−40%
token cost
median across 8 tasks, one production org
Anthropic PartnerAWS × Anthropic Accelerator
CLIENTELL CONTEXT GRAPH
The Context Graph: a deterministic map for enterprise AI agents.
Most AI agents guess at how your business works, then guess again on the next call. The Context Graph reads your CRM, data warehouse, files, and tickets once, then hands every agent the exact, verified path to what a task needs. A lookup, not a guess.
ExtractGraphTraverseServe
+28%
task accuracy
median across 8 tasks, one production org
2×
faster tasks
median across 8 tasks, one production org
All three figures: Median across 8 tasks, one production org. Your environment will differ, which is why we benchmark one of your own workloads before you expand.
Pilots look like magic. Production looks like a bill and a support ticket.
Missing context turns every agent call into three production costs: re-reading, guessing, and decisions nobody can replay.
The environment gets re-read on every call.
Schemas, files, and tool definitions are loaded again and again.
The agent guesses at dependencies it cannot see.
Invisible dependencies create wrong calls, retries, and rollbacks.
The same prompt can take a different path twice.
Without a stable path and replay, repeatable decisions are impossible.
Every agentic call starts from zero.
01Re-read02Guess03Act04ForgetNo memory between runsand it starts over
- 01
Re-read
Load schemas, files, and tools.
- 02
Guess
Infer hidden dependencies.
- 03
Act
Call a tool with partial context.
- 04
Forget
Discard the structure after the run.
A larger model does not fix this. The missing asset is a persistent map of the environment.
What is an AI context layer?
An AI context layer is a persistent, governed map of the systems an agent needs before it acts. Clientell maps schemas, tools, automations, and lineage once, then serves only the dependency path required for each task. Every node and relationship keeps its source, version, confidence, and timestamp. Your models and source systems stay in place.
Drop-in
Works beneath your current stack.
Model-agnostic
One graph serves any model or agent.
Persistent
Map once; update as systems change.
Inspectable
Source, confidence, version, and replay stay attached.
Your environment
Data lakes · CRM · ERP · code · docs · tickets
Clientell context layerSource · version · confidence
Your agents
Only the path required for this task
The problem
Retrieval finds passages. It does not find paths.
An AI context layer stores the relationships an agent needs before it acts. Retrieval finds similar passages. A context graph follows the actual path between objects, automations, jobs, and outputs.
A question such as “Which reports break if this field changes?” has no single passage to retrieve. The answer is a dependency path: field → object → automation → job → dashboard.
Why the usual context fixes stall
Bigger context windows
Fit more of the environment into a single prompt.
You pay to re-read a larger pile, while the relationships inside it remain implicit.
Vector databases and RAG
Return the chunks whose wording is closest to the request.
Similarity does not show which system calls another, which definition is live, or what breaks next.
Memory layers
Recall text and events from previous sessions.
Past text is not a current, typed map of the environment the agent is about to change.
Retrieval ranks what sounds related.
Three strong matches. Three separate passages. The dependency between them is still missing.Which downstream reports break if we change this field?
FIELD
FIELD OF
OBJECT
FIRES ON
AUTOMATION
TRIGGERS
JOB
FEEDS
DASHBOARD
Dashboard reached
A context graph walks what is actually connected.
The same evidence becomes a typed route from the changed field to the dashboard it affects.How it connects
Structure travels. Your data does not.
Read-only extractors map each system inside your estate. Only metadata and relationships move to the graph plane; row-level data stays where it is.
Extractors available for
One graph, not one per system
Each new system extends the same graph instead of creating another context silo.
Each connector holds read-only scope. It can see how a system is put together. It cannot change it, and it does not copy records out.
Email, calendars, files, chat, and your own internal APIs connect through the same extractor pattern. Anything with a readable API can be mapped, and a Clientell engineer builds the extractor as part of the engagement.
SnowflakeDatabricksSalesforceNetSuiteGitHubJiraConfluence
GOVERNED
CONTEXT GRAPH
Data stays in your estate
Extractors map structure without copying records out.
Deployed in your VPC
The graph runs under your network and residency controls.
