Seven frameworks for understanding the economics, sovereignty, and governance of enterprise AI agent deployments. Developed by Tony Wenzel, Co-Founder & CEO of Excipio, and published in The Enterprise Token Economy on Substack.
Excipio is a private memory layer that sits between every AI surface, apps, agents, tools, MCP servers, and the models they call. It intercepts each agent query at a drop-in gateway, serves semantically similar answers from a local cache, and passes genuine cache misses through to your configured model. The result is sub-10ms responses on cache hits and a customer-owned knowledge graph that compounds with every interaction, with no changes to agent code. Savings are measured per mechanism and workload class against an instrumented baseline.
Every LLM-bound agent query is captured at the gateway before it reaches a frontier model. The query is converted to a vector embedding locally, inside your perimeter, and matched against a per-agent index. A semantic match is served in under 10ms; a miss passes through to your configured model. Every resolved query is added to a private, enterprise-owned knowledge graph. It is the same carrier-class intercept-and-cache architecture our engineering team has shipped for more than 20 years, built in C++ for the AI agent token economy.
Prompt caching reduces the cost of re-sending context, but your intelligence still flows out to the vendor's model. A routing gateway switches traffic between models. Excipio caches semantically rather than by exact string, and passes misses through unchanged, but its defining difference is sovereignty: queries are resolved against a knowledge graph the enterprise owns, so validated answers compound inside your perimeter instead of inside someone else's model.
Excipio is a drop-in proxy. You change one BASE_URL environment variable so agent traffic points at Excipio instead of directly at the model API. There is no rearchitecting and no change to agent code, so most deployments are live in hours rather than weeks.
Excipio is model-agnostic and cloud-agnostic by design. Cache misses can route to AWS Bedrock, OpenAI, Anthropic, Azure or Vertex, or a local open-weight model, and you can run on AWS, Azure, GCP, or on-premise, at the same time. You can switch or add models without touching agent code.
Excipio is built for enterprises running AI at scale, especially regulated industries such as banking, insurance, and healthcare where cost, latency, model portability, and data governance all matter at once. It speaks to four buyers at once: the FinOps and finance owners focused on token spend, the CTO on performance and portability, the CISO and General Counsel on data protection, and the Chief Data Officer and board on owning institutional intelligence.
A cache hit is when an incoming agent query is semantically similar to one Excipio has already answered, so the stored answer is served instead of calling a frontier model. Hits return in under 10ms at a fraction of the cost of a fresh call. The share of queries that resolve from cache depends on how repetitive the workload is, and it compounds as the knowledge graph grows. Excipio does not quote a single blanket savings percentage, because none exists until it is measured at your deployment. Savings are measured per mechanism against an instrumented baseline: system-prompt caching against repeated context, exact-match resolution, retrieval deduplication, and validated semantic caching where the workload supports it. Industry telemetry shows the scale of the opportunity: 69% of input tokens are repeated context, yet only 28% of calls use any caching (Datadog State of AI Engineering 2026). Those are market priors, not product claims. Your own telemetry produces the number that matters.
Honestly: response caching earns almost nothing on agentic traffic, because agent loops repeat in form but almost never in content. We measured it and we say so. What agentic workloads do gain is substantial: every loop resends the same system prompt and repeated context at full price, which prompt caching captures; and every call passes through the governance and telemetry layer, which earns with zero cache hits. The workload determines the mechanism, and your telemetry shows which mechanisms pay before you commit to anything.
Excipio combines an annual platform commitment with a consumption component tied to the savings we measure, so cost scales with the value delivered. Contact us for specifics.
On a cache hit the query is resolved locally and zero bytes leave your perimeter, so agent query intent and proprietary logic are not transmitted to an external model on those calls. Per-agent sensitivity tiering lets each agent carry its own data-classification and routing policy. Excipio is designed to support GDPR, HIPAA, and SOC 2 at the architecture level, helping preserve confidentiality as a property of the design rather than a contractual promise.
Every validated cache hit adds a query-answer pair to a private knowledge graph that belongs to the enterprise, not to Excipio and not to any model provider. It grows more accurate and more valuable with every agent interaction, turning traffic that would otherwise leak intelligence to a frontier model into a compounding, owned asset. This is what firm sovereignty means: controlling where your institutional intelligence compounds.
The cost enterprises pay when AI agents re-process answers they already have, billing at full frontier rates for queries already resolved.
Read the full definition →The institutional intelligence enterprises transmit to external LLM providers on every agent query that exits the perimeter.
Read the full definition →The ratio of necessary spend to total spend in an enterprise AI agent deployment. Necessary spend divided by total spend.
Read the full definition →The strategic property of an enterprise controlling where its institutional intelligence compounds. Named by Satya Nadella at Davos, January 2026.
Read the full definition →The principle that AI compliance constraints should be enforced at the infrastructure layer, not through policy documents or contractual promises.
Read the full definition →AI agent memory lets agents reuse what they have learned. Sovereign memory architecture is the enterprise form that keeps governed memory inside the perimeter, invariant by design.
Read the full definition →Any AI identity in the request path, an agent, a service account, an application, or a user-facing assistant, whose model-bound traffic is audited, governed, and controlled.
Read the full definition →The same seven ideas, in the words people actually use.
Frontier models are stateless. Every session starts from nothing, so a question your system resolved last week gets computed and billed again this week, and again next week. Nothing is wrong with your setup. The architecture has no memory. The name for that recurring cost is the Rediscovery Tax. Read the full definition →
Point agent traffic at a memory layer that resolves repeat questions locally before they reach a model. One BASE_URL change, no edits to agent code. The metric worth tracking is not how many tokens you bought but how many the work actually required. That ratio is Token Yield: necessary spend over total spend. Read the full definition →
Not on every call. Genuinely new questions have to go somewhere. But repeat questions can resolve inside your own network, and on those calls nothing is transmitted at all. That distinction is Governance by Architecture: the control holds because of how the system is built, not because of what a contract promises. Read the full definition →
Every query carries context: how your business works, what your people ask, what your policies constrain. Individually each call looks harmless. In aggregate it is a description of how your firm operates, transmitted to a system you do not own. That accumulating transfer is the Disclosure Tax. Read the full definition →
It is a governance problem that caching makes visible. Any stored answer needs three things: lineage showing where it came from, invalidation rules saying when it stops being true, and source precedence deciding which input wins when two disagree. Without those, stored answers rot silently. That failure has a name: the Stale Answer Tax.
Whoever holds the memory. If validated answers accumulate inside a model provider's system, the compounding value accrues there. If they accumulate inside your perimeter, it accrues to you. That question, where your institutional intelligence compounds, is Firm Sovereignty. Satya Nadella named it from the Davos stage in January 2026. Read the full definition →
Prompt caching matches identical text and lowers the cost of resending context. It does not help when the same question is asked in different words, and the intelligence still leaves your network. Semantic memory matches meaning rather than characters, resolves the repeat inside your perimeter, and keeps the resulting graph as yours. The enterprise form of that is Sovereign Memory Architecture. Read the full definition →
One BASE_URL change against a single agent. No rearchitecting, no data leaving your perimeter, and nothing to unwind if you walk away.