MODULE 2 · LESSON 2
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolUse Case Catalogue: Customer-Facing Functions
What follows is a catalogue of patterns that repeatedly work, organised by function. Read it as a prompt for recognition, not a menu — the version that works for you will be the one where your own operation has the four signals.
One point applies across this whole lesson. Customer-facing AI carries risk that internal AI does not, for a simple reason: a mistake is visible to someone outside the organisation, it may be public, and it may be regulated. Everything here needs a tighter exception path than an equivalent internal use case.
Marketing
Content drafting at variant scale. Producing twenty variants of an ad, a landing page or a subject line for testing. The value is not writing quality; it is that testing many variants was previously uneconomic. A new capability case wearing a cost case's clothes.
Audience and campaign analysis. Summarising performance across channels, surfacing which segments respond to what. Genuinely useful, and note that it is a data science task — insight for a human — not an automation task.
Repurposing across formats and audiences. One piece of research becoming a blog post, a customer email, a sales one-pager and a social sequence. High volume, low risk, easily reviewed.
Search and content brief generation. Structuring what a piece needs to cover before a writer starts.
Where the risk sits: anything published without human sign-off. Marketing is the function where confidently invented facts most easily reach the public — a fabricated statistic in a campaign is a reputational problem and, in regulated sectors, a compliance one.
Sales
Call and meeting summarisation with CRM updates. Consistently among the highest-adoption use cases anywhere, because it removes admin salespeople genuinely hate. Volume is high, the outcome is verifiable by the person who was in the meeting, and errors are cheap.
Lead qualification and routing. Scoring inbound enquiries and routing them. Strong four-signal fit where historical outcomes are recorded — and they usually are, because CRMs record what happened to a lead.
Proposal and quote drafting. Assembling a first draft from a specification and prior proposals. The efficiency gain is real; the exception rule matters — pricing and legal terms should never be generated freely.
Research before a call. Assembling a briefing on an account from internal history and public sources. Low risk, immediately useful, requires no integration to start.
Where the risk sits: anything that states a price, a commitment or a delivery date. Those must come from systems of record, not from a generative model.
Customer service
Reply drafting for agents. An agent-assist pattern — the system drafts, the agent edits and sends. The most consistently successful service use case, because it keeps a human accountable while removing most of the typing.
Ticket classification and routing. The canonical back-office-style case applied to a front-office function. Excellent four-signal fit.
Knowledge retrieval for agents. Surfacing the relevant policy or article for the case in front of the agent. Frequently more valuable than reply drafting, because the agent's real problem is often finding the answer rather than typing it.
Post-contact summarisation. Wrap-up notes written automatically. Small saving per contact, very large in aggregate, and it improves the record quality that everything else depends on.
Fully automated customer-facing chat. The one everybody asks for first and the one that most often disappoints. It is not that it cannot work; it is that it needs a much higher standard, a genuine escalation path, and a tolerance for the tail of unusual questions.
Organisations almost always want to start with the fully automated customer chatbot, and almost always should not. The reasoning is worth having ready, because you will need to make this argument.
The tail is unbounded and public. Internal ticket routing has six possible answers. A customer conversation has no fixed set — customers ask about things you do not sell, describe situations you have not seen, and arrive mid-crisis. Every unusual case is handled in front of the customer.
The error is asymmetric. A misrouted internal ticket costs half a day. A confidently wrong answer to a customer about their cover, their refund rights or their medication can be a complaint, a regulatory matter, or worse. The same model accuracy carries a completely different consequence.
Agent-assist produces the data that makes automation possible later. This is the sequencing argument, and it is the strongest one. When an agent edits a drafted reply before sending, that edit is a labelled example: here is what the system proposed, here is what a competent human actually sent. Thousands of those are precisely the dataset you would need to automate safely — and you cannot obtain them any other way.
So the sequence that works is: assist first, measure the edit distance, automate the categories where the drafts stop being edited. After six months you will know exactly which question types the system handles at human standard, because the evidence is in the edit history. Those are the ones to automate, and you will be automating them on evidence rather than on hope.
The organisations that go straight to full automation skip the step that would have told them which parts were safe to automate — and then find out from customers.
Why does the agent-assist pattern in customer service usually make a better first project than full automation?
Customer-facing risk
Click to flipMistakes are visible outside the organisation, sometimes public and sometimes regulated. Every customer-facing use case needs a tighter exception path than its internal equivalent.
Click to flip backAcross marketing, sales and service the reliable patterns are drafting, summarising, classifying, routing and retrieving — with the value often coming from doing something at a scale that was previously uneconomic. Customer-facing work needs a tighter exception path than internal work, because errors are visible and sometimes regulated. Resist starting with the fully automated chatbot: agent-assist keeps a human accountable and produces the edit history that later tells you exactly what is safe to automate.