Context engineering for enterprise AI

Agents fail on context, not intelligence

Ariesnet designs and deploys the Context Supply Chain so enterprise AI agents run on governed, validated, operational context, not ad hoc prompts and scattered data.

Request a briefing See the engagement ladder

Engineering software for enterprises since 1997. Systems that had to keep running under real load, real constraints, and real consequences.

The Context Supply Chain

Sources
Structure
Semantics
Validation
Retrieval
Agent Operations
Learning Loop

Engagements

Start where the evidence is thinnest

Most buyers start at the audit because it is the smallest commitment that produces something a CFO can act on.

Advisory

Executive Briefing

A facilitated half-day for the group that controls the budget. We walk the seven-stage context supply chain against your estate, place you on the L0–L5 maturity scale with reasoning shown, and cost the options — including doing nothing. Bring the skeptic.

One causal story of where context breaks, examined by the whole room
Directional L0–L5 placement, reasoning written down
The next engagement costed honestly — with what it would not fix
One-page summary for stakeholders who were not in the room

engagement

Diagnostics

Token Economics Audit

The entry diagnostic — approvable without a steering committee. Two weeks of read-only telemetry work producing cost per unit of work, a quantified waste analysis, and a remediation list with modeled recovery. Built so a finance reviewer can reproduce the arithmetic without calling us.

Cost per unit of work, not cost per million tokens
Five waste patterns hunted and quantified in your telemetry
Recovery modeled as ranges with assumptions stated — never guaranteed
Ends with a clear go or no-go for the Assessment, including “your spend is fine”

assessment

Diagnostics

Context Supply Chain Assessment

Measurement across all seven stages: sources sampled for freshness, retrieval tested against ground truth, semantics checked for drift, and the learning loop asked when a failure last changed anything. Diagnose establishes what is true; Blueprint decides what to do — and you can take the first without the second.

L0–L5 scorecard with evidence notes on every score
Targets set per stage per use case — not L5 everywhere
Implementation quarters costed before you fund them
A written go or no-go — including “do not implement yet”

assessment

All five engagements

Why this works

The graph your enterprise already wrote

Every serious AI context effort eventually needs a structural map — a graph of what your business objects are and how they relate. There are three places that graph can come from.

A model can extract it from your documents. Fast — and a statistical estimate. The graph inherits every error the extractor made, and no amount of downstream reasoning makes it authoritative afterwards.
Specialists can engineer an ontology by hand. Authoritative — and famously slow and expensive.
Or it can be recovered from structure your organization already authored: the schemas, data models and canonical definitions your teams have fought over in meetings and shipped to production. Deliberate, governed, and authoritative by construction, at a fraction of the cost of either alternative.

We build on the third source. The 10–100x indexing overhead you have heard attributed to 'graphs' belongs specifically to text-first extraction — reconstructing by inference structure that was never captured. Where structure already exists, the graph is not an expense to justify. It is an asset being recovered.

It is also why the conformance gate in Stage 4 can be authoritative at all: content is checked against a contract your own organization authored, not against a model's guess.

One honest boundary: this presupposes structure exists. For genuinely unstructured prose, vector retrieval may be entirely adequate — and we will tell you when it is. Discover by similarity; retrieve by structure.

Evidence

What an engagement looks like

Results

Establishing a retrieval baseline before scaling an agent rollout

An assessment engagement that stopped a rollout, measured the actual failure, and restarted it on a defensible baseline.

Client not named — engagement covered by a confidentiality term.

Problem

An internal support agent was performing well in demonstration and poorly in production. The team had assumed a model limitation and had begun scoping an upgrade. No evaluation harness existed, so nobody could say which answers were wrong or why.

Approach

We assessed the seven stages against the live estate. Sources and structure placed at L0 and L1: the knowledge base held superseded policy documents with no supersession marking, so retrieval was returning them as readily as current ones. We built an evaluation set from known-correct answers and measured the baseline before proposing any change.

7 stages

Assessed against live systems

Scope of the Context Supply Chain Assessment as delivered.
L0–L1

Sources and structure placement

Maturity placement recorded during assessment, evidence attached in the report.
3 weeks

Kickoff to blueprint

Actual elapsed engagement time.

The finding was not that the model was weak. It was that no stage of the supply chain distinguished a current policy document from a superseded one, so the retrieval layer had no signal to rank on.

In their words

How this gets described to us

The audit told us which half of the spend was buying nothing. We had been arguing about that number internally for two quarters.

Director of Data Platform Financial services, Client not named

Almost every engagement we open gets described to us as a model problem. Almost every one turns out to be a sources problem wearing a model problem's clothes.

Ariesnet Engineering Delivery practice

The ladder

Fixed scope, fixed fee, priced before we start

Engagement ladder: rung, engagement, price and basis
Rung	Engagement	Price	Basis
1	Executive Briefing	$7,500	Half-day engagement
2	Token Economics Audit	$15,000	2-week fixed engagement
3	Context Supply Chain Assessment Capacity-limited to 3 engagements per quarter	$32,500	3 to 4 weeks — Diagnose $15,000 plus Blueprint $17,500
4	Implementation	$60,000 – $120,000	Per quarter — scope fixed at quarter start
5	Managed Operations	$8,000 – $15,000	Per month — quarterly rolling commitment

What each engagement includes

Reading

The thinking behind the method

Ariesnet originals and syndicated excerpts, with sources marked.

Ariesnet original

Why AI agents fail on context, not intelligence

The model is rarely the weak link. What reaches it usually is — and unlike a model, that part is engineerable.

Jul 5, 2026

Syndicated · aramai.net

Look up before you make up

The cheapest reliability discipline available to an agent system: check before you assert.

Jun 20, 2026 Originally published on aramai.net

All resources