Why this guide contains almost no statistics
Key Takeaways
For busy support leads: the question is not whether AI can answer your customers. It can. The question is whether you have documentation worth retrieving, a named person who owns that documentation, and an honest definition of a resolved conversation. With those three things, expect a working system inside two weeks. Without them, an AI layer just makes your documentation gaps louder and faster.
- 1Retrieval beats model choice. Which frontier model you use is close to irrelevant next to what the system retrieves before it answers.
- 2Run the arithmetic on your own numbers. Two inputs decide everything: minutes per human conversation and your loaded cost per agent hour.
- 3Escalation is the product. The exit door to a human is the feature customers judge you on, not the answers.
- 4The pricing model matters more than the price. Per resolution billing rises exactly as your automation succeeds.
- 5Budget maintenance from week one. An unmaintained knowledge base turns a good AI into a confident liar within a quarter.
Most AI support articles repeat the same handful of percentages. Sixty to seventy-five percent auto resolution. Thirty percent deflection. Eight to fifteen dollars per ticket. Chase any of them back to a primary source and the trail usually ends at another blog post quoting a third blog post. Earlier versions of this guide did the same thing, and we have deleted those numbers rather than dress them up with a link.
What is left is arithmetic you can run on data you already have, plus instructions for which inputs to measure. If a vendor quotes you a resolution rate, ask three questions before you write it down. Resolution measured how? Over which mix of conversations? Measured how long after launch? The answers are more useful than the headline number, and most vendors will not give you all three.
A small number of figures did survive that test, and they are here with their limits attached. The largest field study of AI in support tracked 5,179 customer service agents at a single software firm and found that access to an AI assistant "increases productivity, as measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers" (Brynjolfsson, Li and Raymond, NBER working paper 31161, 2023). One firm, one deployment, method published. That is what a usable number looks like, and note how much smaller it is than the numbers in the pitch decks.
What AI customer support actually is in 2026
Strip the marketing away and a modern AI support system is five layers stacked on top of each other. The model is one layer, and it is the layer you should worry about least. Understanding the stack tells you where your project will actually fail.
| Layer | What it does | What goes wrong |
|---|---|---|
| Ingestion | Pulls in help articles, docs, past tickets, product copy | Stale content, PDFs nobody parses, tribal knowledge that lives in Slack |
| Retrieval | Finds the passages most relevant to the question asked | Returns a near miss: a similar article that answers a different question |
| Generation | Writes a conversational answer grounded in the retrieved text | Fills gaps with plausible invention when retrieval returns nothing useful |
| Control | Decides tone, forbidden topics, and when to escalate | Boundaries left at defaults, so the AI answers refund policy questions it should not |
| Feedback | Routes bad answers back into the knowledge base | Nobody owns it, so the same wrong answer ships for months |
This is retrieval-augmented generation, usually shortened to RAG. The important consequence is simple. Your AI cannot be better than the documents it can reach. Teams who spend their evaluation period comparing models, and none of it auditing their own content, are optimising the wrong layer.
What is AI support actually worth? Run this arithmetic
Here is a worked example. Every number below is an assumption we are making up for illustration, not a benchmark. Replace each one with yours before you make a decision.
Assume 1,000 conversations a month. Assume each human-handled conversation consumes 9 minutes end to end, counting the reply, the follow-up, the context switch and the notes. Assume a loaded cost of $35 per agent hour, which is salary plus benefits plus tooling plus management, divided by hours actually worked rather than hours on payroll.
That gives 150 agent hours a month, or $5,250. Per conversation, $5.25.
Now assume AI closes half of them. The naive saving is 500 times $5.25, which is $2,625. That number is wrong, and here is the honest version.
| Line item | Assumption | Cost per month |
|---|---|---|
| Closed by AI | 500 conversations, no agent time | $0 |
| Escalated to a human | 500 at 10 minutes, the usual 9 plus 1 minute to read the AI transcript | $2,917 |
| Quality review | 50 conversations sampled at 2 minutes each | $58 |
| Knowledge base upkeep | 4 hours a month | $140 |
| Tool | Flat rate, $99 | $99 |
| Total | $3,214 |
Against a $5,250 baseline, that is roughly $2,036 a month, not $2,625. The missing $589 is the part vendor calculators skip: agents now read a transcript before they start, someone samples answers for quality, and someone maintains the content the whole thing depends on.
