Ariesnet Start an engagement

The method

The context supply chain

A context supply chain is everything that has to be true before an agent gives you a good answer.

The industry has a name for this work now — context engineering. The context supply chain is how we do it.

Where the content came from. What shape it is in. Whether two teams mean the same thing by the same word. Whether anyone checked it before the agent read it. What got selected, what the agent did with it, and what happened the time it was wrong. Seven links, and a chain is worth what its weakest one is worth.

The encouraging part is that all seven fail in ways you will recognize the moment you read them, and all seven have fixes you can point at afterwards. Not a workshop and a slide deck. A source register, a schema set, a term map, a rejection log, a provenance record — things your team owns and keeps.

Agents fail on context far more often than they fail on intelligence. Not never on intelligence: models do have limits, and we will tell you plainly when you have hit one. But if you are deciding where the next quarter of budget goes, context is where the returns are, and it is the part you control.

Worth being sceptical about

The stages are the map. The artifacts are the work.

Be sceptical of the seven-stage picture on its own. Every stage in it lines up with something large organizations already buy, and 'our framework is proprietary' is not an argument, it is a logo on a slide. We use the chain because it is a good map: it tells you which conversation you are in and which stage owns a failure.

What is worth paying for is what you are holding when we leave. A register with names on it. Schemas your own systems already read. A gate that refuses content failing its contract and tells you which part it failed. A record of exactly what an agent read before it answered. Those can be inspected, argued with, and handed to the next team. A framework cannot.

So every stage below names the artifact it produces and the gate it enforces. A stage that only described an activity would not have earned a section.

The seven stages

What breaks, and what you get back

For each stage: the failure you will recognize, what the stage does about it, and the artifact and the gate you are left with.

Stage 1

Sources

What content exists, who owns it, and whether it is current.

You will recognize this

Legal signed the master agreement in March. In June an agent answered a customer from a draft someone had emailed the previous autumn — confidently, with the wrong indemnity clause, and with nothing in the answer that would have let you notice.

What the stage does

Every kind of content your agents touch gets one designated system of record and one named human owner. Copies stay copies. Where no authoritative source exists yet, that absence gets written down instead of quietly tolerated.

What you receive

  • A source register: each content domain, its system of record, its owner by name, how often it is expected to change, and where the unofficial copies are still circulating.
  • An agreed exceptions list — the domains that have no authoritative source yet, each with the owner who accepted that.

The gate

Content from a source nobody designated does not enter the chain. It lands on the exceptions list with a name against it, which is how 'we didn't know that was in there' stops being available to anyone.

Stage 2

Structure

Whether that content has machine-legible shape.

You will recognize this

The travel policy is a forty-page PDF. The agent could not tell a rule from an exception, and could not work out which of the three effective dates in it applied to the person asking.

What the stage does

Content is given a shape a machine can act on — typed records with fields, versions and effective dates — instead of prose that every agent has to reinterpret from scratch, slightly differently, every time.

What you receive

  • A published schema set: the shapes your content must take, exported in formats your systems already read — JSON Schema, JSON-LD, ShEx, XSD.
  • A converted corpus — your priority content as typed records, not a folder of documents with good intentions attached.

The gate

Every record is checked against its shape before it counts as converted. One that misses a required field comes back with the field named and the record identified, rather than passing with a footnote.

Stage 3

Semantics

Whether a term means one thing across systems.

You will recognize this

Sales says 'enterprise'. Finance says 'strategic account'. Support says 'tier one'. The agent treated them as three unrelated things, and nobody noticed until two decks disagreed in the same meeting.

What the stage does

Contested terms get one definition that wins — without asking any team to give up the word it actually uses. Your teams keep their own vocabulary; the systems still agree on what a customer is.

What you receive

  • A term map: one entry per contested concept, the definition that wins, each team's local name for it, and the identifiers that resolve to it.
  • Alignment to public vocabulary wherever public vocabulary already exists, so your private extensions cover only the part that is genuinely yours.

The gate

A new term cannot arrive quietly. It either matches an existing entry or is registered as a genuinely new concept with an owner — no drift of unrecorded synonyms accumulating in the corner.

Stage 4

Validation

Whether anything checks the content before an agent reads it.

You will recognize this

The agent cited last year's pricing policy. The document was real, it was in the right system, and it had been superseded in April. Nothing in the path had asked whether it was still the current one.

What the stage does

Content is measured against a contract — currency, ownership, conflict with other approved content, fit with policy — at the boundary, before it is available to an agent at all. Publicly this is conformance and efficiency protocol work; the contract itself is written down and agreed with you.

What you receive

  • A conformance gate standing between your content and your agents, with the contract each item is measured against stated in terms your owners can read.
  • A rejection log: every item that failed, what it failed on, and who owns the fix. Clients tend to use this one hardest, because it is the first honest inventory of what is actually wrong.

The gate

This stage is the gate. Non-conforming content does not reach retrieval — not ranked lower, not there. And you can read exactly why it was refused.

Stage 5

Retrieval

Whether the right material actually comes back.

You will recognize this

The correct answer was in the wiki. The agent returned a near-identical page belonging to a different business unit, and sounded exactly as certain as it would have if it had been right.

What the stage does

Selection happens on meaning, shape and entitlement rather than keyword proximity. The agent is handed the slice its task justifies and the person asking is allowed to see.

