A sample Gate Zero Readout.

Page 1 · Decision memo

Gate Zero Readout — Arvela Logistics · March 2026 · Prepared by Luca-Ștefan Tamaș, LT Strategy Partners

Ship the shipment-document assistant. Consolidate the knowledge base before any customer chatbot. Don't build the pricing engine.

Of the three use cases assessed, one is ready to build now: an assistant that reads booking emails and attached shipment documents and drafts structured entries for the operations team to confirm. The other two fail Gate Zero today — one temporarily, one on design.

The document assistant clears the gate. Five operations staff spend roughly nine hours a week each re-keying data from PDFs and emails — about €54,000 a year at a conservative base case, using Arvela's own loaded-cost figures. The grounding content exists, is accessible today, and every output is confirmed by a human before it enters the TMS — so the error budget is honest.
The support chatbot is “not yet”, not “no”. The answers customers need live in seventeen inboxes and a wiki last updated fourteen months ago. Built today, the assistant would confidently repeat stale answers. The unlock is organizational, not technical — and it's named on page 4.
The pricing engine fails on stakes, not on ambition. A wrong extraction gets caught by the operator who confirms it. A wrong price goes straight to a customer. With spot-quote history living in email threads, the error rate cannot be validated to the tolerance the workflow demands — this is where AI won't pay off for Arvela as scoped.

What leadership already suspected, confirmed with numbers: the re-keying burden your operations lead has flagged for two years is the strongest AI case in the company — stronger than either of the two ideas that came from vendor conversations.

Owner & 90-day step Operations director. Scope the document assistant against the gate checklist on page 3; in parallel, assign one owner to the knowledge base with a 90-day consolidation target.

Page 2 · The opportunity portfolio

Three use cases, one map.

Value at stake → Production viability →

Dimension scores at walkthrough depth

Gate Zero dimension	A · Documents	B · Chatbot	C · Pricing
Value at stake	G	A	A
Data reality	G	R	R
Grounding & retrieval	G	R	R
Error tolerance vs stakes	G	G	R
Validation & verifiability	A	A	R
Cost, latency & ownership	G	A	A
Security & EU AI Act exposure	A	A	G

G On footing — no blocker found at walkthrough depth
A Addressable gap — named condition must be met
R Gating problem — blocks a build verdict as scoped

Map badges show each use case's provisional EU AI Act tier. Additional flags — like margin-critical on C — mark business stakes; they are not AI Act tiers.

A / Validation (amber): extraction accuracy needs a 200-document ground-truth set before launch — one week of operator time, scoped in the gate checklist.
A / Exposure (amber): shipping documents carry personal data; processing must stay in the EU, and the extraction step should run self-hosted or in an EU region with no third-country transfer.
B / Value (amber): the interruption cost is real but unmeasured — the base case needs two weeks of ticket counting before it can be banked.
B / Data reality (red): no single, current source of the answers exists — they're scattered across seventeen inboxes and a wiki untouched for fourteen months.
B / Grounding & retrieval (red): fragmented, stale content is precisely what retrieval tuning cannot rescue — the score follows the anchor until the knowledge base is consolidated.
B / Validation (amber): correctness is checkable by sampling against the ops team — but only once a consolidated source defines what a correct answer is.
B / Cost & ownership (amber): query volumes and per-answer economics can't be modeled until the assistant has a source to answer from, and no owner is named yet.
B / Exposure (amber): customer-facing, so EU AI Act transparency obligations apply, and transcripts would carry customer personal data — both addressable at design time.
C / Value (amber): the margin upside is plausible but can't be validated from unstructured quote history — value that can't be measured can't be banked.
C / Data reality (red): spot-quote history lives in email threads; no structured record exists to train or validate against.
C / Grounding & retrieval (red): the pricing context that would ground a quote — lane, season, client history — isn't retrievable from any current system.
C / Error tolerance (red): a wrong price reaches a customer with no human between the model and the harm — the workflow as designed has no error budget.
C / Validation (red): with no ground truth, the error rate cannot even be measured to the tolerance the workflow demands.
C / Cost & ownership (amber): per-quote economics look fine on paper, but they're meaningless while the error rate is unmeasurable.

The binding constraint at Arvela: data consolidation, not ambition, gates everything after use case A.

Gate Zero weights are fixed before scoring, by design — what matters most to your business is agreed with you earlier, during the diagnosis. No aggregate score is reported — the lowest dimension gates; averages hide.

Page 3 · Top use case — brief & formal gate

A · The shipment-document assistant.

Business process: Inbound booking requests: email + attached PDFs (bills of lading, packing lists, customs forms) re-keyed into the TMS by the operations team.
Impact, conservative base case: 5 operators × ~9 hrs/week re-keying × Arvela's loaded cost ≈ €54,000/year, before error-correction time. Assumption declared: volumes from the ops lead's estimate, not measured logs.
Buy / integrate / build: Integrate: an extraction model behind a review UI, wired to the existing TMS API. No custom model training warranted at this volume.
Data readiness: Green at walkthrough depth — documents accessible in the shared mailbox today; ground-truth set needed pre-launch (see gate).
Cost band & time to production: Low five figures, 8–12 weeks to a production system with validation, monitoring, and handover. Range, not a quote — a fixed price follows a scoped audit.
Success metric: Hours of re-keying per 100 shipments, measured before and ~30 days after launch.

