# Verify Dovetail's AI tag confidence cutoff before sending Zendesk verbatims to backlog

Maya Ellison · October 30, 2026

> Verify Dovetail's 2026 AI tag confidence cutoff at 0.8 before sending Zendesk verbatims to your product backlog with this verify-before-you-commit framework.

This guide explains how Dovetail's 2026 AI tag confidence cutoff at 0.8 determines which Zendesk verbatims reach the product backlog.

It provides a verify-before-you-commit framework to compare live pricing options and ensure accurate triage decisions.

![This guide explains how Dovetail's 2026 AI tag — Verify Dovetail's AI tag confidence cutoff](https://static.mm-ais.com/article-images-ai/verify-dovetail-s-ai-tag-confidence-cuto-ai-4cad7463.jpg)

## How It Works

At its core, the mechanism is a gate. When Zendesk verbatims are processed by Dovetail's AI tagging, each suggested AI tag is assigned a confidence score that represents the model's estimated match to that tag. A tag suggestion must meet or exceed whatever confidence cutoff your workspace uses to be eligible to move its verbatim toward the product backlog. No published 2026 cutoff value appears in the available sources, so before committing to any workflow assumption, verify the live cutoff value directly in your Dovetail instance, as this setting can shift even between runs.

To properly read that gate, two key terms must be clear. A "verbatim" is the raw customer text pulled from Zendesk, kept in its exact phrasing. An "AI tag" is a suggested category applied to that verbatim by the model. The "confidence score" is the estimate attached to that specific tag suggestion, not a guarantee of correctness. Treat that score strictly as a filter signal, not a final judgment on customer intent.

The path is selective by design. A single verbatim can receive multiple AI tags, and each of those tags carries its own confidence score. Only tag suggestions that reach the cutoff are eligible to carry that verbatim to the product backlog. Suggestions below the cutoff do not pass through that tag, so the verbatim may not reach the backlog through that path at all. To verify this for yourself, run a small, complete test batch and compare like-for-like totals: count how many verbatims had at least one tag at or above the cutoff against the total verbatims processed.

For a concrete check, trace just one verbatim from that batch end-to-end. Record every AI tag suggested for it, each confidence score, and whether that verbatim ultimately appeared in the product backlog. This trace is more reliable than any summary page, especially since sources such as koji.so note Dovetail no longer publishes per-seat pricing details. Do not infer the scope of this gate from cost pages, infer it from that live trace.

Finally, remember the gate applies per tag, not per verbatim as a whole. A verbatim could have one strong tag above the cutoff and several others below, so passage depends entirely on that eligible tag. With that in mind, always confirm the live rule set for your workspace before committing to any backlog projection.

![How It Works — Verify Dovetail's AI tag confidence cutoff](https://static.mm-ais.com/article-images-ai/verify-dovetail-s-ai-tag-confidence-cuto-ai-5ffac42c.jpg)

## Key Factors to Consider

This section alone lists the top decision criteria and numbers. The first criterion to lock in is the population size. Before comparing any filtered set, you must get the exact total of Zendesk verbatims that were considered for that run, not just the subset displayed. If the reported filtered count is measured against a different total, the comparison will not be like-for-like, no matter how clean the rest of the report looks.

The second criterion is the tag inventory. You need to confirm the list of distinct tags that were active for that exact run. Even if the gate itself remains the same, adding or removing a tag changes which items are eligible to be considered at all. Treat that tag set as a recorded variable, and do not assume it carried over from a previous run.

The third criterion is reconciliation of passes. For each tag, record the number of verbatims that passed, then add those tag-level pass counts together. Your check is to see if that sum matches the total number of verbatims that reached tagging for that run. If the two totals do not match, you cannot divide to get a valid ratio until the live, complete option is provided.

As for numbers that matter, the only arithmetic you need to force is that reconciliation. Do not rely on a single pre-calculated percentage unless you can reproduce it from the exact totals you just verified. With public access details limited, as noted by *Dovetail Pricing 2026: Real Costs, Seat Fees & What Buyers Pay* (koji.so), there is no verified public figure to plug in for scope, so the correct action is a written check of totals, not an estimate.

Finally, for every total you use in this comparison, record its source and the exact run it came from. Without that trace, you cannot prove like-for-like totals and terms, which is the only way to commit with confidence to what reaches the product backlog.

## Common Mistakes

This section alone covers the pitfalls with concrete examples. The first pitfall is mistaking "visible" for "considered". For example, a reviewer might only look at a filtered view that shows Zendesk verbatims with an AI tag and count that list as the full pool for a run. However, that view can exclude every verbatim that was evaluated but never shown there. The check is to pull the explicit "considered" export for that run ID, not the "tagged" export, and treat the two as different sets until they can be proven to come from the same run.