What you receive

  • A context package definition: for each task your agents perform, the material they receive, with every item's source, version and scope stamped on it.
  • A retrieval acceptance suite — real questions from your own operators, each with the material that must come back and the material that must not, runnable as tests on every change.

The gate

Material outside the requester's entitlement is never assembled in the first place. Scope is applied when the context is built, not filtered out of an answer afterwards.

Stage 6

Agent Operations

Whether production behavior is governed and observable — including writes.

You will recognize this

The support assistant issued a refund outside policy. Nothing checked the limit at the moment of action, and afterwards nobody could reconstruct what the agent had read before it decided.

What the stage does

Agents run against governed context and governed tools — and, the part most programs postpone, they are allowed to write back into your systems under an explicit policy instead of through a person copying and pasting.

What you receive

  • A write-back policy and its approval record: which agents may write what, to which system, under whose authority, and what became of every proposed write — approved, amended, or refused.
  • A provenance ledger, so each answer carries its receipts: which sources, which versions, which moment.

The gate

A write outside the policy is held for a named human. It is not executed and reported afterwards — that is the version most teams already have, and it is the version that ends in an incident review.

Stage 7

Learning Loop

Whether observed failures actually change the system.

You will recognize this

The same edge case failed for four months. Someone corrected it in a chat thread every time, everybody was helpful about it, and nothing upstream ever changed.

What the stage does

Failures are routed back to the stage that caused them, with an owner and a date. A correction that lives only in a conversation is treated as an open failure, not as a fix.

What you receive

  • A failure register: each observed failure classified to the stage that produced it, with the owner and the change that closed it.
  • A telemetry baseline on the measures your own team chose, so next quarter's argument is about evidence instead of impressions.

The gate

A failure is not closed by a workaround. Closure needs the change to land in the source, the schema, the term map or the policy — and the acceptance question that caught it to pass on a re-run.

On numbers, plainly: measured efficiency results we publish are ours, from our own operations. Anything we model for you is built from your telemetry, and a model is not a promise.

Maturity

Where you are on the path

Six levels, applied to each stage separately. Mixed placements are normal and diagnostic in themselves — a sophisticated retrieval layer sitting on ad hoc sources is the most common shape we assess.

  1. L0

    Ad hoc

    You are here if

    The answer depends on who did the work, and nobody can point to the one place a given fact lives.

  2. L1

    Structured

    You are here if

    Your content has consistent shape and you can query it, but two teams still mean different things by the same word.

  3. L2

    Semantic

    You are here if

    Terms resolve to one definition across systems, and nothing yet checks a document before an agent reads it.

  4. L3

    Validated

    You are here if

    Non-conforming content is caught before it is published, but you could not tell me what your agents actually read yesterday.

  5. L4

    Operational

    You are here if

    Production behavior is measured and attributable, and fixes still depend on someone remembering to make them.

  6. L5

    Self-improving

    You are here if

    A failure seen on Monday reliably changes a source, a schema or a policy — and you can show me the change.

The binding constraint is the lowest-placed stage that the failing work depends on. Spending above it changes nothing you can measure, which is why assessment comes before implementation rather than running alongside it.

Named patterns

The rules of thumb we work by

These have names because we reach for them often enough to need names. They are the practice, stated plainly.

  • Authoritative Source Designation

    Stage 1

    One system of record per kind of content, with a named owner. Everything else is a copy and gets treated as one.

  • Schema Before Schema-less

    Stage 2

    Decide the shape first. Unstructured content joins the chain by conversion, not by hope.

  • Look Up Before You Make Up

    Stages 3 and 5

    Check what already exists before inventing a term, a value or a fact. A great deal of invention is really a missed lookup.

  • Public Commons Private Extension

    Stages 2-3

    Reuse public vocabulary where it exists. Keep private extensions for the part that is genuinely your advantage.

  • Contract at the Boundary

    Stages 4 and 6

    Content and actions are checked where they cross a boundary, not somewhere hopeful downstream.

  • Scoped Context Loading

    Stages 4-5

    An agent receives the slice its task and its entitlement justify. Nothing else travels with it.

  • Governed Write-Back

    Stage 6

    Agents may change your systems, under a written policy, with approval and a record. Read-only is a stage, not a destination.

  • Provenance By Default

    Stages 6-7

    Every answer carries its receipts: which source, which version, which moment. By default, not on request.

  • Temporal Anchoring

    Stage 7

    'Current' is a question with a date attached. Content is anchored in time so it can be superseded instead of silently contradicted.

  • Signal Before Structure

    Stage 7

    Collect the signal that tells you what to fix next before building more structure on top of what you already have.

Where to start

You do not have to fix seven things

In nearly every estate we assess, two stages are carrying the failures and the other five are fine for now. Finding out which two takes a few weeks, not a program, and it is the smallest commitment on the ladder that produces something a CFO can act on.

The engagement ladder runs from a half-day briefing through to running the chain for you. Every rung is fixed scope and fixed fee, priced before we start, and every rung ends in something you keep whether or not you climb to the next one.

See the engagement ladder The Context Supply Chain Assessment

Next step

Find out what your context is actually costing you.

We work with enterprise AI and data leadership teams.

Start a conversation Review pricing

Engineering since 1997