AI Agents
An assistant given a narrow, bounded job — look something up, fill a form, update a record, notify a person — with a fixed list of allowed actions and a hard stop wherever the task falls outside its scope.
What an AI Agent Actually Does Here
An agent is not a decision-maker with open access to your systems. It follows a defined sequence: read a request, retrieve what it needs, take one of a small set of pre-approved actions, and notify or hand over when the task is outside that list. This is different from a chatbot, which only answers questions and takes no action.
Where Agents Are Used
Vendor Onboarding Data Collection
Requesting and checking required documents from a new vendor and flagging what is missing, without approving the vendor itself.
IT Helpdesk Access Requests
Handling routine password resets or standard access requests within policy, escalating anything outside the defined list.
Order Status Lookups Across Systems
Pulling a consistent status by checking multiple internal systems instead of a person doing the same lookup by hand.
Inventory Reorder Point Checks
Flagging stock that has crossed a reorder threshold and notifying the person who places the order, without placing it.
Guardrails We Build In
- A fixed list of allowed actions — no open-ended commands the agent can invent for itself.
- An approval step before anything that commits money, changes a record permanently, or sends an external message.
- A full log of every action taken, the trigger behind it, and the outcome.
- An automatic stop and handover the moment the request falls outside the agent's defined scope.
How an Agent Is Scoped
Task and Boundary Definition
We write down exactly what the agent may and may not do before any code is built.
Action Whitelisting
Each system action the agent can take is enumerated and permissioned individually.
Testing Against Edge Cases
We deliberately try to push the agent outside its scope to confirm it hands over correctly.
Supervised Rollout
Early runs are reviewed by a person before the agent is left to run unattended.
What Stays With a Person
Tasks that need a judgement call not covered by a written policy are not handed to an agent alone. Those stay with your staff, who can be supported by a knowledge assistant for reference instead.
Frequently asked questions
How is an AI agent different from a chatbot?
A chatbot answers questions. An agent carries out a defined sequence of actions across your systems, within a fixed set of permissions, and hands over when it hits the edge of its scope.
What drives the cost of building an agent?
The number of systems it needs to reach, how many distinct actions it is permitted to take, and how much testing those actions need before going live.
What drives how long an agent takes to build?
How clearly the task boundary can be written down up front, and how many systems need integration access.
Can an agent change records in our CRM or ERP without approval?
Only within the specific actions you approve in advance, and anything that commits money or permanently changes a record waits for a named approver.
Who is responsible if the agent gets something wrong?
The action log shows exactly what happened and why, and the approval step is designed so nothing consequential leaves the system unchecked.
Who owns the agent's code and configuration?
You do, along with the action whitelist and logs, so the agent's behaviour is never a black box you cannot inspect.
What do we need to provide to scope one?
A clear description of the task as it is done today, and access to the systems the agent would need to read from and act on.
Tell us what you need.
Send a short brief and one of our engineers will come back to you — usually the same day.
- No obligation
- We reply the same working day
- Your details stay private