Agentic search: your AI agent can follow the cross-references in your procedures. Where do they lead?

With agentic search, an agent follows the cross-references in your procedures. If the target is missing, it may settle on the closest-looking document.

A maintenance technician asks an AI agent for the inspection interval of a transformer. The maintenance procedure answers with a cross-reference: “as per technical note NT-07”. Agentic search changes what the agent can do with that sentence. Beyond the first excerpts it retrieves, it can keep searching, open, reread and search again, until it judges it has gathered enough evidence to answer. The twist fits in one sentence for the steering committee: if a cross-reference in your procedures no longer leads to an identifiable document, a person calls a colleague; an agent that keeps searching may settle on the closest-looking document.

Agentic search is multi-step retrieval driven by the model itself, which decides what to look for, what to open and when to stop. This article explains what changes between an assistant’s classic retrieval and agentic search, and why the links between documents, the “see”, “in accordance with”, “as per note”, become part of your AI agents’ context. It is written for CDOs and agent programme leads preparing go-lives for 2027.

The person this text makes uncomfortable is the CDO who will present the maintenance agent to the go-live committee. The site’s technical director will ask a simple question: “If the agent cannot find the note it cites, what does it do?” The answer depends on the agent, and on where the cross-reference leads.

Two answers seem to cover the topic already. The first is the tool you have bought: a sharper search engine, or a document management system that keeps version history. The engine finds a document whose title resembles “NT-07” faster; the DMS knows a note was renamed or merged, if someone recorded it there, yet the procedure that cites it still carries the old name. The second is the in-house process: the periodic review of procedures, which checks each document’s content. It rereads the text of a procedure; it rarely opens every document that text points to, and that is exactly the path an agent takes.

In our Education series, the 30 October article explained that an agent skill is a procedure that needs approval. This one changes the object: it is about the path the agent travels between documents, and what happens when that path is cut. It also differs from our 7 October piece on withdrawn documents that AI keeps citing: here, the cited document can no longer be found under its name, and the agent looks for a substitute. Finally, our 9 November article was about the statements written on a policy (official version, owner, effective date); this one is about the link between two documents, the cross-reference, which those statements do not repair.

Classic retrieval, agentic search: what changes for the agent

In an assistant’s classic retrieval, often called RAG, the system splits documents into chunks, retrieves those closest to the question, and the model answers from that batch in a single pass. If the batch contains the sentence “as per technical note NT-07”, the model cannot go and open the note. It answers with what it has, or reports that something is missing.

Agentic search gives the model tools to keep going. Mistral AI described it on 20 August 2026 when introducing its Agentic Search service¹, and its best form deserves to be described as the vendor writes it. The model has five tools close to a human reader’s gestures: search the corpus, open a document, move to a section, read a passage, find an exact term in the open document. The vendor explains that one-shot retrieval fails when the model must follow references, compare several documents or verify the evidence, and that the loop lets it refine the query, open a promising document, follow a reference and try another route when the first results fall short. Its documentation states that the loop continues until the model has enough evidence to answer, recommends bounding the number of hops in agent definitions, and notes that these tools can be plugged into any agent compatible with the Model Context Protocol (MCP)²: the mechanism described here is not limited to one vendor’s agents.

These tools address a real limit. An agent that can open the cited note, go to the intervals table and read the right row does what an experienced technician would do. The vendor publishes accuracy gains measured on two public benchmarks. Those benchmarks cover published, dated documents, filings submitted to the US securities regulator and US Treasury bulletins; they measure the ability to navigate. The state of cross-references in an internal document estate is outside their scope.

Mistral AI is not alone in this space, and this article does not compare offerings. It keeps one mechanism that a vendor describes precisely: the agent follows references, and it stops when it judges it has enough evidence. Applying this mechanism to the cross-references between your internal documents is our reading; Mistral AI does not make it.

What the maintenance agent reads, and what changes when cross-references are reliable

Back to the task. The agent prepares the work order for a transformer inspection. It opens the preventive maintenance procedure, reads the “Intervals” section and finds the cross-reference: “as per technical note NT-07”. It searches for “NT-07”.

