A realistic resolution rate for an AI support agent in 2026 is roughly 42 to 80 percent, not the 67 to 85 percent you see in vendor headlines. Where you land inside that band depends mostly on one thing: task complexity. One large vendor measured about 58 percent success on simple, well-defined tasks versus around 35 percent on complex, multi-step ones. The same vendors who publish a 75 to 85 percent figure are usually quoting their own ideal, heavily documented help portal, not a typical first deployment. So the honest answer is a range, and your job before you spend money is to forecast where in that range your own traffic will land.
This guide gives you that forecast, source-backed, so you can plan staffing, escalation, and ROI honestly. If you would rather we do it for you, see how we run AI customer support, but everything below is yours to use whether or not we ever talk.
What resolution rate should you actually expect?
Start with the verified band, not the brochure. Across thousands of real customers, resolution rates commonly land between roughly 42 and 80 percent. Intercom, which publishes a transparent methodology, cites about a 67 percent average across more than 40 million conversations from 7,000-plus customers, and backs a performance guarantee at the 65 percent level. Salesforce reports that its own help portal autonomously resolves about 75 to 85 percent of queries across more than a million conversations a year.
Notice the gap between those two facts. The 67 percent average is a broad population number. The 75 to 85 percent figure is one company's own, exhaustively documented, perfectly grounded help center. They are both true, and they are not the same promise. For planning, anchor on the population average and the wide band beneath it, not the showcase.
A simple way to read the landscape:
| Source | Reported resolution | What it represents |
|---|---|---|
| Verified field range | ~42 to 80 percent | Many real deployments, varied data and scope |
| Vendor population average | ~67 percent | 40M+ conversations, 7,000+ customers |
| Vendor showcase (own help portal) | ~75 to 85 percent | One mature, ideally grounded use case |
If a sales deck quotes you the top of that table as your expected outcome, treat it as the ceiling, not the forecast.
Why is the vendor 67 to 85 percent headline misleading?
Because the headline number is almost always measured on the cleanest possible conditions. A vendor's own help portal has perfect, current documentation, a narrow and well-understood set of questions, and engineers who tune it daily. That is how you get 75 to 85 percent. Your deployment, on day one, has messier knowledge, broader scope, and integration gaps. That is how you get 42 percent until you fix them.
There is also a definitional trick worth knowing. "Automation rate" and "resolution rate" are not the same thing, and the honest math is involvement rate multiplied by resolution rate. Only conversations the agent fully closed on its own should count as resolved. Procedure handoffs and deterministic scripted answers do not. If a vendor blends "touched" and "fully resolved" into one big number, the headline inflates while the real outcome stays flat. Always ask what is in the denominator.
None of this means the technology underdelivers. It means the published numbers describe the vendor's best case, and you need your own forecast for your worst-documented Tuesday.
What drives the difference between 42 and 80 percent?
Three things move your rate more than the choice of model:
- Task complexity. This is the dominant factor. The same vendor that hit 58 percent on simple tasks dropped to about 35 percent on complex, multi-step processes. The split between these two numbers is, in practice, your single best predictor.
- Data quality and grounding. An agent answering from a clean, current knowledge base plus live CRM and ticketing data resolves far more than one guessing from stale docs. McKinsey is blunt that value comes from rewiring the workflow so AI owns Level 1 end to end, paired with well-structured content, not from bolting AI onto the old process.
- Escalation design. A graceful handoff does not lower your resolution rate; it protects your CSAT on the cases the agent should never have tried. Trapping a customer in a loop instead of escalating is how a decent resolution rate still produces angry reviews.
Prefer to run it yourself? You can Hire AI Agents and put one to work today.
The practical takeaway: if your queue is mostly repetitive, documented Level 1 work, forecast toward the upper half of the band. If it is judgment-heavy and multi-step, forecast toward the lower half and plan your human capacity accordingly.
How do you forecast your own resolution rate before spending?
