Make your analytics organization agent-ready.
Your company already has most of the knowledge an analytics agent needs.
It's spread across data models, dashboards, metric definitions, analyses, documentation and decisions.
Propperly reconstructs what your analytics organization has actually concluded — what is current, what has been superseded, where sources conflict, and what remains unresolved — then helps keep it current as the organization changes.
No platform migration. No production integration required to start.
Active customer = logged in within 30 days
Move active-customer window to 28 days
decisionStill calculated using 30 days
conflict3 dashboards still pending
openStart small. Then build it and keep it current.
The model isn't the whole system.
Anthropic recently described what it took to run most of its own business analytics through Claude.
But there was a much bigger gap underneath those numbers.
Without the procedural Skills Anthropic built around its data environment, Claude's accuracy did not exceed 21% in its evaluations.
With those Skills, Anthropic says accuracy consistently exceeded 95%.
Anthropic's conclusion was that analytics accuracy is primarily acontext and verification problem.
Source: Anthropic, “How Anthropic enables self-service data analytics with Claude,” June 3, 2026. These are Anthropic's reported internal results, not Propperly customer results.
Read what Anthropic's ~95% result actually required →Find out what your agent would actually need — before you integrate anything.
You shouldn't need a warehouse migration, production connectors or weeks with an engineer just to learn whether your analytics context can be reconstructed.
Propperly reconstructs what appears current, what has been superseded, where sources conflict, what remains unresolved and what evidence supports each conclusion.
Your team decides what to accept, correct, or leave unresolved. Nothing becomes organizational truth because an AI produced it.
Start with one domain where wrong answers matter
Choose an area where your team already has real questions, real definitions and real historical work.
What you're actually testing
- Can the context be reconstructed?
- Can the team verify it?
- Does it improve representative agent answers?
- Can it stay current when the organization changes?
Without an integration project first
Prove the reconstruction before you build the infrastructure around it.
What that ~95% actually required underneath.
Anthropic's analytics architecture combines strong data foundations, governed sources of truth, procedural Skills, evaluation and continuous maintenance.
We turn that requirement into a bounded implementation project.
- 01
Governed sources
Identify which datasets, definitions and references are canonical, which are fallback sources, and which should no longer guide an agent.
- 02
Business context
Capture the definitions, scope, exclusions, terminology and business meaning an experienced analyst uses without thinking about it.
- 03
Current decisions
Separate what is current from what was previously true. Preserve changes, rejected interpretations, unresolved conflicts and the evidence behind them instead of flattening everything into one document.
- 04
Agent knowledge
Turn the approved domain knowledge into structured, agent-readable references that help route questions toward the right source.
- 05
Analyst workflow
Encode how a strong analyst approaches the domain: what to clarify, where to look first, common failure modes, validation steps and when not to answer.
- 06
Evaluation & maintenance
Define representative questions, expected behavior and provenance requirements — then establish how the context changes when the business changes.
You shouldn't need to rediscover this architecture from scratch.
Turn what your analytics team already knows into an agent-ready foundation.
Agent Readiness Sprint
We start with one bounded analytics domain and the work your team already has.
Propperly reconstructs the domain and surfaces what needs human judgment.
Your team decides what stands.
Nothing silently becomes organizational truth because an AI inferred it.
What you leave with
The sources an agent should prefer, use cautiously or avoid.
Metrics, entities, grain, scope, exclusions and terminology.
Current decisions, superseded decisions, disagreements and unresolved questions with supporting evidence.
Structured domain references designed for agent consumption and retrieval.
The workflow, routing logic, validation rules and known traps an experienced analyst would teach a new team member.
A bounded evaluation set, expected behaviors, provenance requirements and a clear handoff into the agent environment you're building.
From scattered work to reviewed agent context.
Data, definitions, dashboards, analyses, documentation and decisions.
- What appears current?
- Which sources agree?
- Which sources conflict?
- What was superseded?
- What lacks enough evidence?
- Governed sources.
- Business definitions.
- Current decisions.
- Procedural knowledge.
- Explicit uncertainty.
- Evidence.
Not another summary of everything your company has ever said.
A reviewed representation of what the agent can rely on now.
Building it is only half the problem.
A reliable agent can become unreliable without changing the model at all.
Anthropic saw this happen internally.
Its Skill documentation described a data model that changed every day. Without active maintenance, Anthropic reports that offline accuracy drifted from approximately95% to 65% in one month.
They eventually treated Skill maintenance as an engineering problem. Anthropic reports that roughly 90% of its data-model pull requests now include a Skill update in the same change.
Your data changes.
Your definitions change.
Your organization makes new decisions.
Your agent context has to change with them.
Source: Anthropic, “How Anthropic enables self-service data analytics with Claude,” June 3, 2026. These are Anthropic's reported internal results, not Propperly customer results.
The company changes. The agent's understanding has to change with it.
Managed Context Maintenance
The initial build creates an approved starting point.
Maintenance keeps it from becoming another stale knowledge asset.
When underlying sources change, the domain can be manually recommissioned. Propperly surfaces relevant differences against the reviewed current state and routes what requires renewed judgment for human governance.
- A new definition.
- A changed model.
- A decision that supersedes an old one.
- A conflict between sources.
- A correction from an analyst.
