AI · AI at work
Adopting AI at work: From a pilot to a service you can run
Measure an AI pilot against the time it takes to produce a usable result. Compare platforms, calculate the cost per case and put AI literacy into practice.
In this article
Reviewed on 4 November 2026. A support team wants to spend less time answering routine questions. Measure the time from opening a ticket to having a correct reply ready to send. Timing the model’s first draft misses the work that a plausible but wrong answer creates.
Design a pilot that can fail cheaply
Pick a task whose output a named person can judge. Establish how long it takes today, which errors occur and what data it actually needs. A returns enquiry may require the approved policy and the status of a particular order. It rarely needs the customer’s entire history.
Build a small evaluation set with ordinary and difficult cases. Include an expired returns window, a goodwill exception and a message with no order number. Ask staff which answers they would send. Keep some cases out of development and check whether the time saving and quality hold up on those, too. Planning an SME pilot · Adoption within the team
Four platforms and four questions for procurement
Ask each supplier to price the same job: read an enquiry, retrieve two sources and draft a reply. Find out who grants source access, how staff permissions are revoked when roles change and how to export the process. The price of one credit tells you little until you know how many that job consumes. Microsoft · Salesforce · Google · ServiceNow
Calculate the cost of a usable reply
Suppose the team handles 600 cases per month at six minutes each. At a fully loaded €30 per hour that costs €1,800. In a pilot, 480 cases take three minutes each, including review. The other 120 take eight minutes because someone discards the draft. That adds up to 40 hours or €1,200. Assume another €300 for licences and usage. That leaves €300 per month to cover integration, ongoing operation and any additional exceptions before the saving disappears.
Count the discarded outputs. Caching can help when context repeats, while batch processing can suit work with a flexible deadline. Actual rates depend on the model, plan and billing route. A cheaper model that creates more review work can increase the total cost. How AI costs add up · Context engineering
Give people practice and give the system boundaries
Article 4 of the EU AI Act requires measures to support AI literacy, taking account of experience, the setting and the people affected. Give the support team an invented refund policy to spot. A recruitment team also needs to understand when a tool’s scores influence candidate selection. Written rules about inputs are useful to both teams, but each deployment needs its own assessment. Article 4 as amended · AI Act dates and duties
Start with read access limited to the sources the pilot needs. An incoming email might claim that a customer has approved payment to a new bank account. Even if the model follows that instruction, the tool should have no way to change the account details. If you later add payments, show the amount, recipient and purpose at the approval step. That gives you a concrete permissions test to run before launch. Research on separating data and control · Prompt injection and agent security
Decide what to do with the results
Agree before the first test who can own the service and what result would justify expansion. The support team might approve drafts for standard cases while keeping goodwill decisions manual. Once you know correction time, quality, cost and access requirements for each group, you can decide whether to extend the pilot to another workflow.