You do not have to guess. Pull your last few thousand contacts and run a quick, honest classification:
- Sort by topic and volume. Group contacts into inquiry types and count them. The big, repetitive buckets (order status, password resets, returns, billing, hours, basic troubleshooting) are your candidates.
- Label each bucket simple or complex. A bucket is "simple" if it can be answered from a documented policy or a single record lookup. It is "complex" if it needs multiple steps, judgment, or systems the agent cannot reach.
- Apply the benchmark rates. Use roughly 58 percent for the simple buckets and about 35 percent for the complex ones as a starting assumption.
- Weight by volume. Multiply each bucket's expected rate by its share of total volume and add them up.
The result is a defensible first forecast. If 70 percent of your volume is simple and 30 percent is complex, a blended starting estimate is around 51 percent (0.7 × 58 + 0.3 × 35). That is your conservative planning number. A well-grounded, well-scoped deployment will usually beat it, but you should staff and budget as if it will not until you have measured otherwise.
This is also why scoping the agent narrowly matters. If you point it only at the simple buckets and let the rest escalate cleanly, your measured resolution rate on the cases it actually handles goes up, even though the blended forecast above stays honest about total volume.
What does the resolution rate mean for staffing and ROI?
Resolution rate is not a vanity metric; it is a capacity plan. McKinsey's research suggests AI can address up to about 60 percent of addressable care volume and free productivity worth 30 to 45 percent of customer-care function cost. But "address" is not "resolve unattended," and that distinction is your staffing model.
Work it through with your own forecast:
- The share the agent resolves unattended is your forecast rate (say 51 percent on the example above). That is the volume your humans no longer touch.
- The rest still needs people, just fewer of them on repetitive work and more of them on the complex, emotional, high-stakes cases where humans remain strongly preferred.
- Your ROI is the freed human time plus 24/7 coverage, minus build and run cost. It is real (AI adopters report being far more likely to see high ROI), but it follows the resolution rate, so a number you inflated on day one will overstate the savings on day one too.
Forecast conservatively, staff for the forecast, then let measured improvement free up capacity over time. That sequence keeps a CFO happy and a support queue covered.
How do you push your resolution rate up after launch?
The biggest gains come from grounding and scope, not from swapping models. A reliable improvement loop:
| Lever | What it changes | Where it shows up |
|---|---|---|
| Clean, current knowledge base | Fewer wrong or "I do not know" answers | Higher resolution on existing scope |
| CRM and ticketing access | Agent can act, not just talk | Resolves transactional cases, not just FAQs |
| Tighter topic scope | Agent stops attempting what it cannot win | Higher rate on what it does handle |
| Designed escalation ladder | Bad cases leave gracefully | Protects CSAT, prevents loops |
| Weekly transcript review | Failure patterns get fixed | Compounding gains week over week |
Set a real exit bar before you launch: the resolution rate and CSAT you need to see on a narrow scope before you widen it. Then read transcripts every week, fix the recurring failure modes, and only expand scope once the current scope clears the bar. The documented large deployments grew exactly this way, one of them reaching 70 percent resolution across email and chat within twelve weeks by phasing channel by channel rather than launching everything at once.
So what number should you put in the plan?
If you need a single planning number today: assume your blended starting rate, weighted by your simple-versus-complex volume split, somewhere in the 40s to low 50s for a typical mixed queue, and treat anything above 70 percent as a target to earn, not a promise to make. Anchor on the 67 percent population average, respect the 42 to 80 percent band, and remember that the 85 percent showcase is a destination, not a starting line.
The honest version of this question is the one worth answering, because a forecast you can defend beats a headline you cannot hit. If you would rather skip the assembly, we plan, build, and run the voice and chat agents inside your systems, ground them in your data, design the escalation path, and operate them with measured resolution and CSAT. Book a free consultation below and we will forecast a realistic resolution rate for your own traffic before you commit a euro.