In June, Anthropic's data science team published a post with numbers most data leaders have only seen in vendor slideware: 95% of the company's business analytics queries now run through Claude, at roughly 95% accuracy in aggregate (Anthropic, 2026).
If you run a data team, or you fund one, it is worth ten minutes. Not for the headline number, but for how honest the post is about what made that number possible, and about the one failure they say they still cannot reliably catch.
Read this as a pairing, not a rebuttal. We build Modern Gravity to run with frontier models like Claude, not instead of them: the model does the reasoning, and Gravity organizes and certifies the enterprise knowledge it reasons over. Anthropic's post proves the engine works at company scale. This post is about the layer that makes its answers safe to act on.
So here is the thesis: what buyers should demand next is simple to say and hard to build. The right number every time one exists, and never a wrong number dressed up as a right one. Delivering that takes a verification layer built as a product, not a footer.
The receipts worth reading twice
The Anthropic post is unusually specific about what worked and what did not. Four of their findings deserve to be pinned to the wall of every data team:
Accuracy without curated context was 21%. Pointing a frontier model at a warehouse and letting it run produced wrong answers four times out of five on their own evals. Curated context, in their case a skills layer over governed data, lifted it above 95%.
Documentation decayed from 95% to 65% accuracy in one month. Their docs described a data model that changes daily. Untended, the system quietly got worse every week until they wired doc maintenance into the same pull requests that change the data models.
Letting the model write the metric definitions made things worse. In their words, it "encoded the very ambiguities we were trying to eliminate, and was net-negative on our evals versus a smaller, human-curated layer." Their rule now is that the model drafts documentation and a human owns the definition.
Access was not the bottleneck. They gave the agent full read access to thousands of historical queries. Accuracy moved by less than one point, even though the right answer was present about 80% of the time. The agent saw it and still did not use it. Structure, not access, was the constraint.
Why analytics is harder than code
The post makes an argument we have been making for a year, and makes it cleanly: data is not software.
A coding agent works in an open solution space where tests and compilers catch most hallucinations. An analytics agent usually has exactly one correct answer from exactly one correct source, and no compiler exists to prove it found either. Their three failure modes: the agent picks the wrong entity for the concept, the knowledge has gone stale, or the right answer exists and retrieval misses it.
Every one of those failures produces output that looks finished, which is the trap. A wrong SQL join does not look wrong. It looks like a number, formatted correctly, delivered confidently.
Their playbook, our layer
I will be direct about why this post got my attention. Modern Gravity is our answer and verification layer for enterprises, and its promise is the delivery-first one above: the right number, consistently, with the evidence attached. Customers bring their own enterprise LLM, Claude very much included. The division of labor is clean: frontier models reason, Gravity grounds and certifies what they reason over.
Anthropic's post reads like external confirmation of bets we were already making. Some of them, like the certified vault and answers that carry receipts, predate their post in our own decision log. Others we locked in the weeks after it published, and the post sharpened our conviction. We are not claiming we got there first on every point. The post is useful either way.
| Anthropic's published practice | What Gravity ships |
|---|---|
| A human owns every metric definition; the model drafts docs | Every certified metric is anchored to a named business owner with a documented source contract |
| Agents are structurally required to hit the governed semantic layer first | Certified sources are consulted first, structurally, on every question |
| Staleness treated as an engineering problem with enforcement | A stale source contract cannot certify an answer; freshness failures close the gate rather than passing quietly |
| A provenance footer on every response | A visible assurance label on every answer: safe to use, draft, or deferred with the reason shown |
To be clear about what this table does not say: it says nothing about what Gravity scores on its own benchmark. That number stays internal until it is independently audited, and we explain why below.
The unsolved problem, and the promise it points to
Now the part that matters most. Anthropic names the failure they have no robust solution for: the silent wrong answer. Plausible, confident, wrong, and used without objection. Their mitigations are a provenance footer, human sign-off on anything leadership-bound, and daily sanity checks on top KPIs. And they say, plainly, that the footer "doesn't make the answer more correct."
Sit with the aggregate number for a second. Roughly 95% accuracy means about one answer in twenty is wrong, and nothing in the system tells you which one, or whether the wrong one was a lunch-spend rollup or the revenue figure in Thursday's board deck. Plenty of companies could genuinely live with a labeled five percent. Unlabeled, every answer inherits the doubt, and the error rate stops being a statistic and becomes a trust problem.
