Bluestream announce Adobe partnership and FrameMaker Connector
- Written by Rahel Bailey
Technical documentation teams are expected to do more with less, work smarter rather than harder, and somehow protect quality while the resources to maintain it shrink. They are usually the first to be cut and the last to be consulted. For years the squeeze has been uncomfortable but survivable. Agentic AI changes the math entirely, but not in the direction you might assume.
The reflex among leadership is to treat AI as a way to reduce headcount. The more interesting argument looks in the other direction: the software development lifecycle is becoming agentic, swarms of agents are already scanning thousands of tickets, interpreting requirements, writing code and proposing commits, and the volume of documentation this acceleration produces will overwhelm any team that tries to keep up by hiring. The pressure is not “do the same work with fewer people.” It is “absorb a tenfold increase in output without drowning.” That reframing is important. It determines who gets to define the future of the role. Right now, if documentation professionals don’t claim that ground, other stakeholders with often competing interests will define it for them.
AI alone is not agentic
Most documentation teams currently asked to use AI, without much guidance other than that. Teams are given an LLM, without analysing which LLM is best for the task, and expected to start prompting. That was the first generation, where for a brief moment we thought that prompt engineering might become its own career path. It was the equivalent of digging with a hand shovel to build the foundation for a skyscraper. The second generation was building workflows—for example, synthesising resolved support cases into knowledge base articles to deflect repetitive queries. These workflows were definitely useful, but not agentic.
What makes a system agentic is something more specific: an orchestrated set of agents working in concert, each with defined roles, defined skills, and memory. A single model has no memory of what has already happened in a transaction, however much it appears to. An agentic content supply chain has many agents operating across the length of the workflow, from a product’s inception to the delivery of its content. Crucially, the skills those agents carry should be defined and maintained by the documentation professionals who actually understand what good documentation requires, rather than developers or engineers whose expertise lies in brute-force automation.
The Precision Paradox
Generative AI is probabilistic. It finds the closest match across vectors and does so impressively well. That was tolerable when a chatbot was expected to be right 70 or 80 percent of the time. In an agentic system, where many agents must each make correct decisions in sequence, a 20 percent error rate can be catastrophic. Mistakes cascade down the chain and can multiple exponentially as the number of agents replicate and magnify them.
This is the precision paradox: once output passes roughly 80 per cent quality, the final stretch is the hardest to close, and tolerance for error collapses the nearer you get to 90 or 95 per cent. The lesson is not that the model is inadequate, because it is good at what it is supposed to do, but that it cannot do the work alone. Content has to be curated so that the model doesn’t receive extraneous context it shouldn’t handle. Validation is required on both the input and the output: the “garbage in” must be caught before it becomes “garbage out”, and what comes out must be checked for accuracy and tested against the actual product.
This is also where the distinction between probabilistic and deterministic systems becomes practical rather than academic. Earlier expert systems were deterministic and traceable; you knew where an answer came from. Agentic documentation needs to recover that provenance, which is why knowledge graphs matter so much. A graph is deterministic; it preserves the relationships, links, metadata and semantics that a vector database discards when it shreds content into fragments. For example, you will want to preserve the difference between a warning and a caution, or between an instruction for a particular product in a product line. Querying a graph lets a team retrieve exactly the right piece of content to hand to a model, which is both more accurate and also dramatically cheaper, as it uses far fewer tokens.
Our Recent Webinar
Before we go further, this is exactly the shift we unpacked in our recent video on Agentic AI: why it is not just “AI for documentation,” but a different way of thinking about content operations, governance, knowledge structures and the role of documentation teams in an agent-driven SDLC. If you want the broader argument in a more conversational format, including what this means for content teams trying to protect quality at scale, you can watch the video here: Agentic AI video.
Pushing Responsibilities for Knowledge to the Source
One of the most consequential shifts is organisational rather than technical. For years, documentation professionals have acted as researchers, sleuthing for information that subject-matter experts didn’t, or didn’t realise they should, provide. For effective agentic workflows, this needs to be inverted. Responsibility for supplying the product knowledge, from requirements to prototypes, design files, code, and even recorded product walkthroughs, should move up the chain to the product managers and SMEs who have and own that knowledge. If they approve poor input and bad documentation results, accountability for accuracy needs to sit with them, not with the documentation team. Their own responsibilities should be concentrated on what they know best: ensure that the output of the agent is relevant, accurate, informative, timely, engaging, and standards-compliant (meets all technical and regulatory requirements).
To make that work, teams need governance: clear criteria for what good input looks like, agents that check collateral against those criteria, and a structured way to manage the prompts, templates and skills the agents use. Externalising those skills to keep them under the direct control of documentation professionals is what keeps ownership where it belongs.
New Roles, Not Disappearing Ones
Jobs for technical writers whose jobs are focused on drafting content from scratch will shrink. The role will moves toward user research, information curation, knowledge engineering, working with taxonomy and ontology managers and other context engineers.
This begs the question as to whether the role will still be called a technical writer, or whether the title will change to reflect the areas that these professionals are actually responsible for. Their roles will likely span operational governance, quality assurance and, most importantly, the design of the agentic systems themselves. The job starts to blur with that of a content operations strategist, who translates between operational needs and what engineers build, designing the workflow, and then governing it. That lets engineers concentrate on implementing what documentation professionals specify rather than trying to build by mind-reading what is needed.
While teams don’t need to learn how to code, they will need to understand the concepts well enough to direct the people who do. It will be important to be able to understand how to help engineers to build a knowledge graph of the content and specify the constraints and the relationships that matter. That requires a deep understanding of structured content, how ontologies and context graphs work, and emerging standards. The soft skills needed are how to negotiate in collaboration with engineers and other stakeholders. We know from experience that developers assume the documentation process is simple and may cut corners, accumulating content debt that becomes expensive to remediate later.
AI alone is not agentic
Technical documentation teams are expected to do more with less, work smarter rather than harder, and somehow protect quality while the resources to maintain it shrink. They are usually the first to be cut and the last to be consulted. For years the squeeze has been uncomfortable but survivable. Agentic AI changes the math entirely, but not in the direction you might assume.
The reflex among leadership is to treat AI as a way to reduce headcount. The more interesting argument looks in the other direction: the software development lifecycle is becoming agentic, swarms of agents are already scanning thousands of tickets, interpreting requirements, writing code and proposing commits, and the volume of documentation this acceleration produces will overwhelm any team that tries to keep up by hiring. The pressure is not “do the same work with fewer people.” It is “absorb a tenfold increase in output without drowning.” That reframing is important. It determines who gets to define the future of the role. Right now, if documentation professionals don’t claim that ground, other stakeholders with often competing interests will define it for them.
Build the Content Foundation for Agentic AI
If your content team is starting to explore what agentic AI could mean in practice, the most important step is not choosing a model. It is getting your product knowledge into a structure that agents can reliably use.
At Bluestream, we use DITA to structure content for our XDocs DITA CCMS, helping organisations build the governed, reusable and semantically rich content foundations needed for agentic workflows.
