When I speak to people about their approach to AI deployment in their businesses, there are a few hesitations that everyone has. Here are the common ones:
- "Doesn't it always hallucinate?"
- "How can I be sure that it's accurate?"
- "I don't know how to deploy it."
- "Won't it take our jobs?"
The future of AI deployment is not the chat interface as we know it. It is built-out automation and systems that use proprietary data to do the jobs AI is much better at than us: research, ingesting data from many sources, and data analysis and processing. But if you're giving it the keys to decision-making, how can you be sure it makes the right call? The answer is asynchronous AI.
Before we get to that, let's understand the issue.
AI hallucinations
AI hallucinations, when large language models confidently provide false or fabricated information, have been a long-standing and well-documented issue. They range from subtle factual inaccuracies to completely invented events, citations or reasoning. While they're often written with remarkable fluency and authority, their unreliability creates real limitations when AI is expected to perform tasks requiring consistency, reliability and trust.
The issue becomes even more complicated when you're no longer operating within the friendly confines of a chat interface. While humans can detect and course-correct a hallucination mid-conversation in the UI, the story is different when you're using the AI through an API, where inputs and outputs are meant to be automated, invisible and assumed to work.
A test of consistency, output and accuracy
I ran a simple but revealing test that evaluated outputs across three metrics: consistency, output and accuracy. We prompted GPT-3.5 Turbo 100 times through the API (application programming interface) with the same input and observed how stable the outputs were.
These were the input prompts:
- Test 1, basic "Acting as an SEO Expert & copywriter, look at this page https://www.theguardian.com/football/2025/jul/15/bigger-better-more-often-infantino-wont-let-up-on-his-ambition-for-club-world-cup and summarise in a short sentence what it is about"
- Test 2, mid "Acting as an expert in SEO, provide recommendations for optimising this page (the same Guardian article) for more evergreen search volume. The output should be a list of 5 keywords with reasons for each of them"
- Test 3, advanced "Acting as an SEO expert and sentiment analysis. Look at this page: (the same Guardian article) provide me with a detailed list of 3 key changes you'd do to make the page target more search behaviour and keyword opportunity and then analysis the sentiment of the page to summarise 3 key points from the sentiment analysis"
Each one escalates in complexity, and each is the sort of question you'd ask a junior in your team to complete. To grade the results we used this framework:
- Consistency (structural): did the AI produce different styles, tones and formats across runs?
- Output: were the results in the requested format across the requests?
- Accuracy: sometimes facts were present and correct. Other times they were fictional.
Here were the results:
On a one-off basis through the UI, this variation is manageable. You notice it. You click regenerate. You fix it. No problem.
But through the API, where automation depends on deterministic and predictable behaviour, this is a breaking issue. In production, where human oversight isn't practical for every request, hallucination and inconsistency aren't just annoying, they're dangerous.
The importance of thresholds
This leads us to thresholds: the invisible standards that decide whether an AI response is good enough to be used. Think of a threshold as the AI's quality gate. How well does the output need to align with factuality, task specificity or user tone before it's deemed acceptable?
Take a playful but telling example. If you ask an AI to "tell me a story about a mop", you might get three different levels of threshold:
- Low threshold "Once there was a mop. It cleaned floors. The end." Technically accurate. Completely uninteresting. Functionally useless.
- Mid threshold "The mop had dreams of being a dancer, twirling across the linoleum like Fred Astaire. But it was stuck in a janitor's closet… until one night…" Creative and engaging. A solid answer.
- High threshold "In 1973, amidst the oil crisis, a factory in Detroit built a mop with an experimental polymer head that would later be considered revolutionary. This is the story of how that mop ended up in the Smithsonian…" Original, researched, deeply structured. Too ambitious, but better.
Setting and maintaining the right threshold is critical. To do that reliably, you need more than one-shot AI output. You need an architecture that can evaluate, refine and structure, on its own.
Asynchronous AI: a team model for machines
Now imagine AI not as a single black box that returns a string of text, but as a distributed asynchronous system, like a team of people with specialised roles:
- The Briefer Interprets the prompt and defines the goals.
- The Executor Does the work: writing, coding, summarising and so on.
- The Reviewer Checks the output for quality, accuracy and tone.
- The Outputter Packages the final result in the desired format.
This is asynchronous AI, where each role can be played by separate instances or phases of the model, running in sequence or in parallel, evaluating and improving each other's output.
It mimics the way high-performance teams work: distributing complexity, enabling specialisation, and introducing checks and balances, with different thresholds and standards to make sure the desired deliverables are met.
And just like a human team, this system doesn't assume perfection in the first draft. It builds in refinement as a feature, not a patch.
Why this matters: context windows and limitations
In a conversational UI, a lot of this happens invisibly. Context is preserved in your chat history. The model remembers your earlier preferences. It self-corrects, adds nuance and even "learns" over the session, within limits. But that context, the glue holding everything together, doesn't exist in the same way through the API.
When using the API, context windows become a hard constraint. Everything the model needs to understand has to be included in the payload: your prior prompts, any preferences, the response history, all of it. If you don't manage this carefully, the model responds like it has no memory, because it doesn't.
This is where asynchronous, multi-agent, team-like AI becomes not just helpful but necessary. It lets you simulate long-term memory, enforce standards, manage context and run multi-step reasoning, without assuming the model will just get it from a single shot, and with several attempts at an accurate version.
Final thoughts: the path forward
The future of AI isn't about pushing harder on single-shot prompt engineering. It's about orchestrating AI like a team, thinking in systems, and designing workflows where several agents with defined roles collaborate, asynchronously, to meet quality thresholds, manage context and produce reliable outputs.
Hallucinations, inconsistency and brittle context limits aren't just bugs. They're signs that we're still thinking too linearly. The solution isn't just better models. It's better architecture.
Asynchronous AI is that architecture. And it's how we'll go from clever answers to trustworthy systems.
Jeremy McDonald, founder of We Are Pioneers
