Context engineering: are your documents ready to become a production dependency for your AI agents?
Forrester says agent context becomes a production dependency. So do the procedures behind it: every revision changes what the agent applies.
On 19 October, Forrester published its definition of context engineering and of a companion discipline, ContextOps¹. One sentence captures what is at stake: as agents start executing business processes, the context they receive “becomes a production dependency rather than a supporting asset”. Among the sources of that context, Forrester explicitly lists enterprise documents.
We agree, and the consequence for documents is direct: if context is a production dependency, the procedures, policies and reference files that feed it are production dependencies too. If a procedure an agent reads can change without anyone knowing which agents read it, every revision becomes an undeclared release to production. And whatever a revision carries outside the text itself (the cover note, the effective date announced by email, the training session) stays invisible to the agent.
This is context engineering for AI agents seen from the document side. For a CDO preparing the first agents of 2027, the question comes before go-live: do your documents have the attributes of a production dependency? Two answers come to mind. The existing document approval workflow, which already has every version reviewed and signed off. And the agent platform, whose observability tools show what the agent consulted. Both are useful; neither tells the owner of a procedure, at the moment they revise it, which agents will apply the new version the next morning. Our position: this work is documentary, it is done once, and every later agent benefits. It also shortens go-live, because an agent whose documents carry their effective date does not need to be suspended or blindly retested at every revision. If you have no agent planned before 2028, the argument can wait.
What does “context becomes a production dependency” mean?
Forrester defines context engineering as the practice of designing, orchestrating and delivering structured, relevant and timely context to AI systems so that their outputs are accurate, grounded and compliant¹. Its four elements: retrieval over enterprise documents and data (RAG), function calls to applications, memory, and a context model that unifies them. ContextOps is the discipline that governs, monitors and improves that context across its lifecycle, including offline updates.
Forrester adds that agents “cannot rely on intuition, institutional memory, or human judgment”. For documents, that is the decisive point. An employee reading a revised procedure knows it has just changed; they may have received a cover note, sometimes a training session. An agent only has the text.
Applying the production-dependency idea to document revisions is our reading; Forrester does not make that application.
What does a procedure lack compared with a production component?
A software component that a production service depends on usually has four attributes: an identified version, an owner, a list of what depends on it, and a change process that warns those dependents before the switch. If your procedures carry a version and an owner, they have the first two. The other two had no reason to exist while their readers were people, who were informed through other channels.
Take a constructed example (no observed incident). An agent pre-approves purchase requests by reading the delegation of authority policy. Finance revises the policy: the dual-signature threshold goes down, applicable to orders placed from 1 January. That effective date sits in the cover note sent to buyers; the policy text itself does not mention it. On publication day, the agent reads the new version and may apply the new threshold to orders already in progress. Buyers know it only applies in January. The concrete result: committed orders sent back for a second signature, delayed deliveries, and a procurement team working around the agent to hold its deadlines. The agent read an incomplete text correctly.
The mechanism generalises: whatever a revision entrusts to a channel other than the text (effective date, transitional provisions, scope, superseded version) is invisible to the agent.
Why don’t the tools already in place settle the question?
Each existing mechanism covers part of the problem, at its best:
- The document approval workflow has every version reviewed and approved by the right people. It validates the content for human readers. It does not ask which agents read the document.
- Document management versioning records who changed what, and when. It does not say who reads the document, or from when the new version applies.
- Agent platform observability can show, after the fact, which document an agent consulted for a task. It does not come into play when the document owner decides to revise it.
- Context engineering itself, as Forrester defines it, organises how context is delivered to the agent. It assumes the documents delivered carry their status and effective date; it does not write them in place of their owner.
What is missing comes down to two things: information written into the document, and one question added to the revision workflow.
How do you prepare your documents to become a production dependency?
There is no need to cover the whole document estate. Start with the first agent planned for 2027.
- List the documents its task depends on. Before go-live there are no usage logs yet: the list comes from its instructions, its configuration and its declared sources. For the procurement agent: the delegation of authority policy, the procurement procedure, the list of approved suppliers.
- Write into each document’s text what an agent cannot guess. Version in force, effective date, transitional provisions, superseded version. One line is enough: Delegation of authority policy, version 7, applies to orders placed from 1 January 2027; supersedes version 6; owner: Finance. On 9 November we discussed the mentions that give a policy parity with an official figure; here, the mention that matters is the date from which the rule changes.
