against relying on prompts alone. Written instructions need technical and operational controls behind them. Joint international guidance also recommends tightly restricted access and cautious adoption.
For a first pilot, check:
- Connections: use an approved integration with permissions appropriate to the job; avoid pasting your main account password into a chat.
- Scope: restrict the tool to the data and actions it needs. Where supported, give it a separate identity and time-limited access.
- Approval: put an enforced review step before higher-impact actions, rather than relying on the assistant to remember to ask.
- Visibility: keep an activity record that lets the owner see what happened and investigate a mistake.
- Disconnection: test how to revoke access and return to the manual process.
Controls reduce exposure; they do not guarantee security or remove the need to assess the tool and the data it handles.
Run a small, measurable trial
As a starting point, choose one workflow, one owner and a review date after the first working week. Use suitable test information first. Record time spent reviewing outputs as well as time apparently saved, and check whether replies are accurate and useful.
For example, a fast draft is little help if it misses the customer’s location or promises a service you do not offer. A slower but dependable process may be commercially better. Expand only when the evidence supports it.
CREATE’s AI consultancy and automation services focus on finding practical ways to reduce repetitive work. If you have one task that keeps swallowing time, tell CREATE what it is. We can work through the process, the connections and the boundaries before a build starts.
Sources
- Dashlane’s original research announcement, published 19 November 2026; survey findings and methodology.
- UK NCSC: Managing the cyber risk of agentic AI, existing technical guidance.
- Canadian Centre for Cyber Security: Careful adoption of agentic AI, joint international guidance including the UK NCSC.
Sources checked 20 November 2026. This article is practical business and technical guidance, not a statement that any named business has suffered a breach.