The gate

Build now — subject to the checklist below.

Capability target: ≥95% field-level extraction accuracy on the 200-document ground-truth set before go-live; every output human-confirmed until then and sampled after.
Safety target: no auto-commit to the TMS — operator confirmation stays in the loop until three consecutive weeks inside the error budget.
Cost at scale: unit cost per processed shipment stays under the agreed ceiling at 2× current volume; measured, not assumed.
Sign-off domains: operations owner · security/data-protection · engineering · the operators who use it.
Walk-away conditions: if the ground-truth set can't reach the accuracy floor, or EU-only processing can't be met at acceptable cost, stop — the readout's verdict flips to “not yet” and says why.
What would change this verdict: a TMS migration mid-build, or volumes 3× below the estimate used in the base case.

Confidence: scores marked at walkthrough depth only — verified against described systems, not inspected pipelines. The paid audit tests these against the actual mailbox and TMS before any build.

Page 4 · Where AI won't pay off for Arvela

The honest no's — and what would change them.

Not yet B · Customer support assistant

The value is real — status queries interrupt the same team that re-keys documents. But the grounding content fails Gate Zero: the answers live in seventeen inboxes and a wiki last touched fourteen months ago. An assistant built on that would answer confidently and wrongly, in front of customers, under the EU AI Act's transparency obligations for customer-facing bots. The unlock: one owner, one consolidated knowledge base, a 90-day target — then this use case re-enters the gate with a realistic path to “build”.

Don't build C · Dynamic pricing engine

Not because the model can't produce prices — because the workflow has no error budget and no validation path. A wrong extraction is caught by an operator; a wrong price reaches a customer directly and costs margin or trust. The spot-quote history that would train and validate it lives in unstructured email threads, so the error rate can't even be measured to the tolerance required. The cheaper answer that survives scrutiny: a rules-based margin floor with quarterly review — nearly free, no surprise behavior. What would change the verdict: two years of structured quote/outcome data and a human approval step that doesn't erase the speed advantage.

Page 5 · Gaps, next steps, and the map forward

What happens next — with us or without us.

Assign the knowledge-base owner (yours to do, this week). No engagement needed — one name, one 90-day consolidation target, and use case B stops being blocked.
Build the ground-truth set (yours to do, ~1 week of operator time). 200 recent shipments, fields verified by hand. Whoever builds the assistant will need it — us or anyone else.
Scope the document assistant against the page-3 gate. The open questions — actual extraction accuracy on your documents, EU-only processing cost — are exactly what a fixed-scope review of the candidate against these gate conditions tests, on your real documents.
Re-gate the chatbot in 90 days. If the knowledge base lands, B re-enters at “build” candidacy; if it doesn't, that tells you something too.

If you want help, this is how each gap maps

Extraction accuracy & pipeline reality → a fixed-scope review of use case A against the page-3 gate conditions, on your real documents.
EU-only / no-transfer processing → a Load-Bearing Review (security & EU AI Act), scoped by the exposure tier already assigned above.
Taking A from verdict to production → implementation and delivery: one workflow, built, integrated and handed over, measured against the page-3 metric.

This document is yours either way.

Appendix · The Gate Zero rubric

The anchors are the method. Nothing behind the curtain.

Every dimension is scored 1–5 against observable anchors — the same anchors in every readout, fixed before scoring begins. Score 5 and score 1 are shown here for each dimension; the full anchor set travels with the readout.

Value at stake: 5: euros or hours attached to a named workflow, from the client's own figures. 1: “efficiency” with no workflow named.
Data reality: 5: grounding content exists, is clean, and is accessible today. 1: it exists in someone's head.
Grounding & retrieval: 5: structured, current, chunkable content a retrieval system can serve at production quality. 1: fragmented or stale content no retrieval tuning can rescue.
Error tolerance vs stakes: 5: wrong outputs land on a reviewer with time to catch them. 1: wrong outputs reach a customer, a regulator, or a ledger directly.
Validation & verifiability: 5: correctness checkable automatically or by cheap sampling. 1: no way to know it's wrong until it costs something.
Cost, latency & ownership: 5: unit economics hold at 2× volume and a named owner exists. 1: per-query cost unknown and nobody would be paged.
Security & EU AI Act exposure: 5: no personal data, minimal-risk tier, deployment unconstrained. 1: personal or regulated data with an unresolved transfer, or Annex III adjacency unexamined.

Exposure scoring cross-references the EU AI Act (Article 6 / Annex III), the NIST AI Risk Management Framework, and the OWASP Top 10 for LLM applications. Declared assumptions from the session travel with the readout as a register.

This is the document your leadership team receives — built on your numbers, your systems, and your use cases.

Start with the Business Diagnostic