AI support results depend on your documentation and rules, not on the model
Wonderchat, a vendor of AI support software, published research in October 2026 on 5,848,078 support conversations from more than 1,000 businesses, collected between March 2023 and August 2026. Of those conversations, 2,794,330 could be scored for resolution. Among 687 businesses with enough volume to compare, all on the same platform, models and release, the share of conversations the AI resolved ran from 46% to 88%, with a median of 69%. Wonderchat attributes the spread to documentation coverage, system access and escalation rules. CustomerThink carried the announcement.
Does the model matter less than vendors imply?
Wonderchat's data shows capability improving for everyone while the spread between the best and worst operators stayed put, which points to how each business runs support rather than to the underlying model.
Resolution by the AI agent alone reached 75.3% in 2026, up from 56.7% in 2024. The share it could not complete fell from 43.3% to 24.7%. Across that period the distance between strongest and weakest deployments held roughly steady.
A constant gap is suggestive, not proof. The mix of businesses and question types also changed over those years, and the excerpt I read doesn't say how that was controlled for.
What is actually causing the unresolved conversations?
In a sample of 5,000 failed conversations, 58.5% failed because the answer was never documented anywhere the agent could reach. Outdated content explained only 1.2%.
The other categories matter too. 12.6% fell outside the scope the team had configured, and 8.3% were too ambiguous to answer without a clarifying question. A further 19.3% could not be attributed to any cause from the conversation alone.
The practical point is that auditing existing help articles targets the smallest bucket. Writing the missing answers targets the largest. The unattributable 19.3% is also a reminder that the classification is itself a judgement, and the excerpt doesn't say who made it.
How far should I trust a vendor-run study?
Treat it as a strong hypothesis, not an independent benchmark. Wonderchat chose the businesses, defined resolution and scored the conversations, using only its own customers' data.
Some questions are open. Only 2,794,330 of the 5,848,078 conversations could be scored, so a large share was excluded, and we aren't told whether those looked different. The 687 businesses are a subset of those with enough volume. The 2026 AI-alone figure of 75.3% sits above the 69% median across businesses, which suggests the two numbers use different weightings. Check which is which before quoting either.
The same caution applies to the vendor-supplied case studies we have run, such as the 8×8 and Premier Guarantee one. A well-documented method is not the same as outside verification.
What should I ask my own vendor?
Ask for your resolution rate and its exact definition, plus a breakdown of failed conversations by cause: missing content, out-of-scope, ambiguous, or unknown.
Also ask where you sit against a 46% to 88% range on comparable software, and whether the vendor can show an unanswered-questions report. If it can't, you can't tell a documentation problem from a configuration one. Ask who owns writing the missing answers, and what system access and escalation rules are currently set. Wonderchat names both as factors.
If your vendor answers with a model upgrade, ask for evidence that the upgrade moves your failures, not the average.
Frequently asked questions
What resolution range did the study find?
Between 46% and 88% across 687 businesses on identical software, with a median of 69%.
Does the study compare different vendors?
No. It covers a single platform, Wonderchat's, so it says nothing about how rival products perform.
Should I rewrite my existing help articles first?
Probably not. Stale content explained only 1.2% of sampled failures, against 58.5% for answers that were never written.
Sources: CustomerThink, reporting Wonderchat's research.
Research identifies documentation coverage, system access, and escalation rules as the key factors separating effective AI support from underperforming implementations…