Why AI Shouldn’t Handle Long Tasks: Confidence Without Context
Have you ever tried to solve a complex issue with a chatbot, only to watch it devolve into nonsense?
After a handful of turns, the model starts losing the plot; forgetting details, erasing nuance, and becoming increasingly inaccurate. Yet, it maintains a deceptive confidence as the workflow quietly collapses.
This isn’t user error. It’s a structural mismatch between how LLMs process information and how complex workflows operate. Standard web interfaces treat LLMs as stateless engines: they have no persistent model of the world, only their training data and whatever temporary context the conversation provides. As the thread grows, this context becomes fragile. The model doesn’t 'forget' in the human sense; it loses signal-to-noise ratio, causing early constraints to fade into the background. The 'memory' features you see in web interfaces are external layers. Fragile text injected into the response’s prompt, rather than integrated into any sort of world model.
A major 2025 study’s title sums it up: “LLMs Get Lost In Multi-Turn Conversation.” Conducted by researchers from Microsoft and Salesforce, the abstract illustrates this succinctly:
“Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover.”
Even this late into 2026, large-language models don’t fare particularly well on jobs that require multiple complex information threads to be held concurrently.
Why Comms Teams Are at Risk
For comms teams, this is particularly noticeable. Our workflows are inherently multi-turn and iterative.
A typical comms task isn’t a single message; it requires collaborative back-and-forth. This is where chatbots are weakest. Tasks like ensuring brand consistency across channels, refining tone based on stakeholder feedback, managing crises, or adapting to breaking news demand consistency that AI struggles to maintain.
Take brand consistency, for example. It requires holding a massive amount of soft, nuanced client feedback alongside hard constraints to maintain clean deliverables. AI may generate an impressive first draft, but it often fails to deliver a consistent voice. This is one of the reasons why AI slop has become such a problem.
The Hidden Cost of "Fixing" It
The amount of machinery to try and force an LLM to be consistent can be immense.
I’m currently running an experimental project that demands strict adherence to a complex, multi-threaded workflow. To keep the model on track across multiple sub-tasks. What did it take?
78 memory entries with hard rules and operating procedures.
50 explicit disciplines to enforce consistency.
15 automated checkers (scripts) that outright rejected known failure modes.
14 named process protocols and 9 "must-read" files to ground the model.
This is the state after just two months. The list is ever-growing, and the system remains more fragile than I’d like. And even with all this engineering, a competent human would likely perform these tasks far better with minimal instruction, simply because they don’t forget what you told them two weeks ago if it mattered.
The Takeaway: Bounded Tasks Only
If this is what it takes for a two-month side project to stay somewhat on track, imagine the impossibility of doing this for a client-facing, high-stakes comms campaign.
The solution isn’t more tooling. It’s understanding that LLMs aren’t good at long-chain tasks, and designing your workflow around it.
Some things to keep in mind:
Never ask an LLM to "handle a project." Outline the specific steps and deliverables you want from a given task. Don’t expect it to understand your intent. Be explicit about exactly what you want it to do.
Because the AI doesn't have all the context, it can act like a fresh pair of eyes. It might notice things that you, as the person who has been staring at the draft for three hours, have stopped seeing.
Accept the Limit: LLMs can be very good at bounded, explicit tasks with well-defined rules. They are poor at long-chain, implicit context tasks.
Make the “rules” explicit: If something matters, don’t rely on the model to infer it from previous conversations. Spell out things like audience, tone, goals, forbidden language, required information, and success criteria.
The Bottom Line
For most multi-step jobs, the tech debt of keeping an LLM on track is greater than the value it provides. Even with an entire workflow dedicated to keeping an LLM on task, a single-dedicated human will almost always outperform them on things that require long-term knowledge and critical thinking.
This doesn’t necessarily mean LLMs don’t have a place in complex workflows. It means their role needs to be scoped to their capability. As we explored in Grounding AI: Why Your Chatbot Needs an Open-Book Exam, AI works best when given clear boundaries, reliable sources, and a human to verify the result. Clear tasks with well-defined deliverables tend to work best. The more complex the workflow, the more room the AI has to confabulate and make mistakes.