05AI implementation
Customer service AI implementation, without losing the customer
Support is the function with the loudest claims attached to it, so it is worth being plain about the limit before describing the work.
01
Start with the deflection ceiling
A well built support assistant does not close every ticket. It closes the repetitive ones. How much of your volume that turns out to be depends almost entirely on how repetitive your incoming questions actually are, not on anything the software adds.
If your tickets are mostly order status, password resets and opening hours, a large share of that volume never needs a person. If every ticket is a different technical problem in a different customer configuration, expect closer to none of it, and plan on that basis.
Anyone quoting you one confident percentage before reading your tickets is quoting you their marketing, not your business.
02
Deflection and routing are different jobs
Deflection
Deflection means the customer's question was fully answered and the ticket never reached a person. Real, measurable, and capped by how repetitive your volume is.
Routing
Routing means the ticket still reaches a person, but faster, in the right queue, with the history summarised and the account details already attached. Less impressive to talk about. Frequently worth more money.
03
This distinction decides whether the project is worth doing.
Routing has no ceiling, because every ticket benefits from it. Most of the value in a support build comes from routing, and most of the disappointment comes from a buyer who was sold deflection and measured only that.
04
One ticket, followed end to end
A customer writes in at 11pm about a charge they do not recognise.
The assistant reads it, identifies the account, checks the billing record, and finds a prorated charge from a plan change. It explains the charge in plain language with the dates, and asks whether that resolves it. If yes, the ticket closes and never becomes a human's morning. That is deflection.
If the customer says the plan change was not authorised, the assistant stops. It does not argue, does not guess, and does not offer a refund. It writes a summary, tags it as a billing dispute, attaches the account timeline, and puts it at the front of the queue for a person who starts at 8am. That is routing, and stopping is the correct behaviour.
05
The escalation rule
Escalation is not a fallback for when the assistant fails. It is a designed path with explicit triggers, and we write those triggers with you before anything goes live.
Standing triggers: the customer asks for a person, at all, ever. Anger or distress in the message. Money, refunds and disputes. Anything about health, safety or legal exposure. Two failed attempts on the same question. Anything outside the knowledge base.
The customer never has to work out the magic word. "Talk to a human" is on screen from the first message.
06
The knowledge base is the real project
The uncomfortable part. An AI support assistant is a reader. What it reads is your help centre, your macros, your past ticket resolutions and your policy documents. If those are out of date, contradictory or half written, the assistant will confidently repeat the mistake to every customer, at speed.
So the real work is usually not the model. It is auditing what your support team actually knows, writing down the parts that only live in someone's head, and building a routine that keeps it current as policies change. Companies that skip this get an assistant that is wrong at scale, which is worse than the queue they started with.
We do this part first. It is also the part that keeps paying off if you later change tools.
07
Support has more published cost and savings research pointed at it than any other function on this site, and nearly all of it turns out to trace back to a vendor selling the thing being measured, or to a number that cannot be found in the document it is credited to. We checked the figures originally planned for this page against their original sources. Most of them did not survive that check, so they are not on this page. The one that did is below.
08
The one published figure worth printing
That is a real, peer-reviewed number, and it is the only one in this section we would put the firm's name behind. It says support agents got measurably faster with help. It does not say what that is worth in your currency, and it is not a forecast for a deflection or cost-reduction project on your business. Every figure used anywhere on this site, including the ones dropped for not holding up, is listed with its publisher and date on the sources page.
15%
more issues resolved per hour by 5,172 customer support agents, at one company, given access to a generative AI assistant
The Quarterly Journal of Economics, Generative AI at Work (Brynjolfsson, Li and Raymond), vol. 140, issue 2, published . Read back .
The gain was concentrated among the least experienced agents; the most experienced agents saw small gains in speed and small declines in quality. This is a measured productivity result inside one company's support organisation. It is not a cost saving, and it is not presented as one here.
09
What must never be automated
No automated final decision on a refund, a cancellation or an account closure. No automated handling of a safety complaint, a medical question or a legal threat. No automated response to a customer in visible distress. No synthetic reviews or replies written in a customer's voice. And no bot that claims to be a person.
Questions
What should you know before the first call?
- What percentage of customer service tickets can AI actually handle?
- Published deflection rates for AI customer service vary widely because they depend on how repetitive the incoming questions are. Businesses with highly repetitive contact reasons sit at the top of the published range; businesses whose tickets are each technically different sit near the bottom. Any estimate made before reading a real ticket sample is a guess.
- What is the difference between deflection and routing in AI support?
- Deflection means the customer's question is fully answered without a person ever touching the ticket. Routing means a person still handles it, but the ticket arrives in the right queue with the history summarised and the account already identified. Deflection is capped by how repetitive your volume is. Routing improves every ticket, which is why it often delivers more total value.
- How do customers reach a human if an AI is answering first?
- A "talk to a person" route should be visible from the first message, and escalation should also fire automatically on requests for a human, on anger or distress, on money and refund questions, on safety or legal matters, and after two failed attempts at the same question. A customer should never have to find a magic phrase.
- Why does the help centre matter more than the AI model?
- An AI support assistant answers from your existing help centre, macros, policies and past ticket resolutions. If those are out of date or contradict each other, the assistant repeats the error to every customer at speed. Cleaning and maintaining that material is usually the largest part of the work and the part that keeps its value if tools change later.
- Which helpdesk platforms does this work with?
- Support automation is normally built on the helpdesk already in place, which in practice means Zendesk, Intercom, Freshdesk, Gorgias, Help Scout or Salesforce Service Cloud, joined to the order, billing or account system the answers depend on.
- What should never be automated in customer service?
- Final decisions on refunds, cancellations and account closures, anything touching safety, health or legal exposure, and any conversation with a customer who is distressed. Those carry consequences that a wrong answer cannot be walked back from, so a person makes the call.
Where this leads
This page is one part of the whole offer: automate the rest of the business too.
Start with a free consultation
Tell us what your customers write in about and what happens to a ticket at 11pm. We will tell you what can be answered automatically, what must reach a person, and what we would build first. The call costs nothing and ends with a plain list either way.
No obligation · a scoping conversation first