One more caveat, and it is the one that ends most business cases. Freed agent hours are capacity, not cash. They become cash only if you genuinely do not hire the next person, or if the freed hours go into work that earns something. If your team simply spends the recovered 66 hours a month on the same backlog with less pressure, that is a fine outcome, but do not put it in a budget as savings.
What you need in place before you turn it on
Three prerequisites, in order of how often their absence kills the project.
First, documentation that answers your top questions in the words customers use. Pull your last 200 conversations, cluster them by topic, and check that the top 20 topics each have an article. Write the articles around problems, not features. "Why was I charged twice" retrieves better than "Billing overview".
Second, an escalation path that is one click and always visible. Not buried, not gated behind three clarifying questions, not conditional on the AI deciding it has failed. Customers forgive an AI that cannot help. They do not forgive being trapped. When it hands over, it should pass its full transcript across and get out of the way, without asking the customer to repeat anything.
Third, a named owner. Not a team, a person. The single best predictor of whether an AI support deployment still works in month six is whether one identifiable human is responsible for the knowledge base.
One compliance note that stopped being theoretical this month. Article 50 of the EU AI Act became applicable on 2 August 2026 and requires that AI systems intended to interact directly with people are built so that "the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious" (EU AI Act, Article 50). Breaches sit under Article 99(4)(g), which sets fines of up to 15 000 000 EUR or 3 percent of total worldwide annual turnover, whichever is higher (EU AI Act, Article 99). If you serve EU customers, label the assistant.
Which conversations should the AI be allowed to answer?
Use one heuristic. If being wrong costs more than the conversation costs to handle, a human handles it. A wrong how-to answer costs a follow-up message. A wrong answer about a security incident costs a customer, and possibly a lawyer.
| Conversation type | Let the AI answer? | Reason |
|---|---|---|
| How-to questions with a documented answer | Yes | Cheap to get right, cheap to get wrong |
| Plan, feature and pricing questions | Yes, if pricing is public | Documented and stable |
| "Is the service down?" | Yes, wired to your status page | Volume spikes exactly when humans are busiest |
| Order or account status lookups | Yes, if it can read the record | Deterministic answer, high volume |
| Refunds and billing disputes | Draft only | Money plus emotion, needs a decision |
| Bug reports | No, route and summarise | Needs investigation and reproduction |
| Security, privacy and data deletion | No | Legal exposure, one wrong answer is expensive |
| Cancellations | No | Last chance to keep the customer |
| Anything with angry or distressed tone | No, escalate immediately | Correct answers do not fix anger |
Notice that "draft only" is a real setting and an underused one. The AI writes, an agent approves and sends. You get most of the speed with none of the exposure, and you get a training set of corrected answers for free.
Per resolution, per seat, or flat rate: which pricing model fits
The pricing model shapes your behaviour more than the sticker price does. Three exist, and each punishes a different kind of company.
| Model | Bill grows with | Fits | Punishes |
|---|---|---|---|
| Per resolution, as Intercom's Fin is sold | Automation success | Low or spiky volume, unpredictable months | High steady volume, and anyone whose success raises their bill |
| Per agent seat, as Zendesk is sold | Headcount | Small stable teams | Growing teams, and anyone wanting engineers or ops in the tool |
| Flat rate | Nothing | Predictable budgets, teams of any size | Very low volume, where any subscription beats zero |
The published units are worth reading side by side, with the date attached because these move. As fetched on 2 August 2026, Intercom prices its Fin agent at $0.99 per outcome, charged once per conversation and with no additional seat cost, while Zendesk's own explainer on outcome-based pricing states that "Zendesk charges $1.50 per automated resolution", a figure that appears nowhere on its actual pricing page. That page lists Support Team at $19 and Suite Team at $55 per agent per month paid yearly (Zendesk, fetched 2 August 2026). Two vendors, two meters, and one of them taxes your headcount while the other taxes your success.
Per resolution deserves the sharpest question. If it works, you pay more. Redo the arithmetic above at $1 per resolution, a round number for illustration rather than a quote: 500 resolutions costs $500 instead of $99, and net saving drops from about $2,036 to about $1,635. Then ask who decides what counts as a resolution. Usually the vendor. Per seat has the opposite problem: it prices your team out of adding the people who would use the tool best, which is why we argue for flat pricing and why we wrote a direct comparison against Zendesk.