Read-only scopes
Each connector can observe structure, not change it.
Metadata, not row data
Schemas, dependencies, and automations travel. Customer records do not.
The lifecycle
Six stages, every system, every run
Every connected system runs the same six stages, keeping the graph current and inspectable.
Your systemsOne direction, nothing rediscoveredThe agent
010203040506
Extract
Read objects, schemas, dependencies, and automations at the source.
Graph
Turn that structure into typed nodes and relationships.
Score
Label each relationship by evidence and confidence.
Traverse
Follow the dependency path required by the question.
Serve
Send only the nodes on that path to the model.
Learn
Use outcomes to update confidence without hiding history.
The agent
SOURCE
EXTRACTED
AGENT
A stated fact, read directly from the system.
A declared foreign key or a trigger that names its object.Edge honesty
The agent can tell the difference between what it read and what it inferred.
Every relationship carries its evidence level. Uncertainty gets escalated, not averaged away. Stated facts, supported deductions, and unresolved conflicts never look identical to the agent.
Safe to act on without asking.
The label travels with the answer, so whoever reads it downstream sees the same evidence level the agent saw.
Attachment points
Four ways in, one graph behind all of them
Attach through an SDK, MCP server, plugin, or proxy. Pick the surface that fits your stack; every option reaches the same graph, labels, and replay history.
Four ways in
MCP serverSDKPluginProxy / gateway
One Context GraphEvery door returns the same thing
Same graphSame labelsSame replay
EXTRACTEDINFERREDAMBIGUOUS
The attachment changes. The graph, labels, and replay do not.
Flight recorder
Full replay of what any agent saw before it acted.
Every traversal records the exact context served to the agent at decision time.
Replay any past run against the graph as it stood then. Review the decision from evidence instead of reconstructing it from memory.
See the production benchmarkTRAVERSAL / 7F32A9REPLAY READY
01
Nodes served
The nodes handed to the model on that run.
02
Versions
The version of each node at decision time.
03
Weights
The weight each relationship carried.
04
Provenance
The source and confidence label for every node.
Production benchmark
Same task. Same model. One variable: the graph.
Eight tasks ran both ways on one production environment. Every task used fewer tokens with the graph.
Without graphWith graph
−67%
token cost20.2 min → 2.1 min−40%median token cost
+28%median task accuracy
2×median task speed
Same workload, same model, same prompts. Measured end-to-end, including failed calls and retries.
The 32.6-minute tech-debt run was capped and still returned estimates, so its true baseline cost is higher than shown.
One production org, US manufacturing, 2026.
Last updated September 1, 2026
Accuracy benchmark
The same model, asked the same questions, with and without the graph.
110 structural questions about one production Salesforce org, asked 3 times each, graded by Claude Opus. Same model, same questions, same grader. The only variable is whether the graph was there to query.
44.2%working without it
86.5%reading the graph
A controlled result on one org, not a live customer average. It measures answer accuracy on structural questions, which is a different thing from the end-to-end task benchmark above.
Use it under internal work or inside the product you ship.
One governed graph supports two enterprise paths: employees using AI internally, or customer-facing agents inside your product.
Give everyone the same map in the tools they already open.
Employees query one governed map from the browser, IDE, admin console, or internal scripts they already use.
BrowserIDEAdmin
One company graph
Same context
Fewer tokens, visible dependencies, and repeatable answers from one shared layer.
Give every customer agent a complete, isolated environment map.
Serve every customer agent from an isolated graph while keeping your product orchestration unchanged.
Customer ACustomer BCustomer C
Isolated graphs
One product API
Lower context cost per task, consistent behavior, and strict tenant boundaries.
Have a production workload in mind?
Book a workload benchmarkStart with one workload, not a transformation project.
A Clientell engineer stays inside all six stages. Your team signs off on the estate, the security boundary, the graph, and the benchmark before the next stage opens.
010203040506Your teamClientell engineerYour sign-offthe estatethe security boundarythe graphthe benchmark
Estate audit
List every system, its owner, and every place AI already runs.
Access
Complete the infosec review, residency decision, and deployment boundary.
