Skip to content
Tony Wood
Home

Why Your AI Agents Will Choke On Your Docs (And What To Fix Before 2026)

Tony Wood |

Every time I watch someone new start, or look at how teams handle documentation, I think: how does anyone figure this out, especially if we expect AI agents and agentic workflows to sit on top of it.

The tools we use look normal if you have been around for a while. You know which shared Google Drive is “the real one”. You know that one manager keeps everything in local Word files. You know someone else lives inside Confluence, but only in a particular space that they once sent you in Slack.

Humans can survive this. Agentic systems cannot.

If you try to drop autonomous agents into that environment in 2025 and 2026, they will choke on your documents. Not because the AI is “not smart enough”, but because the way we work is built around human quirks rather than clear, enforceable rules.

This is a leadership problem, not a tooling problem.

The Mess You Have Today (And Why New Joiners Hate It)

Let’s be honest about how work actually happens.

  • Some people use a shared Google Drive and dump everything into one catch-all folder.
  • Some keep critical files on their laptop, then share one-off links when asked.
  • Others create PDFs and email them, then never update the original.
  • There are Word documents on local machines and on SharePoint.
  • There are Apple files on personal desktops that no one else can open properly.
  • There are Confluence pages scattered across spaces, sometimes in the “right” navigation, sometimes hidden three clicks deep.

There are policies, in theory. Naming conventions somewhere on an intranet page. A slide deck explaining “how we do documentation here”. In practice, everyone has their own way.

Recent analysis of information sprawl in enterprises is blunt about the cost of this. As one author puts it:

"A single source of truth is essential. Without it, employees waste countless hours searching across multiple platforms and repositories, leading to lost productivity and onboarding frustration."

If you have ever heard a new joiner ask “where the hell is everything”, this is why. They are not slow. They are trying to reverse engineer years of undocumented habits across tools that were never designed to work together.

The uncomfortable truth is that we tolerate this because experienced people can work around it. They know who to ask. They know which version “feels” current. They know the quirks.

Your AI agents will not.

Humans Bend Rules. Agents Cannot.

Under pressure, people do what works.

You have a deadline. So you save the file on your desktop, email a PDF, and tell yourself you will tidy it later. You believe your way is clearer. You route around the official process because you need to ship.

Multiply that by hundreds of people over years, and you get the documentation sprawl you see today.

The research is clear about the gap between how humans behave and how automated systems operate:

"While humans frequently adapt, ignore, or reinterpret rules under pressure, automated systems follow rigid logic, leading to consistent but inflexible outcomes."

That consistency is the whole point of agentic systems. It is also why they fail in human-shaped chaos.

An AI agent cannot decide that this time it will ignore the naming convention because the person who uploaded the file was stressed. It will not think “Jane always keeps the latest pricing sheet in her personal folder, not the shared one”.

It follows the rules you give it. Every time. Exactly.

If those rules do not exist, or only exist on a slide deck nobody follows, the agent fails. It will pick up old versions. It will miss critical documents. It will hallucinate to fill gaps. You will blame the AI, but the culprit is your operating system for knowledge.

This is not a technical bug. It is the inevitable outcome of trying to bolt agentic workflows onto processes that were designed around human flexibility instead of machine clarity.

Why Your Current Tools Block Agentic Workflows

Look at your stack through an agent’s eyes.

Tools like Google Drive, SharePoint, local folders, Apple productivity apps and sprawling Confluence spaces grew up to support human work first. They assume:

  • People can remember where things “normally” go.
  • People will notice when a file is outdated.
  • People can use eye and hand coordination to scan, scroll and guess.
  • People can ask a colleague when in doubt.

Agents do not have that backchannel. They see:

  • Multiple “final” versions of the same document.
  • Inconsistent file types for the same process.
  • Broken links and orphaned pages.
  • No canonical location that is always correct.

Industry leaders are starting to call this out explicitly. As one recent perspective on agentic AI success explains:

"Agentic AI systems demand standardization and seamless data flow; inconsistent or human-defined processes introduce errors and prevent effective automation."

You can see the pattern. The same sprawl that frustrates new joiners is exactly what blocks your AI agents from being useful.

So when leaders ask “why can’t we have agents triage support tickets, prepare board packs or onboard staff”, the answer is often awkwardly simple. Your documentation and processes are not ready.

The Tipping Point: From Human-Friendly Chaos To Agent-First Clarity

We have reached a tipping point.

For years, we optimised tools for human comfort.

We let everyone organise their own world. We tolerated “special ways of doing things”. We assumed people would remember, adapt or fix it on the fly. That was painful but survivable when only humans were involved.

The moment you want agentic workflows to take on meaningful work, that trade-off breaks.

An agent cannot learn your culture in the way a human can. It cannot sense office politics and adjust. It cannot interpret fuzzy phrases like “we normally keep those in the sales folder unless it is a partner deal”.

To make agentics useful, you have to flip the logic.

  • Start from what an agent needs: one place to look, one rule to follow, one format to rely on.
  • Design your documentation and processes around that.
  • Then let humans flex on top of a clear, reliable base, not instead of it.

We are seeing the first generation of tools and patterns that support this shift.

  • Knowledge platforms that explicitly tackle information sprawl and enforce a single source of truth.
  • Collaboration tools that use AI summarisation, auto tagging and intelligent search to turn scattered pages into structured, actionable knowledge that both people and agents can work with.
  • Enterprise knowledge graph approaches that connect data, documents and events into a coherent, queryable network for agentic AI.

