“Which of our customer contracts…?” Ask for an AI answer that shows it checked every one

Ask an AI assistant which of your customer contracts allow a price adjustment at short notice, and you get a confident list. Nobody can see which contracts it skipped. You can act on an answer that also states how many were checked, which could not be read and which need a person's judgement. That answer comes from an agent that queries structured contract records in a business ontology, not one that searches documents.

A completeness receipt is the tally attached to a portfolio answer. It sorts every record into compliant, in breach, ambiguous or unreadable, and the four groups must add up to the number scanned. AWS published the idea on 2 October 2026 in its Adjudicated Query pattern, where every compliance sweep produces one.

Why does a document-search assistant miss contracts?

Document-search assistants rank the passages most similar to your question. AWS notes that similarity search "has no threshold that means all of them." A model writing its own database query adds a quieter risk: one invented filter can silently shrink the set being counted while the number still looks exact.

Research on whole-collection questions agrees. On the GlobalQA benchmark (October 2025), existing RAG methods performed poorly on counting, extremum, sorting and top-k tasks; on Qwen2.5-14B the strongest baseline scored 1.51 F1. AGGBench calls these find-all questions; S-RAG answers them by structuring the corpus at loading time. For a contract portfolio the lesson is direct: structure the documents before you ask portfolio questions.

Population questions (which of our contracts, how many, are all of them) are wrong if one contract is missing, so they need a receipt. Lookup questions, such as what clause 7 says, do not; ordinary retrieval serves them well.

How a business ontology makes a complete answer checkable

The Adjudicated Query pattern gives the chat model two jobs: translate a question into a call on a fixed set of typed operations, and narrate the result. A deterministic engine applies versioned rules, and the sample code asserts that the four groups add up to the number scanned before a sweep is saved. On AWS's synthetic lease data, one sweep returned 10,111 violations, 689 ambiguous and 20 unreadable. Those figures show what a receipt looks like, not real-world accuracy.

FAIS applies the same discipline through its foundational business ontology. An ontology is a shared model of your company's things (customers, products, contracts), how they relate and which actions are permitted on them. Each customer contract becomes a record linked to the customer and product records in your ERP and CRM, holding the few terms your decisions depend on, each tied to its source clause. FAIS builds agents that answer portfolio questions by running predefined queries over those records. You gain:

• Answers you can forward. The tally travels with the list, so a director sees coverage at a glance.
• Less re-reading. Each term is extracted once and reused.
• Named exceptions. Ambiguous and unreadable contracts go to a named reviewer instead of dropping out. In AWS's sample, filing a review or override is a human act that needs a named reviewer and a reason; no tool lets the model do it.
• A reusable foundation. The same records serve renewal checks, price notices and the agents you add next year.

A price-review question, before and after

Consider a Taiwan component maker whose material costs have risen. Its sales director asks which customer contracts allow a price adjustment on 30 days' notice or less.

Before: the assistant returns the contracts whose wording best matches the request. Nobody knows whether older scans were read or a one-calendar-month clause was counted. Sales re-checks every contract by hand before calling anyone.

After: each contract record holds its notice period in days. One predefined query returns four groups (meets the condition, does not, ambiguous, unreadable) that together equal the active portfolio. A clause promising only reasonable notice goes to the account owner; unreadable scans go to whoever holds the originals. The director starts calls from the clear list while exceptions are resolved.

What to require before you trust a complete answer

1. An exact portfolio boundary. A receipt proves coverage of the population you defined, so name it precisely: for example, active customer contracts signed by your Taiwan entity.
2. A few clean terms. Numbers, dates and yes/no fields a rule can compare. Leave judgement terms to people.
3. A receipt that adds up. If the four groups do not sum to the total, there is no answer yet.
4. A named owner for exceptions, with a date to clear them.
5. Planted cases as an acceptance test. Add contracts with known answers, including one unclear clause and one poor scan, and require the agent to find each. AWS's sample seeds its synthetic corpus with a counted set of violations.

Where to start this quarter

Pick the portfolio where a missed document costs most. FAIS can help you separate clean comparisons from terms that need judgement, model them in your company ontology and connect an agent that reports what it checked and what it could not read.

Trust an AI answer about your whole portfolio only when its receipt adds up and it has found every case you planted.