- Add one question to the revision workflow: which agents read this document? The answer is the list built in step 1. The document owner notifies the agent owner before publication, the way teams are warned before a component changes.
- Before publication, check what the agent will do with the new text. Three real cases are enough: an order in progress, a new order, an edge case. If the agent cannot decide from the text alone, it is the text that must change.
The expected effect: an agent that goes live earlier because its documents are ready, fewer gaps at the switch, revisions published without freezing the agent, and work that serves every later agent reading the same document. We have not measured this on revisions read by agents; it can be checked at the first revision of a document an agent already reads.
What does field evidence show about bringing a corpus back into shape?
This case serves here to show that remediation could be carried out with limited resources, which matters if the exercise has to be repeated at each revision.
At TotalEnergies Retail Power & Gas, the customer chatbot relied on around 500 pages of official web documentation². 19% of pages needing correction were identified by K-AI, including cases impossible to spot by eye at scale. 53% of cases were resolved in 3 weeks, prioritising the most critical, with limited resources: 1 to 2 experts, half a day to a day per week.
What this evidence does not establish: it is a first diagnostic, on a chatbot and not on an agent executing a task, and it does not measure the cost of revisions followed over time.
Where does a Document Knowledge Platform fit?
A Document Knowledge Platform (DKP) is the document quality layer that runs upstream of agents, on the documents they consume. It governs (ownership, authority, lifecycle), cleans (duplicates, obsolescence, competing versions) and only activates the corpus for AI once those two steps hold.
It does not replace the agent platform, ContextOps tooling or document management. It works on what they read. A DKP does not belong in the ContextOps category Forrester describes: context engineering is the channel through which the document problem becomes visible, and the DKP deals with that problem at its source. In this work, K-AI counts anomalies and identifies the applicable version and obsolescence. Linking a document to an agent’s task takes judgement: the tool proposes candidate documents, the team confirms. Processing covers document content only, within a contractual scope approved by the business owner and the CISO or DPO, with no reuse for model training.
Definitions of context engineering will evolve. The place of documents, upstream of any context given to an agent, does not depend on an analyst’s next publication.
In short
The agents of 2027 will read your procedures as a production dependency. Before they arrive, audit the documents their task depends on, clean up what a revision currently entrusts to channels other than the text, then monitor every revision by asking which agents will read it.
Frequently Asked Questions
What is context engineering for AI agents?
It is the practice of organising what an agent receives to carry out a task: documents, application data, memory and rules. Forrester defines it as delivering structured, relevant and timely context so that outputs are accurate, grounded and compliant. Internal documents are one of its main sources.
Does an AI agent notice that a procedure has been revised?
It reads whichever version it can access. It does not receive the cover note, the training or the hallway conversation that explain the revision to teams. Anything not written in the text, such as an effective date, does not reach it.
Doesn’t the agent platform already handle context updates?
It can refresh its index and trace what the agent consulted. It cannot write, in the owner’s place, the effective date of a rule or the version it supersedes. That information belongs in the document.
Who should keep the list of agents that read a document?
Each agent’s owner draws up the list of documents it depends on; the document owner checks it before each revision. The CDO arbitrates when a document has no owner.
Who approves the document scope analysed?
The business owner of the documents and the CISO or DPO, jointly, never IT alone. The analysis covers document content only, with no reuse for model training.
Notes
¹ Forrester, Charlie Dai, Leslie Joseph and Indranil Bandyopadhyay, “Context Engineering And ContextOps: The Missing Foundation For Scaling Agentic AI”, blog, 19/10/2026. ² K-AI, Customers page and TotalEnergies Retail Power & Gas case study.
Sources
- Forrester, “Context Engineering And ContextOps: The Missing Foundation For Scaling Agentic AI”, 19/10/2026: forrester.com
- K-AI, Customers page: k-ai.ai/en/clients
- K-AI, TotalEnergies Retail Power & Gas case study (PDF): k-ai.ai
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. The document list of a first agent already says a lot; a one-hour conversation shows which of those documents carry their effective date and owner, and which could change without the agent knowing. 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.
