To tell whether your AI agent is improving, track the same measures over time. Improvement is a trend, and a trend requires consistent measurement from the start. Why single measurements mislead A month's figures reflect that month's mix of requests. * An easier month read as improvement
To stop an agent answering outside its scope, define what it should do when it does not know. An agent given no instruction for uncertainty will answer anyway. Why out-of-scope answers are the dangerous ones They are indistinguishable from good ones until someone checks. * A confident answer drawn
To run several AI agents, give each one a defined scope and an owner. A set of agents with overlapping material produces inconsistent answers to the same question. What goes wrong with several agents They are created independently and never reconciled. * Two agents answering the same question differently * Documentation duplicated
To give an AI agent a voice, define the tone explicitly. Left unspecified, it defaults to a generic register that sounds like every other assistant. Why tone matters more than it seems Customers notice the register before they evaluate the answer. * Excessive formality where the brand is casual * Excessive friendliness
Understanding how AI can support a sales team means separating administration from persuasion. The administration is substantial, repetitive, and worth removing; the persuasion is not something to delegate. Where sales time actually goes Very little of a salesperson's week is spent selling. * Updating records after conversations * Categorizing and
To run a service desk with AI, the agent has to do more than answer. Handling a request end to end means classifying it, replying, updating the record, and closing it when the requester is satisfied. What partial automation leaves behind Answering is one step of five. * Someone still classifying
To use AI in financial document processing, apply it to transcription and keep it away from authorization. The volume is in the first; the risk is in the second. Where the volume actually is Most finance document work is reading and typing. * Invoices arriving from many suppliers in different formats
To answer HR policy questions with AI, attach the policy documents to an agent. The questions are high-volume, repetitive, and answerable from written material — which is exactly the profile. What HR spends time repeating The same questions arrive continuously from different people. * Holiday entitlement and how to request it
To build an internal chatbot for a department, keep its scope narrow. An assistant that answers everything badly is less useful than one that answers a defined set well. Why a single company-wide bot underperforms Different departments have different documentation, tone, and audiences. * One knowledge base covering unrelated topics
Choosing between an agent or automation depends on whether the step needs interpretation. A rule that can be stated exactly should be an automation; an agent is for what cannot. Why this choice matters An agent applied to a deterministic rule adds cost and variability for nothing. * Using AI where
To fill records from documents automatically, the reader creates the record rather than only extracting text. The file arrives and the row exists, populated. What manual entry from documents costs It scales with volume and none of it is judgement. * Values typed from a PDF into a form * Transcription errors
To use AI to triage an email inbox, the classification happens on arrival. Nobody reads a message in order to decide who should read it. What manual triage costs One person spends their day routing work they cannot do. * A first responder reading everything to sort some of it * Delay