16 September 20268 minute read
An AI agent creates value when it moves work forward. That sounds obvious, and it is the test most “we should add AI” conversations fail. A plausible paragraph in a chat window is not an operational outcome. A faster first response to a known class of enquiry, a complete draft for a human to check, a correctly routed case, or a retrieved policy with a citation can be.
VCS Consulting treats AI agents as an optional layer on software and operations, not as a strategy of their own. The question is always the same: which piece of work is slow, repetitive or easy to get wrong, and could a controlled agent help a person do it better?
The pattern that tends to work
Good first uses share a shape. Demand is frequent enough to matter. The knowledge required can be approved and bounded. The next action is definable. Failure is visible and recoverable. A person remains accountable for anything that commits the business — a price, a legal position, a customer promise, a change to a record that cannot be undone casually.
- Customer and website queries that follow a known script, with escalation when they do not.
- Internal knowledge retrieval across policies, procedures and product guidance.
- Enquiry triage: classify, extract, summarise and route, rather than invent a reply.
- Document handling where the fields are known and a person still reviews exceptions.
- Workflow assistance that prepares the next step in a system the team already uses.
- In-product help that uses the customer’s own context, not a generic chatbot bolted to the homepage.
Where agents usually waste money
They struggle when the work is rare, highly judgement-led, or dependent on information nobody has written down. They also struggle when there is no owner for the knowledge, no permission model, and no agreement about what the agent is forbidden to do. In those conditions you do not get an assistant. You get a new source of confident mistakes.
A public-facing agent with nothing behind it except a marketing site is a common example. If the business cannot already answer the question consistently by email, the model will not invent a reliable service operation for you.
Value is measured in the operation
Judge the work, not the novelty. Useful measures include time to first useful response, the share of enquiries that no longer need a specialist, cycle time through a defined process, rework caused by missing information, and how often the agent must hand over. “We have launched an agent” is not a result.
The implementations that last are connected to systems, content and people: approved sources, clear permissions, a path into the tools teams already use, and a review habit when the work is important. That is slower to pitch than a demo. It is the difference between a feature and a change in how the business operates.
