Researchers from Microsoft Research and Salesforce tested the leading AI models available in two settings: receiving a complete request all at once, and receiving the same request in stages, across several messages — the way a real work conversation actually unfolds. The study, published in May 2025, analyzed more than 200,000 simulated conversations across six task types.
The result: in multi-step conversations, accuracy dropped by an average of 39% compared with the same request made in a single shot. And the effect showed up in the most advanced models on the market almost as strongly as in the simplest ones — this is not a limitation you fix by switching models or paying more.
For anyone deciding how AI gets used across a company, that number changes the question worth asking. It is not "which model is smarter" — it is: what does the tool do once a conversation grows long and full of corrections? Does it restart the right way, or does it stubbornly push forward in the wrong direction until the end?
The mistake that sticks to the conversation
The researchers split the performance drop into two parts. A small loss of raw capability — the model "knows" almost as much after a long conversation as it did at the start. And a much larger jump in inconsistency: the same request, rephrased in slightly different ways as the conversation goes on, starts producing increasingly divergent answers to itself.
The explanation the study gives is direct. The model tends to make assumptions in its very first responses and tries to hand over a final answer too early. When someone corrects a detail midway through, the model carries the wrong assumption forward instead of rebuilding its reasoning from scratch — it stacks the correction on top of the earlier mistake, without erasing the mistake.
When the model takes a wrong turn in the conversation, it gets lost — and it does not recover.
That is the opposite of what intuition suggests. The common expectation is that a longer conversation, with more back-and-forth, leaves the AI "more tuned" to what the person wants. The study shows the opposite: every additional round after a misunderstanding tends to reinforce the misunderstanding, not fix it.
This explains a scene anyone who uses AI daily has lived through: the conversation starts great, but after a few exchanges it starts repeating an answer already discarded, ignoring an instruction given three messages ago, or sticking to a format already corrected twice. That is not the tool "getting tired" — it is the exact effect the study measured, happening in real time.
Restart the conversation, do not push through it
The most common fix is the wrong one: keep explaining, correcting, and reinforcing inside the same conversation, hoping one more message will finally solve it. The study shows that attempt tends to make things worse, because the model is not simply "forgetting" one isolated fact — it is carrying more and more accumulated text, with decisions, corrections, and noise all mixed together.
Official AI-maker documentation describes the mechanism differently, but arrives at the same conclusion. Every AI has a limit on how much text it can process at once in a conversation — the so-called context window, meaning how much it can "read" before the start of the conversation falls out of its reach. Anthropic's own documentation (maker of the Claude model) explains that as the volume of accumulated text in a conversation grows, the AI's accuracy at correctly recalling what was said drops — an effect the company's own engineers named context rot.
The fix these same makers document for long tasks is not "explain again, in more detail." It is to summarize what has already been decided and restart with a clean conversation, carrying only that summary forward — not the entire back-and-forth, word for word. In other words: the problem is not solved by pushing on, it is solved by restarting in an organized way.
That changes the criteria for choosing an AI tool for everyday company use. A good tool is not just one that "converses well" — it is one that makes this restart easy: one that makes clear when it is time to summarize and move on, and that stores what was already decided somewhere that survives past that one specific conversation.
What has to be in place
A workplace ready for long AI conversations rests on a few concrete mechanisms, not on the promise of a "smarter model."
A record of what was decided, not just the raw conversation history. What is worth reusing later is not the fifty messages exchanged — it is the conclusion. A corporate tool needs to make that conclusion easy to find and reuse, separate from the back-and-forth that led to it.
A simple way to restart without losing the thread. When a conversation starts repeating itself or contradicting what was already said, the fix is to summarize what was agreed and open a fresh starting point with that summary — not to keep pushing in the same thread until it becomes unmanageable.
Reuse of what already worked. When someone finally nails the right request for a recurring task, that should become something reusable by any authorized person on the team — not something lost inside one person's conversation from half an hour ago.
An AI model sized to the task. A simple, repetitive question does not need the most expensive, most "reasoning-heavy" model available; a task that requires longer analysis benefits from it. Routing each task to the right model prevents the habit of pushing everything into one giant conversation, trying to "solve it all at once."
This is how Skyller was designed: requests and conversations that worked well can be saved as a reusable template for the team, and agents work through coordinated steps per task, instead of piling everything into one endless exchange of messages.
What the team gains when the request becomes a template
The most expensive cost of a conversation that "loses its thread" is rarely the wrong answer itself — it is the time spent until someone notices the answer is wrong, plus the time that same person (or someone else, months later) will spend again building the same request from scratch.
When the history of a long conversation stays locked inside one person's account or app, every colleague restarts that learning curve alone. The analyst who figured out, through trial and error, the exact way to ask for a bank reconciliation report is the only one who knows how to ask it that way — and if they leave the company, that knowledge leaves with them.
When that same successful request becomes a reusable template inside a corporate environment, the gain shows up in two places at once. First, in time: no one has to redo the work of "teaching" the AI to understand the task. Second, in quality: because the request template was already tested and refined, every new person who uses it inherits the version that already works, not the first noisy attempt.
That is the effect that separates a company that uses AI from a company that has good people using AI on their own. In the first case, the learning stays with the team. In the second, it stays with each person, and resets with every bit of turnover.
A starting checklist
Before the next long AI conversation, consider this simple checklist with the team:
- Set a practical message limit before summarizing. When a conversation passes ten or twelve exchanges without reaching a result, stop and write a short summary of what has already been decided — instead of sending one more message trying to fix the last one.
- Carry the summary into a new conversation, not to the end of the old one. The goal is to start with a clean context window, carrying only the conclusion — not the entire discussion that led to it.
- Save the request that worked. If a specific way of asking for something worked well for a recurring task, save that text as a template for next time, instead of rebuilding it from memory.
- Ask IT where these reusable templates live. If the company already uses a corporate AI tool, there is likely a place built exactly for this — and it is worth finding before recreating the wheel.
