A company buys AI subscriptions. Emails take less time to draft, meetings come with summaries, and customers still wait half a day for a reply. Their enquiries sit in a shared inbox until someone assigns them. The subscription made drafting easier. A well-designed implementation starts by finding out where the customer loses time.
NBER’s summer 2026 survey reached nearly 9,000 executives across seven countries. Some 78% of firms used AI, while 83% of surveyed firms reported seeing no effect on labour productivity over the previous three years. These are executives’ assessments, which show the difference between how widely firms use the tool and how much of an effect they have seen. Firm Data on AI · NBER
1. Counting subscriptions instead of outcomes
Active users and prompt counts show whether staff have picked up the tool. Customers care about when they receive a meaningful reply.
How to avoid it. Pick one process, name the manager accountable for its outcome and record a baseline. For example, the head of sales agrees on a target for the time to a first meaningful reply and sets a date to review the result.
2. Speeding up one task while the queue grows elsewhere
A demo can easily show how AI drafts an email. In daily work, most of the delay may come from an enquiry waiting for someone to pick it up.
How to avoid it. Follow several enquiries from submission to reply. Measure active work and waiting time separately. Start with the step that creates the queue, then check whether it simply moved to the next team.
3. Giving the model work that rules can handle
A standard integration can transfer a name and phone number from a structured form into a CRM. A model can help interpret free text, for example by distinguishing a support request from a new order.
How to avoid it. Use rules to transfer data and route standard enquiries. Let the model propose a classification. Have people review ambiguous cases and approve actions with significant consequences. Test the manual route before launch.
4. Scaling a pilot based on minutes saved
The time freed up may go towards more enquiries, more complex cases or less overtime. Each of those outcomes calls for a different measure.
How to avoid it. Compare the pilot with a similar stream of enquiries or a baseline period, accounting for volume and complexity. Track routing errors and repeat contacts alongside speed. Include integration, licences, manual review, support and security in the cost.
5. Connecting AI to systems before defining its permissions
When a model can read customer emails and update CRM records, an external email could contain instructions designed to make it take an unauthorised action.
How to avoid it. Decide what the model may read, what actions it may propose and who can approve them. Enforce permissions in the business system, log actions and keep a manual route available if the system fails.
Generative AI Risk Management Profile · NIST
Before scaling, I would ask the process owner three questions. What changed for the customer or the team? How did we measure it? What does it cost to sustain the result?