One of the most meaningful changes in AI is not visible in benchmarks or demonstrations. It shows up in how little people now need to explain for a system to be useful.
The usual explanation is better models. That is only part of it. What is really changing is how AI systems engage with intent: how they tolerate ambiguity, ask better follow-up questions and stay with the flow of work instead of interrupting it.
Prompting was a workaround
The rise of prompting revealed a contradiction. Systems built on natural language still required people to turn half-formed thoughts into precise, over-structured instructions. People learned to coax systems into understanding them, often by trial and error. Each exchange became a small detour away from the work itself.
People should not have to package all the context. That responsibility belongs to the system. Prompting was never the destination. It was a workaround for systems that could not yet hold up their side of a conversation.
How people build shared understanding
Linguists and psychologists have studied how conversation works for decades. Herbert Clark and Susan Brennan described grounding: the process by which people in a conversation establish, step by step, that they understand each other. Speakers do not deliver complete information at once. They offer something, check it landed, and repair misunderstandings as they go.
The philosopher Paul Grice described the cooperative principle behind it, in four maxims. Say as much as is needed, and no more. Say what you believe to be true. Be relevant. Be clear. A good human colleague follows these without thinking. A well-designed AI conversation should too.
What a good conversation layer does
In real organisations, context does not arrive neatly. It emerges through partial requests, loosely worded instructions, unspoken constraints and past decisions remembered by people rather than systems. People navigate this easily. Software traditionally does not.
A well-designed conversation layer:
Accepts uneven input: a voice note, a forwarded email, a two-word message, a photograph.
Asks what matters next, not everything at once: one or two questions that unlock the most, following the principle of progressive disclosure.
Adapts to the person: a senior engineer gets technical shorthand; a first-time buyer gets plain explanation.
Draws on what is already known: customer history, past orders and earlier decisions, so people are not asked twice.
Confirms understanding before acting: a short summary of what it understood, especially before anything binding.
Keeps continuity: a conversation picked up next week starts where it left off.
In short, it sources context instead of demanding it.
From conversation to organisational memory
When communication is designed this way, conversations stop being disposable. Every clarification, override, escalation and approval leaves a trace. Over time, those traces add up to a record of how the organisation reasons: which constraints mattered in practice, how ambiguity was resolved, when exceptions were allowed and how precedent shaped outcomes.
That is the difference between knowing what happened and understanding why. A companion piece on decision memory covers how that record is kept and used. Conversation is where it begins.
Why systems of record were never enough
Enterprise systems are good at recording the current state. They show what is true now. They rarely show how a decision came to be. That missing layer has always lived elsewhere: in conversations, judgement calls and experience carried by people.
That is why organisations rely on what engineer Tanya Reilly called glue work: the often invisible effort of people who connect information across tools, chase loose ends, and apply judgement where systems fall short. Glue work keeps organisations running, but it rarely gets recorded or recognised, and it leaves with the people who do it.
When AI is added only at the interface or the analytics layer, it inherits that blindness. It sees outcomes, not reasoning. When AI sits inside the path of communication and execution, it sees context being assembled in real time, and, if designed well, it can keep that understanding for the organisation.
Co-intelligence lives in the middle
The most effective systems are neither fully autonomous nor purely assistive. Ethan Mollick calls this co-intelligence: people and AI working together, each doing what they do best. In practice, people set intent, apply judgement and approve what binds. Systems gather context, handle routine steps, keep the record and bring the right decisions to the right people.
Signs a conversation layer is working
People stop rephrasing requests to be understood.
Questions asked by the system are few, specific and useful.
The same information is not requested twice.
Exceptions arrive with the context already attached.
New colleagues can find how similar situations were handled.
How Mitochondria's agentic AI systems hold a conversation
The approach. Every unit is designed around how people in that workflow communicate: the channels they use, the inputs they send and the questions an experienced colleague would ask. The design draws on the founding team's decade of research and practice in communication.
What it enables. Requests are understood from partial input, clarified with few, well-chosen questions and confirmed before anything binding happens. Each exchange adds to decision memory held in the organisation's own environment.
The impact. Less time spent explaining, fewer loops, and an organisation that keeps the reasoning its people once carried alone.
Read more about Mitochondria's approach.
Sources
Clark, H. H. and Brennan, S. E. (1991). Grounding in communication. In Perspectives on Socially Shared Cognition.
Grice, H. P. (1975). Logic and conversation. In Syntax and Semantics, Vol. 3.
Reilly, T. (2019). Being Glue.
Mollick, E. (2024). Co-Intelligence: Living and Working with AI.
—
Mitochondria builds agentic AI for evidence-led core operations. Research-led and model-agnostic. ISO/IEC 27001:2022 certified, with controls aligned to the GDPR, UK GDPR, India's DPDP Act and Australia's Privacy Act. Amsterdam and Pune. Request a working session →