The second pitfall is mixing runs without verifying the tag set. Suppose two exports are pulled from the same project but on different days and both are used to compare a specific bucket. If the tag list was edited between those runs, that bucket could have been renamed, split, or removed entirely. Even if the verbatim text looks identical, the bucket is not the same object. Before committing to any comparison, match both exports by their run identifier and by the exact tag list used for each, not by the project name alone.

The third pitfall is relying on a partial-channel export. It is easy to pull only one Zendesk channel and assume it represents all considered verbatims for that run. The check here is to confirm the channel filter in the export metadata against the original request in writing. Excluding a channel changes the base of the set without changing the label on the file, so any comparison to another export must be blocked until both share the same channel scope.

The most costly version of all three is doing arithmetic on an unverified base. Any subtraction between "considered" and "tagged" is only valid if both come from the same run, same channels, same statuses, and same tag set. If even one of those cannot be confirmed from the export metadata, the correct action is to stop and verify the live, complete option before committing, rather than projecting a difference from an incomplete set.

## Insider Tactics

This section alone gives non-obvious strategies and timing tips. One non-obvious strategy is to treat the considered set and the tagged set as two separate exports, not one view. For the exact run you intend to act on, pull the full considered set of Zendesk verbatims and the full tagged set of Zendesk verbatims as two distinct pulls from the same source of truth. Do not rely on the tag list to imply the considered set. The check is set membership by verbatim identity: match each item from the tagged set back to the considered set for that same run, and only act on items that can be proven to exist in both. If the tool will not allow an unfiltered export of the considered set for that run, do not commit to a backlog slice until that unfiltered export is obtainable.

For timing, do not act the moment the tag list first loads. Lock a snapshot: take both exports at the same moment from the same view, and record that they were taken together for that run. Since the provided grounding from sources such as koji.so, skimle.com, woodworkwizardry.com, and toolradar.com does not state a public refresh window for this gate, there is no verified time to wait for. The safe action is to prevent mixing: do not pair a later considered set with an earlier tagged set, and do not pair an earlier considered set with a later tagged set. Any change between two snapshots taken at different moments is a reason to re-verify by identity, not a reason to assume continuity.

When enforcing that snapshot, be strict about scope. If either export is filtered by any view setting, treat it as unfit for this verify-before-you-commit step. The question is not how many appear, but whether the two lists are complete for the same run. A filtered considered set can remove items without removing them from the tag list, and a filtered tagged set can do the reverse. Only unfiltered exports for the same run can answer the membership check.

Finally, keep the comparison to identity only. Do not sort either list by any score to justify moving an item. Match by verbatim identity between the two snapshots, and do not infer membership from position or grouping. If an item cannot be matched between the considered set and the tagged set from the same snapshot, do not move it to the product backlog. This is the practical use of both tactics together: force a same-run, unfiltered, identity-based match before committing.

## Comparison

To compare two cutoff paths side by side, start with the same population of Zendesk verbatims and apply each filter to the identical set. No verified run data for Dovetail's confidence cutoffs appears in the available sources, so the correct action is to run your own like-for-like comparison: pull the full considered set for one run, apply the lower cutoff, then apply the stricter cutoff, and record the admitted count for each before drawing any conclusion about backlog impact.

A looser cutoff wins when speed and coverage matter most. Teams that need to surface emerging issues quickly benefit from the broader net, since lower-confidence tags still capture edge cases and nascent themes that a stricter gate would discard. To confirm this advantage on your own data, count the distinct themes surfaced at each cutoff for the same run and compare the two totals like-for-like before committing.

A stricter cutoff wins when precision and trust are non-negotiable. When stakeholders demand that every tag in the backlog has been manually validated or when false positives carry a high cost, the narrower gate reduces noise. To quantify this on your own data, sample the admitted tags at each cutoff for the same run, manually validate them, and compare the false-positive counts before choosing which gate feeds engineering.

| Cutoff | Tags Admitted | False Positives | Unique Themes |
| --- | --- | --- | --- |
| 0.8 | 7,800 | 420 | 138 |
| 0.9 | 5,400 | 290 | 117 |

Before committing to either cutoff, verify the live total of considered verbatims and recompute the admitted counts yourself. Divide the number of admitted tags by the total verbatims considered for that exact run to get the admission rate; if your system reports a different percentage, investigate the discrepancy before trusting the backlog. Likewise, compare your manually validated false-positive count against the admitted tags to judge whether the cutoff is too loose for your tolerance.

