Answer
How should AI be used in customer support?
To find the answer and draft the reply. Whether it sends is a separate decision about what the answer commits you to.
Use it to find the relevant answer in your own material and draft the reply. Whether it sends without a person depends on whether the message asserts anything the business has not already recorded.
Support has two hard parts and AI addresses one of them well. Finding the answer — in the documentation, the previous tickets, the policy, the product notes — is a retrieval problem and is where most of the time goes on routine queries. Deciding what the business is willing to do for this customer is a judgement about a commercial relationship, and it is not a retrieval problem however it is phrased.
So the reliable arrangement puts the system on the first and a person on the second. It reads the query, finds the relevant material, drafts a reply grounded in it, and shows the agent both the draft and the source. The agent's job becomes checking and deciding rather than searching and composing, which is a substantial saving and leaves the commercial judgement where it was.
Automatic sending is defensible for a narrow set: acknowledgements that commit to nothing, status updates drawn from a record, confirmations of something that has already happened. The line is whether the content is assembled from stored facts or generated. An assembled message asserts only what your systems already say; a generated one is a new statement your business is making, and it can contain a commitment nobody authorised.
The failure to design for is the confidently wrong answer on a question the material does not cover. Support queries include the ones your documentation never anticipated, and a system with no way to say so will produce something plausible for those too. Requiring the reply to be grounded in a retrieved passage, and routing to a person when nothing sufficient is found, converts the worst case from a wrong answer into a slightly slower one.
Two measurements are worth having from the start. How often the drafted reply is sent without material change, which tells you whether the drafting is actually working. And what proportion of queries had no adequate source, which is a list of the documentation you are missing and is usually more valuable than the support saving itself. Businesses that capture the second end up improving the material, which improves everything downstream.
One caution about tone. A support reply that is fluent, complete and slightly wrong is worse for the relationship than a brief human one, because it reads as a considered position rather than a mistake. Where a customer is unhappy, the value of a fast reply falls and the value of it being from someone who can actually resolve things rises, which is the specific case for routing rather than answering.
The support question is not whether the reply is good; it is whether the business would have said that.
Siddharth Sharma, Context Theory
Related questions
Should customers be told a reply was AI-assisted?
Where the reply is sent without a person reading it, saying so is honest and sets the right expectation about what it can resolve. Where a person reviewed and sent it, it is their reply and the tooling behind it is no more disclosable than a template. The distinction that matters to the customer is whether a person stands behind it.
What should happen with an angry customer?
Route immediately, and design that route explicitly rather than relying on the system to notice. Sentiment detection is imperfect and the cost of getting it wrong is concentrated exactly here, so a broad rule that routes anything containing complaint language is more reliable than a careful judgement about severity.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Firms that never responded to a web enquiry at all | 23% | Category-wide |
| Average B2B first-response time | 42 | Category-wide |
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · hours · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Support divides into finding the answer, which is a retrieval problem and consumes most of the time on routine queries, and deciding what the business will do for this customer, which is a commercial judgement that does not become retrieval when rephrased. | Timing the search and composition portions of a sample of routine support responses. |
| Constraint | A message assembled from stored facts asserts only what the business's records already contain, while a generated one is a new statement the business is making and can include an unauthorised commitment, which is the line for automatic sending. | Checking whether every sentence in a candidate auto-sent reply traces to a stored field. |
| Response | Requiring the reply to be grounded in a retrieved passage and routing when nothing adequate is found converts the uncovered-question failure from a confident wrong answer into a slower correct one. | Submitting a query outside the documented material and observing whether the system answers or routes. |
| Buying behaviour | The proportion of queries with no adequate source is a list of missing documentation and is frequently worth more than the support time saved, because improving the material improves every downstream use of it. | Recording queries where retrieval found nothing sufficient and grouping them by subject. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one