Suppose the note was merged, during a reorganisation, into a maintenance standard shared by several sites, under another title. The procedure was not changed: its content was right, only its cross-reference aged. The search for “NT-07” returns no exact match. Nothing then stops the agent from doing what the loop allows: it rephrases, searches for “technical note transformer inspection intervals”, and finds a note with a similar title, written for another site, with other equipment and other intervals. It is consistent, signed, current for its own scope. The agent reads the table, may judge it has enough evidence, and fill in the work order. This sequence is a scenario built for the explanation, not a measured behaviour: nothing in the vendor’s documentation describes this case, and the agent’s instructions can prevent it.

Nothing in this sequence is a failure. Every document is valid. The agent followed the path the text gave it, then the nearest route when the path ended. A human technician facing the same dead reference would have called the site’s maintenance lead. The agent has no such reflex unless its instructions give it one.

The second effect is less visible. If the programme team has bounded the number of hops, as the documentation recommends, each search spent looking for a missing document is one step fewer for checking the answer. In the example the vendor publishes, the model limited to a single search finds only half of the requested months, and says so. An agent whose instructions do not require that report may answer with what it has.

What changes when cross-references are made reliable comes down to three conditions written into the documents. The cross-reference names an identifiable target: a stable reference and, where relevant, the words “version in force”. The target exists under that name, or a written mapping says where it now sits: “NT-07, merged into the maintenance standard, section 4, since 1 March”. The target states the scope it applies to: site, equipment, activity. The agent can then open the right document directly, search for the right term, and stop on evidence a person has designated rather than on the closest-looking one.

This is how the agentic transition speeds up. A repaired cross-reference is repaired for every agent that will read the procedure: the maintenance agent today, the planning or spare-parts purchasing agent tomorrow. Each step the agent no longer spends hunting for a missing target remains available to check its answer. This gain is not measured; you verify it with the second agent, by comparing how many searches each one spends on the same procedure.

At the committee, the CDO can then answer the technical director: when the cited note cannot be found, the agent stops and reports it; and for the three procedures in the task, a person has followed every cross-reference once.

What the field shows: 19% of pages to correct in an official reference set

At TotalEnergies Retail Power & Gas, the customer chatbot in production relied on around 500 pages of official web documentation. These ~500 pages were mapped and audited, without migration; 19% of pages turned out to need correction, including cases impossible to spot by eye at scale; 53% of cases were resolved in 3 weeks, prioritising the most critical, with 1 to 2 experts working half a day to a day per week. Thomas Bensoussan, Head of Digital Products, notes that the approach surfaced conflicts that were hard to spot by eye, and that the recommendations indicate which expert each case should go to³.

What this measurement establishes: in an official, published reference set of around 500 pages, 19% of pages needed correction, and 53% of cases were resolved in three weeks with 1 to 2 experts. What it does not establish: the share of faulty cross-references among those corrections, which was not counted separately, or the behaviour of an agent using agentic search, since this was a chatbot. It shows that a maintained, official reference set can contain defects nobody saw, in one case. Not necessarily in yours.

The cross-reference log: three procedures, one row per cross-reference

The gesture is to walk once, by hand, the path the agent will walk.

What it produces. A cross-reference log: one row per cross-reference found in the task’s procedures, and five columns. Source document and section. Target cited, as written. Target status: found as is, renamed or merged with a written mapping, renamed without a mapping, not found. Target scope: applicable or not to the agent’s scope. Owner who corrects the cross-reference.

Where it lives. The agent’s go-live file, kept by the programme team, next to the list of its sources.

When it is written. Before setting, in the agent definition, the library it queries and the number of hops it may use. If the agent is already open to a pilot, before extending it to other sites.

The sample. The three procedures the agent’s instructions cite or assume. To find the cross-references, a text search in each procedure is enough with existing tools: “see”, “cf.”, “in accordance with”, “as per”, “note”, “appendix”. If the task relies on fewer than three procedures, all of them.

A completed row. “Preventive maintenance procedure for transformers, section 3; target cited: technical note NT-07; status: merged into the maintenance standard, section 4, no written mapping; scope: shared standard, applicable; owner: site maintenance lead.”

How to read it. A “found” and “applicable” row: the cross-reference can be followed. A “renamed without a mapping” or “not found” row: the owner writes the mapping or corrects the cross-reference. Meanwhile, the agent’s instructions specify that on this cross-reference it stops, cites the procedure, reports that the target cannot be found, and a person answers.