First graph
Crawl, resolve, and commit one bounded slice of the environment.
Teaching
Let the people who know the estate rule on what extraction could not settle.
Proof run
Run the workload both ways against the rubric your team approved.
Deploy
Expand only after the result earns the next workload and the next system.
Your teamClientell engineerA filled marker is who drives that stage.
Access and the infosec review usually take the longest, because that runs on your team's calendar rather than ours.
Where this came from
Built for a hard problem, then it turned out not to be specific to it.
Clientell first built this layer for its own Salesforce agent across 2,000+ customer orgs. Every org had different schemas, automations, and undocumented dependencies.
The agent needed dependency paths, confidence labels, and a replayable record, not more text chunks. We hardened those primitives across the live environments it served.
The same pattern exists in warehouses, ERPs, and code repositories: declared structure, implied structure, and unresolved structure. The context layer now stands on its own across those systems.
2,000+customer orgs the graph has been hardened across
7systems with extractors available today
3confidence labels on every relationship learned
Common questions
What is the Context Graph?
The Context Graph is a governed map of your CRM and the rest of your enterprise stack: data warehouses, email, files, tickets, and custom APIs. It resolves a specific path to what a task needs instead of flooding a model with your whole business. Every relationship is typed, versioned, and labeled EXTRACTED, INFERRED, or AMBIGUOUS. Every agent and AI client reads from the same graph, so context is built once and reused rather than rebuilt for every task.
Does this store our data?
No row-level customer data is copied into the graph. Read-only extractors send schemas, relationships, dependencies, and automations from inside your estate. The deployment runs in your VPC.
Is this the same as RAG?
No. RAG retrieves passages; a context graph follows typed relationships across systems. They can work together: retrieval supplies relevant text, while the graph supplies the dependency path and confidence behind it.
How is this different from a vector database?
A vector database answers “what looks similar?” This graph answers “what is connected, through which path, and with what confidence?” Every relationship is typed, versioned, and labeled EXTRACTED, INFERRED, or AMBIGUOUS.
What happens when a system changes?
The extractor updates the affected nodes and relationships. Previous versions remain available for replay. Contradictions become AMBIGUOUS edges and are sent for human review.
What happens when the graph is wrong?
Contradictions become AMBIGUOUS edges and go to a named owner on your team rather than into an answer. Every ruling a human makes is recorded, so the correction is visible in the replay of any later run. If a source system itself is wrong, the extractor reports what that system says. The flight recorder shows exactly what the agent was handed at decision time.
What can we actually use today?
Clientell can attach through an SDK, MCP server, plugin, or proxy/gateway. Seven system extractors are prebuilt today, and the same extractor pattern reaches email, calendars, files, chat, and your own internal APIs. All of them map into the same graph, so your attachment choice does not create a separate context silo.
Can we audit what the agent saw before it acted?
Yes. Each traversal records the nodes, versions, weights, and provenance served at decision time. A past run can be replayed against the graph state the agent actually saw.
Where do the performance numbers come from?
They are medians across 8 tasks in one production environment, measured end-to-end including retries. The result was 40% lower token cost, 28% higher task accuracy, and 2× faster completion. Your result will depend on your environment, so Clientell benchmarks one of your real workloads before expansion.
How long does the first graph take?
Weeks, not quarters. Access and the infosec review usually take the longest, because that runs on your team's calendar rather than ours. Once access lands, one bounded slice of the environment is crawled, resolved, and committed, then the workload runs both ways. We scope the timeline against your security review before you commit to anything.
What does this cost?
The proof engagement is scoped to one workload and priced per engagement, not per seat. There is no integration commitment. You get the benchmark and the graph findings whether you expand or not. Pricing beyond that depends on how many systems you map and how the graph is deployed. We quote it after the estate audit rather than before.
One workload · your environment
See what the Context Graph changes before you commit to it.
Pick a real production task. We run it both ways against a rubric your team approves, then give you the benchmark and graph findings whatever you decide.
One workload · same model · same prompts · your numbers
Read-only scopes throughoutRuns in your VPCMetadata only, never row dataNo integration commitment