Large language models are the most general-purpose form of AI available today. They can read, write, reason, summarise, and generate across almost any domain. That generality is their strength. It is also the reason most businesses are still only scratching the surface of what they can do with them.
Right now, the typical adoption pattern looks like this: people use LLMs for tasks. Drafting an email. Summarising a document. Analysing a dataset. Brainstorming approaches to a problem. These are useful. Every business should be doing this today. But they are tasks, not processes, and they are certainly not full deliverables.
Tasks, processes, deliverables
It helps to think about AI capability on a spectrum. At one end, you have the task: a single, bounded unit of work. Draft this paragraph. Extract the key figures from this PDF. Suggest three approaches to this problem. LLMs are already good at this. Most of the productivity gains people talk about today live here.
In the middle, you have the process: a sequence of tasks chained together to achieve a broader outcome. Reviewing a contract involves reading the document, flagging risk clauses, comparing against standard terms, and producing a summary of issues. Each step is a task. Strung together with the right logic, they become a process.
At the far end, you have the deliverable: the complete output that a business actually ships to a client or uses to make a decision. A full tender response. A geological site report. A fund strategy recommendation. A deliverable is typically the result of multiple processes, informed by company-specific knowledge, assembled in a specific order that the business has refined over years.
The interesting question is not whether LLMs can handle tasks. They can. The question is what it takes to move them from tasks to processes, and from processes to full deliverables.
The agentic era gets you part of the way
The current wave of agentic AI is explicitly trying to bridge this gap. Agents can use tools, chain steps together, make decisions, and execute multi-step workflows. This is genuine progress. It moves AI from "do this one thing" to "handle this sequence of things."
But look at where agents are heading first. They are being optimised for the processes and deliverables that are common across all businesses. Your VAT return. Standard legal review. Bookkeeping. Customer support workflows. Compliance checks. These are high-volume, well-defined, and the feedback loops needed to train agents on them exist at scale.
This makes sense from the perspective of the labs building these systems. Reinforcement learning requires environments where you can measure success, run thousands of iterations, and refine the model's behaviour. If a task is performed by millions of businesses in roughly the same way, it is a good candidate for that kind of optimisation. The labs will get there, and when they do, these generic processes will largely be handled.
The long tail is yours
But most of the work that makes a business distinctive does not look like a VAT return. It sits in a long tail of specialised workflows that no frontier lab is going to build RL loops for, because the feedback environments simply do not exist at scale.
Think about what makes your company valuable. It is probably not the ability to summarise a document or draft an email. It is the specific way you move from raw inputs to finished output. The sequence of decisions, the order of operations, the particular way you structure a tender response or interpret site data or assemble a fund recommendation. You have refined that over years of competitive pressure. It is proprietary. It is the reason you exist and not someone else.
An off-the-shelf agent, no matter how capable, does not have access to that. And the labs are not going to train it in. Your niche is too small, too specific, and too dependent on context they do not have.
Two things the model does not ship with
To get LLMs from generic task-handling to producing real deliverables for your business, you need to provide two things.
The first is what I call the trajectory. This is the choreography of your deliverable: which tasks need to happen, in what order, with what decision logic between them. It is the blueprint of your process, made legible to a machine. If you are a design-build firm responding to a tender, the trajectory is: pull the relevant project history, map requirements against your capabilities, draft section responses in your house style, run a compliance check, assemble the final document. Each step might use the general ability of an LLM, but the sequence and the logic connecting them is yours. It comes from your experience of what wins, what gets rejected, and what the client actually cares about.
The second is the context. This is the company-specific information the model needs at each step to do the work properly. Your past project data. Your preferred formats. Your pricing logic. Your interpretation of regulatory requirements as they apply to your sector. None of this is baked into the model. You have to install it, whether that is through retrieval-augmented generation, fine-tuning, knowledge graphs, or simply structured context injection. The method matters less than the principle: you are simulating the tacit knowledge that a long-tenured employee carries around in their head.
Think of it this way. A capable LLM is like hiring the smartest generalist you have ever met. They can reason, write, and learn fast. But on day one, they do not know your processes and they do not have your data. You would not drop them into the deep end and expect a perfect deliverable. You would give them a blueprint of how you work and access to the information they need. The same applies to AI.
This is IP, not just adoption
Here is where it gets interesting from a strategic perspective. When you encode your trajectory and context into a system that an AI can execute against, you are not just "adopting AI." You are creating a new form of intellectual property: machine-legible process knowledge.
This matters because it is durable. The underlying LLM will change. The models will get better, cheaper, and more capable. But the encoded understanding of how your business specifically moves from input to output retains its value regardless of which model sits underneath. Companies that invest in this now are building an asset, not just buying a tool.
Stay open to the loop
One caveat. The point of encoding your process is not to fossilise it. The best implementations treat this as a feedback loop. You encode your best current understanding of the optimal trajectory. But a capable model, precisely because it reasons generally, may surface inefficiencies or alternative paths that you could not see from inside the workflow. The scaffold should be firm enough to guide the model but flexible enough to learn from it.
Your processes have survived competitive selection. They encode real value. But they are also the product of human habits and historical constraints that may no longer apply. The companies that get the most from AI will be the ones that encode their knowledge and then let the system challenge it.