These are not silver bullets. They are signals of the direction of travel. The strategic move is not “buy more AI features”. It is “standardise and structure everything that matters so humans and agents are reading from the same script”.

What Leaders Must Fix Before 2026

If you sit on a leadership team and you want agentic systems to deliver real value in 2025 and 2026, treat documentation and process structure as infrastructure.

Here are practical moves you can sponsor now.

1. Name A Single Source Of Truth For Each Domain

For every core domain, decide what the system of record is.

  • Customer contracts.
  • Product specs.
  • Policies and procedures.
  • Runbooks and playbooks.
  • Training and onboarding assets.

One system for each. One location inside that system. No exceptions.

Then, do the hard cultural work. Turn “helpful duplicates” into anti patterns. Make it easy and fast to update the source of truth. Make it painful to create competing versions.

2. Standardise Folder Structures, Spaces And Page Types

Human friendly does not mean anything goes. It means predictable.

  • Agree a small number of folder or space templates that cover most use cases.
  • Fix a naming pattern for documents that includes date, owner and version.
  • Use the same page types and sections for similar content such as policies, playbooks and post mortems.
  • Retire legacy structures on a schedule, rather than letting them linger forever.

This is what allows both humans and agents to know where to look, without guesswork.

3. Write For Agents As Well As People

When you document a process, assume that a human and an agent will both read it.

  • Use clear, step based language.
  • Avoid “see above” or “you know the drill” shortcuts.
  • Include explicit owners, triggers and expected outputs.
  • Keep status and dates current, so an agent can trust what it reads.

Peer reviewed research on automation performance shows that ambiguity in process descriptions is where most failures and errors emerge. Clarity is not bureaucracy. It is the cost of useful automation.

4. Use Modern Knowledge Tools To Clean The Mess

Agents need structure, but you do not have to build all of it by hand.

There are now tools that help:

  • Summarise long pages into concise briefs for humans and agents.
  • Auto tag content so it is easier to find and link.
  • Apply intelligent search across multiple spaces so you can see where duplicates or gaps exist.
  • Build enterprise knowledge graphs that connect documents, systems and events into a single, navigable model.

Think of these as power tools for your documentation clean up. They do not replace the leadership decisions about “one rule, one place”, but they make it practical to execute at scale.

5. Tighten Onboarding Around The New Rules

Use new joiners as your reality check.

  • Design a simple onboarding path that assumes zero tribal knowledge.
  • Walk them through the real system of record for each domain.
  • Ask them to find specific documents or answers and note where they get stuck.
  • Treat every “I could not find it” as a signal that something is not standardised enough.

Your goal is simple. A new starter should not need to ask “where the hell is everything”. If they do, your AI agents will be asking the same question silently.

What This Unlocks For Humans

This might sound strict. It is.

You are taking away some personal freedom in how people store and structure their work. You are asking them to give up private folder hacks and idiosyncratic file names.

In exchange, you are offering something more valuable.

  • Less time wasted hunting for information.
  • Faster onboarding for every new colleague and every new agent.
  • Higher confidence that the document in front of you is the right one.
  • More capacity for the work humans are uniquely good at such as judgment, creativity, negotiation and messy collaboration.

One knowledge expert summed it up neatly:

"For agentic AI to deliver real business value, enterprise knowledge must be structured, current, and accessible—unlike typical document sprawl which blocks effective automation."

That is the trade you are making. Give up the illusion of freedom inside chaos. Gain real freedom by letting agents take on the routine, rule bound work they excel at, on top of a clean, structured base.

The Next 24 Hours And The Next Few Weeks

If this resonates, do not turn it into a twelve month programme that dies in committee.

In the next 24 hours:

  • Pick one critical domain, for example customer support, sales operations or HR policies.
  • Name the single source of truth for that domain.
  • Communicate that decision clearly to the teams involved.

In the next few weeks:

  • Map all existing documents in that domain into the chosen system.
  • Apply a simple, consistent structure and naming convention.
  • Clean up duplicates and archive old versions.
  • Rewrite one or two key processes in a way that a human and an agent can both follow.
  • Run a small agentic workflow against this “clean zone” and see how much smoother it runs.

Once you can see the difference in one area, expanding across the organisation stops feeling theoretical. It becomes operational common sense.

You do not have to fix everything before 2026. You do need to start moving from human friendly chaos to agent first clarity now.

Links:

Quotes:

  • "A single source of truth is essential. Without it, employees waste countless hours searching across multiple platforms and repositories, leading to lost productivity and onboarding frustration.", https://www.glean.com/perspectives/navigating-information-sprawl, Trust rating: high, Reason: Concisely explains the cost of information sprawl and the importance of a single source of truth for productivity and onboarding, Date written: 2024-09-30
  • "While humans frequently adapt, ignore, or reinterpret rules under pressure, automated systems follow rigid logic, leading to consistent but inflexible outcomes.", https://www.sciencedirect.com/science/article/pii/S2214804325000448, Trust rating: high, Reason: Provides scientific backing for the claim that agents cannot improvise like humans and therefore need clear, stable rules, Date written: 2025-07-21
  • "For agentic AI to deliver real business value, enterprise knowledge must be structured, current, and accessible—unlike typical document sprawl which blocks effective automation.", https://www.superblocks.com/blog/enterprise-knowledge-graph, Trust rating: medium, Reason: Directly links structured enterprise knowledge to successful agentic AI and highlights the blocking effect of document sprawl, Date written: 2025-10-13

Share this post

Keep reading