Every AI change in your data stack, on the record.
AI-generated changes are merging into production pipelines faster than anyone can review them. MergeSight gives data leaders one standing view of what AI shipped, whether it was reviewed and tested, and how it correlates with data incidents.
MergeSight joins four things your data stack already produces — GitHub pull-request metadata, AI-authorship signals, dbt artifacts (manifest.json, run_results.json), and your data-incident feed — into one read-only view. That's AI change assurance.
Free single-repo monthly report · Read-only · Metadata only · No new workflow for your engineers
AI change assurance report
Sep 2026
github.com/harborview-grocery/warehouse
AI-authored changes 47
Human-authored changes 81
AI review coverage 76.6%
Human review coverage 95.1%
| PR | Author | Attribution | Review | dbt models | Tests | Incidents |
|---|---|---|---|---|---|---|
| #482 | d.okafor | Direct | Reviewed | 3 | 14 passing | — |
| #479 | ai-assistant[bot] | Direct | Reviewed | 2 | 9 passing | — |
| #476 | r.chen | Inferred attribution unverified | None | 1 | 2 thin | INC-0417 |
| #471 | ai-assistant[bot] | Inferred attribution unverified | Reviewed | 4 | 21 passing | — |
| #468 | m.silva | Direct | None | 2 | 5 passing | — |
AI code is merging with thin review — and nobody is counting
The numbers are public and blunt. Sonar's 2026 State of Code finds 42% of committed code is now AI-generated — and 96% of developers don't fully trust it. Faros telemetry across thousands of teams shows a 31.3% rise in pull requests merging with zero review. And LinearB's benchmark of 8.1 million pull requests found AI-authored changes accepted at 32.7% versus 84.4% for human-authored ones — after waiting 4.6× longer for a reviewer to pick them up.
Most review processes were designed for human-written code — pull request by pull request, reviewer by reviewer. AI-authored volume broke that math: more changes, thinner review, and no way to see, across a quarter, how AI-authored changes differ from the rest of your merge history.
So when an incident hits, the engineering leader's real question — "which AI-authored changes shipped, and were they reviewed?" — gets answered by hand: ninety days of merge data pulled into a spreadsheet, after the fact, if it gets answered at all. That is the job MergeSight takes over.
How it works
- Connect read-only. MergeSight joins the metadata your data stack already produces — GitHub pull-request metadata, dbt artifacts (
manifest.jsonandrun_results.json), and your data-incident feed. No warehouse access. Nothing written. - Attribute and join. Every change is classified AI-authored or human-authored from commit signals, then joined to review coverage, dbt test and documentation coverage, and the incidents that followed.
- Read the report. A monthly AI change assurance report for your team — shareable — plus a standing dashboard and coverage alerts in between.
What you finally get to see
-
The AI footprint on your pipeline
Every AI-authored change that shipped — which pull requests, which dbt models, who merged them — in one report instead of a week of merge-log archaeology.
-
Review coverage, split by authorship
Merges with no review, AI-authored versus human — the split generic delivery dashboards never give you, because they don't know what AI wrote.
-
Test and documentation coverage on AI-touched models
When an AI-authored pull request touches a dbt model with thin tests or missing documentation, you see it — before it becomes the incident you're explaining upward.
-
AI changes next to the incidents that followed
When a pipeline breaks, see the AI-authored changes that preceded it — correlation you can actually investigate, assembled for you instead of reconstructed from memory.
-
Attribution you can trust because it's honest
AI authorship is detected from commit signals — and labeled "attribution unverified" wherever it is inferred rather than direct. Decision support, not forensics.
-
Read-only and metadata-only
No warehouse access, nothing written, no new workflow for your engineers. The assurance view lives at the manager's desk while the team keeps shipping.
Pricing
Start free with one repository. Scale when the questions get bigger.
-
Free — $0
One repository. One AI change assurance report every month. The whole report, not a teaser.
-
Starter — $150/mo, billed annually
For small dbt-first data teams putting AI to work: the standing dashboard, coverage alerts, and the monthly report across your repositories.
-
Team — $300/mo, billed annually
For growing data teams where AI-assisted development is the default. Everything in Starter, sized for a bigger team.
-
Scale — $500/mo, billed annually
For larger data organizations joining AI attribution to incidents across the whole stack.
Every plan is priced per data team, billed annually. The free single-repo report stays free.
Questions engineering leaders actually ask
What exactly is in the free monthly report?
A complete AI change assurance report for one repository: every AI-authored and human-authored change that shipped, review coverage including merges with no review, test and documentation coverage on the dbt models those changes touched, the data incidents correlated with preceding changes, and attribution labels on every row. It's shareable — built to be forwarded to your VP or into your board deck, not buried in a dashboard.
What does "attribution unverified" mean?
AI authorship is detected from commit signals — co-authored-by trailers, tool telemetry where tools expose it. Detection is not proof: some AI code carries no marker, and markers can be edited. So wherever attribution is inferred rather than direct, every row is labeled "attribution unverified." That honesty is the point. Auditors and boards don't expect line-level certainty — they expect a factual, artifact-level record of what shipped and how it was reviewed.
Who is this for?
Data engineering managers, Heads of Data, and analytics engineering leads running dbt-first teams with AI-assisted code in the workflow — the people who own review policy and get asked the hard questions after an incident, at the quarterly review, or when the AI policy rolls out.
What is MergeSight not?
Not an audit. Not compliance certification. Not a legal attestation. MergeSight is decision support for engineering leadership — a factual, standing record of AI-authored change activity. It's also not an AI code reviewer (it doesn't comment on your pull requests) and not a data-observability platform (it doesn't watch your warehouse at runtime). It's the standing join between those worlds, at the manager's desk.
How is this different from engineering-intelligence dashboards and AI code reviewers?
Engineering-intelligence platforms are built and priced per developer seat for large engineering organizations — they measure delivery, not data. AI code reviewers work one pull request at a time. Data-observability tools watch incidents at runtime. None of them joins AI attribution, review coverage, dbt artifacts, and data-incident feeds into one standing view for a data team. That join is the product.
What does MergeSight connect to?
Read-only connections to the metadata your stack already produces: GitHub pull-request metadata and commit signals, dbt artifacts (manifest.json and run_results.json), and your data-incident feed. No warehouse access, no writes, nothing installed on your infrastructure.
Get your free single-repo report
Launching Q4 2026 — leave your details for more info when we launch, and we'll follow up with everything you need to connect your first repository.