July 2026

Releases Shipped

Versions are listed here as each July 2026 release ships.

Added

ML: Train Model and ML: Score workflow steps — train a machine-learning model on any table and score data with it, entirely inside a workflow. Train fits your choice of seven algorithms (logistic or linear regression, decision tree or random forest classifiers and regressors, or a gradient-boosted classifier) and writes a one-row model table that flows through the workflow like any other table — with the algorithm, parameters, feature list, and training metrics recorded in queryable columns. Score reads a model table and a data table and appends a prediction column. Mistyped algorithm names or hyperparameters are rejected when you save the step rather than being silently ignored. See ML: Train Model and ML: Score.
Alteryx machine-learning pipelines now convert to runnable steps — importing an Alteryx workflow that uses the Machine Learning tool family (Assisted Modeling) now produces native ML steps: the modeling chain fuses into a single ML: Train Model step carrying the original algorithm and settings, and each Predict tool becomes an ML: Score step. XGBoost models convert with a documented approximation note, and anything that can’t convert cleanly is flagged in the conversion report with the detail you need to finish it by hand. See Machine Learning (Assisted Modeling) Conversions.
PlaidCloud Managed Bucket document accounts — add fully managed file storage to Document in a single step. Choose PlaidCloud Managed Bucket as the account type, give it a name, and PlaidCloud provisions and runs the storage for you — there’s no cloud project, credentials, region, or storage tier to set up or tune. Available on the Business and Enterprise plans. See Add a PlaidCloud Managed Bucket Account.
Attach screenshots to support tickets — when you open a ticket or reply in Help & Support (or reply by email), you can attach screenshots and files. Images show inline in the conversation. See Getting Help.
Support console for agents and partners — support agents work a sorted, filterable ticket queue: reply, attach files, add internal notes, claim tickets, set priority, and escalate to PlaidCloud. See The Support Console.
Team notifications and escalation for partners — organization admins can route new-ticket alerts to a Slack, Teams, or email channel, email the ticket owner on replies, and escalate tickets left unanswered too long. See Managing Support.
Analysis paths for AI questions — give the tables you analyze most a friendly name, and set one as your default, so you and your AI assistant can ask about “the Operations Results” — or just “what changed last quarter?” — without naming the project and table every time. Works across Microsoft 365 Copilot and any MCP-connected agent. See Analysis Paths.
AI cost-tracing now tells you how far to trust each answer — when you ask an MCP-connected AI agent why an allocation result changed, the answer now carries a plain-language confidence level and caveats. It flags when a cause is a mix of overlapping factors, when the answer rests on current hierarchy or driver weights, or when part of a change may just be an unfinished data load — so a plausible-but-shaky number is called out before you act on it rather than buried in the prose, and the trust signals travel with the answer even when a follow-up agent summarises it. An optional, administrator-enabled check can also lower confidence and flag an answer when the pool or driver data behind it hasn’t been reloaded recently. For results built from several allocation branches, the confidence now weighs each branch by how much of the change it accounts for, so one small, shaky branch no longer drags down an otherwise-clean answer. See Tracing Allocations with an MCP-Connected AI Agent.
AI allocation cost-tracing over the REST API — the full root-cause analysis (the plain-language headline plus the year-over-year baseline, anomaly, and data-freshness summary) can now be driven directly from the analyze REST API, not only through an MCP-connected AI agent.
Data Preview drawer in the Visual Workflow Designer — see a step’s output data in place, without opening a new window. Hover a step and click the browse glyph (or right-click → Preview Output Data, or the eye icon on an Inspector output card) to dock a preview of the table’s first 100 rows — typed columns, row and column counts, a selector for multi-output steps — at the bottom of the canvas. With the drawer open, clicking another step retargets the preview, the drawer refreshes itself when the previewed step finishes a run, and Open in Table Explorer escalates to the full explorer in one click. See Preview Step Data.
Directory Listing workflow step — write a table listing the files in a document directory, with an optional file pattern (like *.csv) and an option to include subdirectories. Use it to drive downstream steps from whatever files are present. See Directory Listing.
Row Count Assert workflow step — a data-quality gate that fails the workflow when two tables have different row counts, with a customizable failure message. Use it to confirm a transform kept every row or that a reconciliation didn’t drop or duplicate records. See Row Count Assert.

Changed

What-if (“forward impact”) estimates are now real per-target numbers instead of upper bounds. A hypothetical change is split across the results that share a driver by each one’s current share of the pool, so a fan-out’s figures add up rather than each showing the whole change, and you can scope the estimate to a single target. Pointed at a driver or basis table rather than the input pool, the agent now declines the estimate and explains why.
Dimension-hierarchy allocations are now explained rather than left as “can’t attribute this.” At the whole-total level the change is attributed to the input pool; filtered to a specific target, it separates how much came from the pool moving versus that target’s share of it shifting.
Results produced by more than one allocation step are handled gracefully — the agent lists the candidate steps and lets you pick which to trace, or, when a filter narrows to rows only one step produced, traces that step automatically and tells you it did.
Data-freshness checks now include the driver and basis tables, not just the input cost pools, so a stale driver that would silently skew a split is surfaced instead of going unflagged.
A year-over-year comparison with no prior-year data now says so — reported as no baseline available rather than labelling the change “normal”.
Packaged report macros are now first-class Report steps — a report bundle imported from an Alteryx packaged report macro now appears as a hand-editable Report step, carrying its macro definition automatically and rendering through the same path as the other Report steps. See Report: Packaged Macro.
Convert AMP-format .yxdb files for import — Alteryx .yxdb files written by the AMP engine use a proprietary format no tool outside Alteryx can read. A new guide and a downloadable Designer workflow batch-convert a folder of .yxdb files to CSV so their data imports into PlaidCloud losslessly. See Convert .yxdb Files for Import.
Importing an Alteryx workflow now ends on a conversion report instead of a plain “imported” message. It summarizes how many steps mapped with high confidence and how many need review, and lists the lower-confidence or caveated steps with their operation and notes so you can check them before running the workflow. Each imported step also carries its mapping confidence as a memo on the canvas. See Migrate Alteryx Workflows.

Fixed

Anomaly detection now ranks a period’s change against the history of period-to-period changes (so an unusually large swing is flagged), and orders history by real time first on models that label periods by quarter (for example “Q4 2025”) so periods aren’t mixed across a year boundary.
Cost-tracing no longer errors on models whose column names contain spaces, parentheses, or a percent sign (for example “Driver Value” or “Margin %”).
The full analysis no longer errors when a table is given by its friendly name instead of its internal id — the baseline and anomaly summaries resolve the name either way.
Cost-tracing now reads live column names for tables whose cached metadata is empty (for example results built by a model-run step or a project copy), so per-slice attribution works on those models too.
Corrected a dimension-based allocation being mislabelled as a “from zero” jump when its weighting basis happened to start empty, and a percentage change carrying the wrong sign when the prior value was negative.
Cost-tracing now runs on StarRocks-backed workspaces as well as Databend and Postgres.

Removed

Retired legacy import and export steps — the SAS7BDAT import, HTML import, and HTML export steps have been removed. Use the current import and export steps for these formats instead.