How to Set Up AI Customer Service - What Week One Actually Looks Like (2026)

The first useful output of an AI support project is a number you can work out before you have a vendor. It is the share of your monthly ticket volume that already has a correct answer sitting in one retrievable place. That figure is your ceiling. Every resolution rate a salesperson quotes you sits above it, and the distance between the two is unpaid work that lands on your team.

Nobody sells that step, which is why nobody writes about it.

Everyone ranking for this question sells the answer

We ran the search on 31 July 2026 and read the first page. Lindy, Forethought, Decagon, Chatbase, LiveAgent, Duet and Atlassian all publish AI support software. The eighth result, Lillibolero, is an AWS consultancy that implements it for you. Not one page was written by somebody who had run a support queue and had nothing to sell at the end of it.

This shapes the advice in a predictable direction. A setup guide published by a platform begins at the platform. Lindy's page advertises a 15-minute sync with your existing tools and a 60-second signup. Lillibolero costs the whole project at three to four hours, of which gathering your support knowledge gets 30 minutes. Both are describing configuration and calling it setup.

Configuration is genuinely quick. It is also not the part that decides whether the thing works.

The audit that produces a number

Export a month of tickets. Sort them by the question being asked rather than by channel or tag, because tags describe how your team files things and questions describe what the agent will face. Take the top twenty questions by volume.

Now answer each one yourself using only your help centre search, the way a customer would. No internal wiki, no asking the person who has been there four years. Time each attempt. Then drop the question into one of four buckets.

Weight each bucket by ticket volume. The first bucket, expressed as a share of all monthly tickets rather than a share of the top twenty, is your day-one ceiling.

Working the arithmetic

The numbers below are an illustration rather than a measurement, but the shape of the result holds for any queue you run it against.

Take a queue of 4,000 tickets in a month. The top twenty questions cover 2,760 of them, which is 69% of volume. Of those 2,760, say 1,490 are answerable from a single passage, 620 need two or more articles, 520 are not documented anywhere, and 130 are documented incorrectly.

Your ceiling is 1,490 of 4,000. That is 37%. Merge the multi-article answers into single pages, which is a fortnight of editing and no engineering at all, and the ceiling moves to 53%.

Hold that against the quotes. Decagon's setup guide states that AI handles 70 to 80% of routine interactions. Fin publishes a 76% average across 12,000 customers on its own platform, with top performers at 80 to 84%. The illustrative queue cannot reach 70% on day one whichever of them it buys, and roughly 33 points of that gap are questions that are either unwritten or wrong. No model fixes an unwritten answer.

Which is the whole argument for doing this first. The audit costs an afternoon and it tells you whether the difference between a $0.49 platform and a $2.00 one is even reachable in your case.

Where the audit fails

It undercounts one category badly, and in retail that category is usually the largest. Where is my order has no document answer at all. No article can contain it, so every such ticket lands in the not written down bucket and drags your ceiling down, when in fact it is among the most automatable questions you have. It just needs an order lookup rather than a paragraph.

So score those separately. Anything whose answer comes from a system rather than a page goes in a fifth column, and that column is a shopping requirement instead of a documentation task. It also changes which platforms qualify: a tool that only reads help articles is no use to you regardless of how well it writes, and the integration work behind a live order lookup is engineering that quotes rarely include.

The distinction is worth making early because the two columns get fixed by different people on different budgets. One is a writer for a fortnight. The other is a sprint.

The vendors' own timelines already concede it

Decagon publishes a six-week implementation. Week one is discovery and workflow documentation. Week two begins converting existing SOPs into what the company calls Agent Operating Procedures. Two of six weeks, a third of the published project, happen before the agent speaks to a customer, and both are content work rather than software work.

Fin's ROI page is blunter, calling content preparation the single highest-leverage pre-launch activity and reporting that teams who review AI suggestions weekly gain 15 to 20 percentage points of resolution within 60 days. That swing is larger than the gap between most platforms on most shortlists.

Fifteen minutes to six months, for a job everybody agrees on the definition of. What separates the estimates is almost entirely whether they include your documentation.

Fix the stale bucket before the empty one

This is the ordering people get backwards. The instinct is to write the missing answers first, because a gap feels like the obvious problem. It is the wrong instinct.

An undocumented question produces an escalation. Annoying, visible, recoverable. A stale answer produces a confident, fluent, wrong response that the customer acts on, and you find out when the refund request arrives. The failure is silent until it is expensive, and your monitoring will not catch it because the conversation looks like a success on every dashboard you own.

So week two runs in this order: correct the stale bucket, merge the multi-article bucket into single pages, then write the undocumented answers. The third job is the biggest and it can wait, because until it is done those questions simply escalate, which is what would have happened anyway.

Pick the pricing model before the product

Rates published as of July 2026, each collected from the vendor or from Fin's own comparison page. The right-hand column is the same 10,000 billable events priced under each model.

Twenty times between the cheapest and the dearest published rate for the same volume. That spread is decided at signature and no amount of configuration changes it afterwards.

There is a floor worth knowing about too. Gartner puts a self-service contact at $1.84 and an agent-assisted one at $13.50. An AI agent billing $2.00 per conversation is charging more than the traditional self-service it displaces, and only beats the human number by being right often enough to avoid the escalation. If your ceiling is 37%, it will not be.

Escalation goes in before the first customer sees it

Put the route to a human in the first message rather than after three failed attempts. Your deflection figure gets worse and this is the correct trade, because deflection counts a customer who gave up as a win.

LiveAgent reports that 76% of customers who have to repeat information during an AI-to-human handoff rate the experience significantly worse. Repetition, not the bot itself, is what people remember.

Test the handoff, not the conversation

During evaluation, escalate on purpose. Then go and look at what the human agent actually receives: the transcript, the customer record, whatever the agent inferred, and whether any of it survived the transfer.

If the customer has to start again, you have added a step rather than removed one. Platforms differ sharply here and no demo has ever shown it, because the demo ends at the resolution.

Two numbers for the first fortnight

The first is confirmed resolution, meaning the customer said the problem was solved, tracked separately from deflection. The gap between them is your error rate wearing a disguise.

The second is the one nobody tells you to keep: resolution measured against your ceiling rather than against the vendor's benchmark. A queue with a 37% ceiling that resolves 31% is running at 83% of what its content permits. Buying a better model there buys you six points at most. Writing the missing answers buys you thirty.

Most teams never work out which of those two situations they are in, and so they upgrade the software.

We do not benchmark platforms. Every rate and resolution figure above is a published vendor or analyst claim, attributed where it appears. The audit arithmetic is worked from illustrative inputs, stated as such, and reproduces on your own export.

WRITTEN BY AR · UPDATED 2026-07-31

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

Editorial policy