A 14 day rollout that survives contact with reality
Days 1 and 2: import content and fix the top 20 gaps you found in the conversation audit. Do not skip to configuration. Content is the project.
Days 3 to 5: run in suggest-only mode against live conversations. The AI drafts, agents send. Log every draft an agent rewrote and why. This is your real accuracy measurement, and it costs you nothing but attention.
Days 6 to 8: fix what the rewrites revealed. Most fixes are content fixes, not prompt fixes. If you find yourself editing instructions to patch a factual error, you are papering over a missing document.
Days 9 to 14: go live on your lowest-stakes channel with escalation wide open, then review every escalation daily. Expand topic coverage only when a topic has been quiet for a week.
How do you measure AI support without fooling yourself?
Most reported resolution rates are inflated by definition, not by dishonesty. If "resolved" means the customer stopped replying, then everyone who gave up counts as a success.
That is not a strawman, and you do not have to take our word for it. Intercom's own billing documentation defines an "Assumed Resolution": "If a customer disengages from the conversation for 24 hours after Fin's last answer, it is considered an assumed resolution". Silence is scored as success, and on a per-resolution plan it is billed as one. Intercom also markets Fin as "averaging 76% across 12,000+ customers", a vendor claim published without any methodology and resting on that same definition. Ask any vendor, ourselves included, exactly which conversations sit inside their headline number.
| Metric | Honest definition | How it gets gamed |
|---|---|---|
| Auto-resolution rate | No human touched it, and the same customer did not open a new conversation on the same topic within 72 hours | Counting silence and abandonment as success |
| Escalation accuracy | Of escalations, how many an agent agreed needed a human | Escalating everything, which looks safe and helps nobody |
| CSAT for AI conversations | Rated by the customer, compared against your human baseline in the same period | Surveying only the conversations that ended well |
| Cost per conversation | All support cost, including tools and the AI tax, divided by all conversations | Excluding review and maintenance time |
| Repeat contact rate | Same customer, same topic, within 72 hours | Not measured at all, which is the usual case |
Track the repeat contact rate from day one even if you track nothing else. It is the single hardest number to fake, and it catches confidently wrong answers faster than any satisfaction survey.
Where this breaks
Near-miss retrieval is the most common failure and the hardest to spot. The customer asks about the Pro plan seat limit, retrieval returns the Business plan article, and the answer reads perfectly while being wrong. Nobody notices until a customer quotes it back at you.
Stale content is the slow failure. Ship a feature, forget the article, and the AI keeps confidently describing last quarter's product. Wire documentation updates into your release checklist or accept the decay.
Account-specific reasoning is the ceiling. Questions that need multiple lookups plus judgment, such as why a proration looks wrong, are not documentation problems and will not be solved by better content.
And the organisational failure that beats all of these: everybody watches it in week one, some people watch in week three, nobody watches in month three. That is when the numbers quietly stop meaning anything.
When AI support is the wrong choice, including ours
If you handle a handful of conversations a week, skip it. At 40 conversations a month and 9 minutes each, you are spending 6 agent hours. No tool saves enough there to justify the setup time.
If your support is mostly bespoke technical investigation, where every ticket needs logs and a reproduction, AI belongs in your agents' hands as a drafting and summarising tool, not in front of customers.
If you have no documentation and no intention of writing any, buy nothing. The tool will amplify the gap.
Corebee specifically is the wrong choice if you want usage-based billing that starts near zero, if you need on-premise deployment, or if your buying process requires a procurement-friendly enterprise contract. It is priced flat at $99 a month with unlimited agents, which is deliberately bad value at very low volume and deliberately good value at high volume with a growing team.
The five minute version
AI customer support is a documentation product wearing a chatbot costume. The model is commoditised, the retrieval is the differentiator, and the maintenance is the cost nobody quotes. Measure your own minutes per conversation and loaded hourly cost, do the arithmetic including the transcript-reading tax, and decide with your numbers instead of an industry percentage. Keep the door to a human wide open, sample answers weekly, and give the knowledge base an owner with a name.
Ready to test it against your own content? Start a free trial and run the suggest-only week before you let it talk to anyone.