One use case, end to end — an intelligent invoice-processing agent traced through every layer of the framework
The scenario: Meridian Industrial, a fictional ~$8B-revenue industrial company, takes an intelligent invoice-processing agent from first idea to Finance-attested value. Each stage below names the framework page it exercises — this is the framework running as a system, not a set of disconnected diagrams.
This is an illustrative composite built to show how the framework pieces connect. It is not a client story, and Meridian Industrial does not exist.
Where does this play sit in the portfolio?
Classified as Horizon 2 — Operational Reinvention. Not a copilot for AP clerks; a redesign of the invoice workflow with AI as the default path.
Accounts Payable processes 1.4M invoices per year across 3 shared-service centers. Only 38% flow touchless today; the rest need manual matching, coding, or exception handling.
75% touchless processing — roughly doubling straight-through rate while holding error rates flat or better.
Size it, screen it, survive the kill gate, baseline it
T-shirt sized at ~$6–9M/yr — labor capacity release, early-payment discount capture, and duplicate-payment avoidance.
Passes 5 of 6 screens. Data readiness flagged amber: PO data quality in two legacy ERPs is inconsistent, requiring a remediation workstream before build.
2 of 5 sibling candidates in the same intake cohort were killed at this gate — one for thin value, one for an unfixable data dependency. This one advanced.
Detailed model lands at $7.2M/yr run-rate value with an 11-month payback. Finance signs the baseline — the number the project will later be attested against.
How much agent does this problem actually need?
Started as an L1 single-step extractor; exception handling (partial matches, multi-PO invoices, disputed quantities) pushed the design to L2 — planning, working memory, and explicit safety boundaries.
HITL approval required for any payment over $50k; supervised autonomy below that threshold with sampling-based review.
Complexity x autonomy scoring places it in the Medium governance tier — full eval suite and trace capture, but no board-level review required.
The test harness exists before the agent does
Bootstrapped BEFORE build: 1,200 SME-labeled invoices spanning vendors, formats, and currencies, plus synthetic edge cases (handwritten memos, multi-page line items, near-duplicate vendors).
Release gates set at >=97% field-extraction accuracy and <=0.5% wrong-vendor matches. The agent does not ship until it clears both on the golden set.
Adversarial testing targets prompt injection via invoice memo fields — free-text a vendor controls. Injection attempts must be detected and quarantined, not executed.
Canary deployment: 5% of invoice volume, then 25%, then 100% — each step gated on live eval metrics holding above threshold.
Build the use case, reuse the platform
Consumes the enterprise AI gateway — PII redaction, rate caps, and model routing come for free instead of being rebuilt.
Vendor matching runs as retrieval over governed vendor master data, reusing the existing RAG pipeline rather than a bespoke index.
Eval orchestration, drift monitoring, and trace storage are platform services pulled from the catalog, not project code.
9 weeks from design sign-off to production instead of an estimated 6 months greenfield — because the platform components already existed.
Policy becomes architecture, not paperwork
Medium-tier review completed inside the 2-week SLA — governance as a scheduled gate, not an open-ended queue.
The policy "payments over $50k need human sign-off" is implemented as a mandatory HITL approver node in the agent graph. The agent physically cannot release a large payment alone.
Trace capture on 100% of decisions — every extraction, match, and approval is replayable for audit and incident review.
Run it, catch the drift, attest the number
Touchless rate reaches 61% — ahead of ramp plan.
A major new vendor introduces an unseen invoice format; extraction accuracy dips on that segment. Caught by the eval-drift alert, golden dataset refreshed, accuracy restored within days.
Touchless rate at 74% — within a point of the 75% target.
Finance attests $5.8M annualized value against the $7.2M model. Honest gap noted: capacity release in EU shared services ran slower than modeled. The gap is recorded, not hidden.
Freed AP capacity redeployed per the workforce-transition policy — no layoffs. The attested savings fund two H2 follow-on use cases entering the same intake funnel.
The golden dataset and thresholds existed before the first line of agent code — and later caught the month-5 drift.
Gateway, RAG, and eval services from the catalog turned a 6-month build into 9 weeks.
Two sibling candidates died at the gate so this one could get real funding and focus.
The $5.8M is Finance-attested against a signed baseline — including the honest gap versus the $7.2M model.
Freed capacity moved to higher-value work under the workforce-transition policy, sustaining trust for the next wave.
© 2026 Ramesh Kaluri. All Rights Reserved.