Read that sentence of theirs again, because the best-resourced AI data team in the world, operating on their own models, is saying the open problem in AI analytics is knowing when the answer is wrong before someone acts on it.
Gravity's answer to that problem is a delivery promise. Most analytics AI guesses. Gravity actually answers your business questions, with receipts. Numbers are verified against certified sources before an answer earns a board-ready label, so what reaches you is a number you can act on, with the evidence attached.
The proof that the promise is real is what happens at the edge: when a number cannot be verified, Gravity does not guess. It defers, and shows you the receipt for why. In demos, that deferral screen is often the moment buyers lean in, because everyone in the room has been burned by the confident wrong number before.
Curious what a verified answer looks like next to a deferral? See how Gravity answers with receipts.
How we know it holds: nearly 45,000 scored answers
Claims like that need their own receipts, so here is ours. We test Gravity against GravityBench, our internal benchmark for grounded business analysis, built the way Harvey built its benchmark for legal AI: real business questions, known ground truth, and scoring that punishes confident fabrication instead of rewarding fluency.
The current base is 96 benchmark runs producing nearly 45,000 scored answers to real business questions, accumulated between April and July of this year. GravityBench treats a fabricated figure as a failing result no matter how polished the rest of the answer looks, and it scores an honest "I cannot verify this" as a legitimate outcome rather than a loss. In front of a board, it is one.
We are deliberately not publishing accuracy figures from those runs today. Our benchmark governance requires an independent audit before any score leaves the building, and we hold ourselves to that. A public task subset and third-party audit are part of GravityBench's rollout; this post reflects where that process stands today. The volume is the point here: this layer is tuned against tens of thousands of scored answers, not a demo script.
The build-versus-wait question
Anthropic's post ends with questions for teams starting this work, and the first one is quietly loaded: how important is a correct answer today versus in the future? Their point is that some infrastructure becomes unnecessary as models improve.
Our view, stated as our view: the trust layer is not in that category. Better models make generation cheaper. They do not tell you whether the answer in front of you is safe to take to a board meeting, because that question depends on your definitions, your sources, and your owners, none of which a model can know on its own.
Read as an adoption guide, their post is also a bill of work. To run this playbook yourself, their own account says you build four things. Grounding: governed data foundations with canonical datasets and aggressive deprecation of near-duplicates. Context: the sources of truth around them, from a human-curated semantic layer to the business context most teams skip. The agent layer: skills and reference docs that, by their measurement, decay from 95% to 65% accuracy within a month unless maintenance is wired into engineering. And verification: offline evals, adversarial review they measured at a 6% accuracy gain for 32% more tokens and 72% higher latency, and continuous correction harvesting. Four workstreams, each with a named owner and a permanent maintenance loop.
That is the work a customer is signing up for, per Anthropic, and it is the work Gravity delivers as a product. Their post depicts a sustained, multi-layer internal build: one ablation experiment alone, in their words, "redirected months of roadmap," their skill docs decayed for a month before maintenance became an engineering problem, and the silent wrong answer is still open at the end of it. Getting Gravity's verification layer to its current state took us months of full-time build on top of an existing certified-data foundation, and that is our honest reference point for what a team building this internally should expect.
Where this leaves you
We said at the top that buyers should demand the right number every time one exists, and never a wrong one dressed up as right. Anthropic's own post is the proof of where the market stands: 95% automation achieved, silent wrong answers still open.
The strongest analytics stack we know how to build is exactly the pairing this suggests: a frontier model like Claude doing the reasoning, with a certified trust layer like Gravity underneath it so every answer arrives labeled, verified, and safe to act on. Their post is the best public evidence yet for why you want both.
If you own analytics, data, or the P&L that depends on them, the Monday-morning question is concrete: when your AI analytics surface produces an answer that looks right but is not, what in your stack catches it before someone acts? If the answer is "a person, hopefully," you have found your gap.
Read their post: Anthropic, 2026. Then try Modern Gravity to see every number carry its receipt.
And a genuine question for the operators reading: what does your team do today when an AI-generated number looks plausible and is wrong? I am collecting real answers, not hypotheticals.
Results referenced are based on observed patterns in available data and are not guarantees of performance. Individual outcomes vary based on data quality, implementation, and market conditions. Modern AI recommends independent validation of any metric before business decisions are made. Figures attributed to Anthropic are from their published post of June 3, 2026, and are their reported internal results.