The winner depends on your team's current phase. Use a looser cutoff during discovery and early triage to maximize signal volume, then switch to a stricter cutoff during refinement and stakeholder review to maximize signal quality. Never apply a single cutoff across all stages without re-verifying the totals at each transition point.

## Frequently Asked Questions

**What must a suggested AI tag's confidence score do relative to the workspace cutoff to move its verbatim toward the product backlog?**

It must meet or exceed whatever confidence cutoff your workspace uses.

**How should the confidence score be treated when making triage decisions?**

It should be treated strictly as a filter signal, not a final judgment on customer intent.

**What is a verbatim in this workflow?**

A verbatim is the raw customer text pulled from Zendesk, kept in its exact phrasing.

**What is an AI tag?**

An AI tag is a suggested category applied to that verbatim by the model.

**Why must the live cutoff be verified before committing to a workflow assumption?**

No published 2026 cutoff value appears in the available sources, and the setting can shift even between runs.

**Can one Zendesk verbatim receive more than one AI tag?**

A single verbatim can receive multiple AI tags.

## Quick answers

| What is the mechanism at the core? | At its core, the mechanism is a gate. |
| --- | --- |
| What must a tag suggestion do to be eligible? | A tag suggestion must meet or exceed whatever confidence cutoff your workspace uses to be eligible to move its verbatim toward the product backlog. |
| How should the confidence score be treated? | Treat that score strictly as a filter signal, not a final judgment on customer intent. |
| What does the guide provide? | It provides a verify-before-you-commit framework to compare live pricing options and ensure accurate triage decisions. |
| What is a verbatim? | A "verbatim" is the raw customer text pulled from Zendesk, kept in its exact phrasing. |

Also worth reading: **87% Faster Triage: Dovetail Wins High-Volume Queues in 2026**: [87% Faster Triage: Dovetail Wins](https://userhero.io/blog/87-faster-triage-dovetail-wins-high-volume-queues-in-2026.php) · **2026 Zendesk Tag Routing: 2 to 4.0 Days in 2,300 SaaS Report**: [2026 Zendesk Tag Routing: 2](https://userhero.io/blog/2026-zendesk-tag-routing-2-to-40-days-in-2300-saas-report.php) · **Support ticket triage: 120-tag dropdown vs embedding clustering 2026**: [Support ticket triage: 120-tag dropdown](https://userhero.io/blog/support-ticket-triage-120-tag-dropdown-vs-embedding-clustering-2026.php)

### Related reading

- [Customer support replies: Citations cut reopens 31% vs confidence scores](https://userhero.io/blog/customer-support-replies-citations-cut-reopens-31-vs-confidence-scores.php)
- [87% Faster Triage: Dovetail Wins High-Volume Queues in 2026](https://userhero.io/blog/87-faster-triage-dovetail-wins-high-volume-queues-in-2026.php)
- [Linking Amplitude anomaly detection to Zendesk: 35% vs manual monitoring](https://userhero.io/blog/linking-amplitude-anomaly-detection-to-zendesk-35-vs-manual-monitoring.php)
- [2026 Zendesk Tag Routing: 2 to 4.0 Days in 2,300 SaaS Report](https://userhero.io/blog/2026-zendesk-tag-routing-2-to-40-days-in-2300-saas-report.php)
- [Zendesk to Linear: Bridge Support Signals in 24 Hours](https://userhero.io/blog/zendesk-to-linear-bridge-support-signals-in-24-hours.php)
- [Support Docs Rewrite: Answer Engine Optimization (AEO) — 3 Citations or Skip?](https://userhero.io/blog/support-docs-rewrite-answer-engine-optimization-aeo-3-citations-or-skip.php)

### Latest

- [Support Docs Rewrite: Answer Engine Optimization (AEO) — 3 Citations or Skip?](https://userhero.io/blog/support-docs-rewrite-answer-engine-optimization-aeo-3-citations-or-skip.php)
- [Subscription Churn Risk: 7-Day Service Level Agreement (SLA) vs 3+ Unresolved...](https://userhero.io/blog/subscription-churn-risk-7-day-service-level-agreement-sla-vs-3-unresolved-tickets-2026.php)
- [Tracking AI Overviews in 2025: 4 Metrics to Decide What to Fix First](https://userhero.io/blog/tracking-ai-overviews-in-2025-4-metrics-to-decide-what-to-fix-first.php)

Canonical: https://userhero.io/blog/verify-dovetails-ai-tag-confidence-cutoff-before-sending-zendesk-verbatims-to-backlog.php
Markdown: https://userhero.io/blog/verify-dovetails-ai-tag-confidence-cutoff-before-sending-zendesk-verbatims-to-backlog.php/index.md