Give it access, and know who can revoke it.

Letting an artificial intelligence (AI) model use tools and look up your own records sounds like a small step. In practice, it is the difference between a chat window and a colleague with a key to the filing cabinet.

Most people's first experience of an AI model is a text box. You ask a question, it answers from what it was trained on. Giving it tools means something different. It can check your stock levels, look up a customer's order history, read your calendar or update a record, then act on what it finds. It is no longer giving an opinion. It is doing part of the work itself.

What changes once it can look things up

The advantage is real. One AI research laboratory tested its own internal analytics agent and found accuracy rose from about 30% with no business context to about 90% once it could see that context (Venturebeat). The same test showed accuracy falling to about 60% two months later, once the context it relied on had gone out of date (Venturebeat). The lesson is not connect it and forget it. The lesson is that the lookup has to stay current, or the model's answers quietly rot.

The shift is already changing how businesses are approached from outside. At one company, AI agents acting on behalf of customers have begun dealing with the business directly rather than through a person, a pattern that started in Asia. That leaves the business without the usual trail of historical data to plan against (SiliconANGLE).

Where it goes wrong

The trouble is rarely the AI model itself. It is the access sitting around it. A survey of organisations already running AI agents in production found that half allow some sharing of login credentials between different agents (Venturebeat). That means several tools are using one identity, so nobody can say afterwards which one did what.

Taking access away again is just as weak. In a separate survey, only 14% of organisations said a named owner could revoke an agent's access within minutes (Venturebeat). For most businesses, switching off a tool that has gone wrong is slower than switching off a departing employee's log in.

Used well, the same access can do real work. At one company, AI agents in customer support more than doubled containment, meaning more enquiries were resolved without a person, and this cut one team's headcount by about 30% through attrition alone (SiliconANGLE). That is the upside businesses are chasing. It only arrives safely when the access behind it is scoped, owned and checked.

So before you agree to any of this, ask your supplier a short list of questions. Which records can the AI model read, and which can it change. Does it log in as itself, separate from any other tool or person. Who can switch off its access, and how fast. How often is the information it relies on refreshed, and who checks it is still correct.

At Yunaris, we scope each of those answers before we build, and quote the work on what your business needs the AI model to see and do.

Sources

1. Scaling AI agents requires inventory, contracts and maintained context, Venturebeat, 20 October 2026
2. NiCE bets the company on AI, data and accountability, and the legacy business buys it time, SiliconANGLE, 20 October 2026
3. AI cloud security needs to prove control before agents reach production, Venturebeat, 20 October 2026