- A previously unresolved question that now has an answer.
Governance is targeted: surface what requires attention, not a full rewrite. Governed history persists — the state updates without erasing what came before.
The loop
The principle
Don't ask humans to maintain another giant context document.
Route only what requires their judgment.
Measured on a sealed longitudinal corpus.
We evaluated whether a maintained governed state improves temporal answer reliability — the ability of a downstream agent to produce answers that reflect what the organization currently treats as valid, rather than what any source said at some prior point.
Corpus and metric
893-object sealed longitudinal corpus from the Mozilla codebase, covering a period of documented organizational change. Evaluation set: 49 scored instances across 29 families. Metric: family-macro Temporal Answer Reliability (TAR). Evaluation type: retrospective Exploratory v2.
Result
Propperly's maintained governed state scored 26–27 points above two summary baselines — a Fresh Summary baseline and an Automated Maintained Summary baseline. Those differences were statistically robust in the retrospective evaluation, though both baselines had representation limitations.
Limits of this evaluation
Its advantage over a historical hybrid-RAG baseline was not statistically conclusive. A Human-Reviewed Maintained Summary arm (C2) was pre-registered and deferred — representing a stronger human-maintained alternative not yet measured in this evaluation. This is a retrospective technical evaluation, not commercial or customer validation.
Bring your agent. Keep your stack.
We're not replacing the systems that contain your data. We're making the context around them usable by agents.
Propperly isn't another text-to-SQL chatbot, a replacement warehouse, or an autonomous system that silently declares company truth.
We work on a different problem:
What should that agent be allowed to believe?
We reconstruct the context behind analytical work, preserve where it came from, represent changes instead of overwriting them, and keep human authority at the point where evidence becomes organizational truth.
The result isn't more access to information.
It's a better foundation for using that information reliably.
What analytics buyers ask first.
Straight answers on how this works commercially — and where it fits with what you already run.
Is Propperly software, a service, or both?
Propperly is a productized system, delivered today through a hands-on engagement. The system performs the reconstruction and maintenance; today the initial deployment and evaluation are hands-on — the Agent Readiness Sprint — rather than a self-serve platform you provision yourself. It isn't generic analytics consulting: you leave with a reviewed, maintainable current-state, not a report. The Local Evaluation Kit is not publicly self-serve yet.
Do we need production warehouse access, or a big integration, to start?
No. A bounded reconstruction works from the analyses, SQL, dashboards, definitions, tickets and decisions your team already has. No warehouse migration and no production connector are required to see whether the context can be reconstructed and verified.
What if we already have a semantic layer, catalog, or RAG/search?
Keep them. Propperly governs a different object — the analytical conclusions and decisions — and defers metric definitions to your semantic layer and metadata and lineage to your catalog. Search and RAG make more findable; Propperly maintains what currently governs. See the full boundary on Why Propperly →
Who decides what becomes "current"?
People do. Propperly reconstructs and proposes; a reviewer approves, edits, rejects, or leaves it unresolved. A model's inference never silently becomes organizational truth.
What happens when sources disagree?
Conflict is kept as a first-class state the agent can see, rather than quietly collapsed into a single answer. It stays visible until a person resolves it — and the same is true of questions the organization hasn't decided yet.
What is maintained after the initial Sprint?
When underlying sources change, the domain can be manually recommissioned. Propperly surfaces relevant differences and routes what requires renewed judgment — a changed definition, a new or superseded decision, a conflict, a correction, a now-answered question — for human governance. The governed state updates without erasing history. The agent consumes the updated governed state via read-only MCP access.
Do you build the agent, or can we use our own model and agent platform?
Use your own. Propperly prepares and maintains what an agent relies on; it doesn't require a particular model or agent platform, and it isn't a text-to-SQL chatbot or a replacement warehouse. Bring your agent — Claude or otherwise. The governed state is served to agents via read-only MCP access; the agent queries the governed analytical understanding rather than reconstructing organizational truth from raw artifacts on every call.
What does "Run it yourself" mean today?
The Local Evaluation Kit lets a team try a reconstruction on its own material. It is not publicly self-serve yet — today it's arranged with us — so treat "Run it yourself" as a hands-on evaluation rather than an instant download.
Building an analytics agent?
Start with the part it needs to trust.
Bring one bounded domain and the work your team already has.
We'll explore what it would take to turn it into a reviewed, maintainable foundation for reliable AI analytics.
Read the research →Request a focused pilot
This is a conversation about a bounded pilot on a slice of your own analytics history — not immediate access to a production enterprise platform. Tell us the problem worth testing and we'll follow up to scope it.
Please don't paste dashboards, data, credentials, or sensitive definitions — a short description of the problem is all we need to start.
We use these details to review and respond to your request. Do not submit project files, credentials or sensitive information. Forms are processed through Netlify.
Privacy details
We collect only the contact and qualification details you enter here, and use them solely to review and respond to your request. Please don't submit project files, source code, API keys, credentials, secrets, or other sensitive information. Submitting this form doesn't create an account. You can request deletion atassaf.m@propperly.ai. Submissions are processed through Netlify. We measure aggregated website traffic and interaction events through Plausible; the contents of this form are not sent to analytics.
Request received.
I'll review it and email you if there is a clear fit for a focused pilot.