How it fails. The log only sees named cross-references. An implicit one, “according to the rules in force”, “in line with site requirements”, has no target to follow: it goes on a separate row, “unnamed cross-reference”, which the owner specifies or accepts. The dated log also becomes a record: it shows that on a given date, the company knew a cross-reference led nowhere. Each open row therefore gets an owner and a date; the legal weight of such a record is for your legal department to assess.

Its budget. The 2027 agent programme, under source preparation. Correcting a cross-reference is a minor change to the source document, through its owner’s existing revision cycle.

Audit, clean, monitor

The log covers three procedures. An agent programme opens hundreds, and each may point to several others. That is the purpose of a distinct category, the Document Knowledge Platform (DKP): Govern, governing the document estate (ownership, authority, lifecycle); Clean, detecting and handling anomalies, duplicates, obsolescence and competing versions; Activate, opening the corpus to AI systems only once the first two steps hold. A DKP replaces neither an enterprise search engine, nor an agentic search layer, nor a DMS, nor an agent platform. It runs before them, on the documents they consume.

Audit. Count obsolete documents and those whose applicable version cannot be identified, in the libraries the agent will query. To link a cross-reference to its likely target, the tool proposes candidate documents; the team confirms. The reading covers document content only, within a contractually defined ingestion scope approved by the business Document Owner and the CISO or DPO, with no usage logs or conversations, and no reuse for model training.

Clean. Route each case to the document’s owner, who decides. Business experts keep control of the decision.

Monitor. A note that is merged, renamed or withdrawn breaks the cross-references that cite it. The log is updated at each such revision; otherwise the June agent will follow the November path.

The CDO can then answer the technical director about the agent and about every cross-reference in its task.

Frequently Asked Questions

What is agentic search, in one sentence?

It is multi-step retrieval driven by the model: it searches, opens a document, reads a passage, searches again, and stops when it judges it has enough evidence to answer. Classic retrieval answers in a single pass from the first excerpts found.

Why do cross-references between documents matter more with an agent?

Because an agent using agentic search can follow a cross-reference to the cited document, where a one-pass assistant stopped at the sentence containing it. If the target cannot be found, the agent may continue searching and settle on a nearby document that nobody designated.

Doesn’t capping the agent’s number of steps solve the problem?

Capping steps limits the time and cost of the search. It says nothing about where the cross-reference leads: a capped agent that hits a missing target simply has fewer steps left to check its answer.

Who should fix a cross-reference that leads nowhere?

The owner of the document containing the cross-reference, through their usual revision cycle, with help from the owner of the cited document when it was renamed or merged. The agent programme team keeps the log; it does not edit the documents.

Does a DKP platform read the content of our internal procedures?

Yes, within a contractually defined ingestion scope approved by the business Document Owner and the CISO or DPO. Only document content is read, with no usage logs or conversations, and nothing is reused to train models.

Notes

¹ Mistral AI, “Introducing Agentic Search”, 20/08/2026: five tools (“search, open, navigate, read, and grep”); limits of one-shot retrieval when the model must “follow references, compare multiple documents, or verify the underlying evidence”; FinanceBench (SEC filings) and OfficeQA Pro (US Treasury bulletins) benchmarks. The vendor’s published figures are not repeated here. ² Mistral AI, “Agentic Search” documentation, read on 11/11/2026: “The loop continues until the model has enough evidence to answer the question”; “Bound the number of hops in your agent definitions to limit the loop depth”; usable “as a set of Model Context Protocol (MCP) tools that drop into any agent”. ³ K-AI, Customers page and TotalEnergies Retail Power & Gas case study.

Sources

Where to Go From Here

K-AI Corpus Diagnostic — 10 business days on your document estate, full report of the 20 most critical anomalies, money-back guarantee if no meaningful anomaly is found. A one-hour conversation starts from your cross-reference log: it locates, among the libraries your first agent will read, those where obsolescence and unidentifiable versions are likely to concentrate, and clarifies which procedures to start with. Reach the K-AI team: contact@k-ai.ai. The scope of every diagnostic is validated jointly by the business Document Owner and the CISO/DPO, never by IT alone.

K-AI already works with CMA CGM, Veolia, PwC, BNP Paribas, TotalEnergies and CEVA Logistics. Partners: AWS, Snowflake, Microsoft, Wavestone, Devoteam.