Models Are Bought. Harnesses Are Built.
Whether AI works reliably inside a company depends largely on the layer around the model, often called the harness. Anyone can buy the model; the harness has to be written bit by bit.
Models are bought. Whatever model one company can buy, its competitors can buy too, and roughly every six months a better, cheaper one comes out. The layer wrapped around the model, often called the harness, is where a company’s own rules live: how customers are tiered, how budgets are split, who has to approve what, how things are cleaned up after a mistake. None of that can be bought, and it’s hard for anyone else to copy.
So this layer is most of what a company can build up for itself. Training its own model or rushing to buy the most expensive platform is worth less than writing its decision processes into the harness, one at a time.
Four words, four ways to divide the work
Chatbot, Copilot, Agent and Harness are often lined up as an evolutionary path, as if each were more advanced than the one before. I tend to see them as four ways of dividing work between AI and people. What separates them is how far the AI gets to decide.
| Role | What AI does | What people do |
|---|---|---|
| Chatbot | Answers questions | Ask, then judge and act themselves |
| Copilot | Makes suggestions | Accept or ignore them |
| Agent | Carries out a task it’s been given | Assign the task, check the result |
| Harness | Keeps making decisions within set boundaries | Set the rules, handle exceptions |
There’s no need to climb these one by one. A product-knowledge Q&A tool is fine as a Chatbot; there’s little point letting it act on its own. A more common problem is labels that don’t match reality: a system that can only answer questions is called an “agent platform,” and a workflow where a person confirms every step is marketed as an “autonomous agent.” The name itself doesn’t matter. The trouble is that management then believes decision rights have been handed over when they haven’t.
The further down the table, the more power the AI holds and the less the model itself can guarantee. A harness was originally the gear strapped onto a horse so its strength could be used safely. By the last row, people no longer watch each action. What they manage is the harness.
What goes into the harness
Managing an AI raises much the same problems as managing a new employee. So every part of a harness has a counterpart in the way a company already manages its people.
| Part | What it governs | What the company already has |
|---|---|---|
| Context | What the AI can see before it decides, such as who the customer is and the company’s official position | Job manuals, handover notes |
| Tools | Which systems it can use, and what it can read and write | System accounts and permissions |
| Workflow | How many steps a task has, and which steps use the model and which use code | SOPs |
| Boundaries | What it can decide alone, what needs approval, what it must never touch | Approval flows, authority limits, compliance red lines |
| Log | The basis and outcome of each decision, and how to reverse it | Business registers, approval records |
| Evaluation | Ongoing checks on how well it’s doing, with permissions adjusted to match | QA, audit, performance reviews |
When new employees struggle, it’s often because nobody explained the background to them. That’s what context is for. Which systems IT opens up for them on day one determines what they’re able to do, and that’s tools. A business without records can’t be audited, and an AI whose decisions aren’t recorded can’t be held to account when something goes wrong. The last three rows together are how an organization gives AI more authority step by step.
Workflow deserves a few more words, especially the question of who calculates the numbers. Both products I’ve built split the work this way: the language model understands the question, plans the steps and explains the result, while amounts, targets and rankings all come from deterministic code rather than being “estimated” by the model, so the results can be recalculated and audited. Business users will forgive AI for imperfect wording. They find it much harder to accept the same metric coming out as two different numbers.
Will stronger models make this layer unnecessary?
A strong objection is that models will keep improving until this layer isn’t needed. That’s partly true. The parts of a harness that exist to make up for a model’s weaknesses, such as prompting tricks, will gradually be absorbed by stronger models.
A company’s own rules won’t be. However capable a model becomes, it won’t know which expenses at a particular company need which level of approval, which customers can’t be tiered the usual way, or how things were handled after the last mistake. That knowledge lives only in the company’s processes and in the heads of long-serving employees. Writing it down, in enough detail that a system can follow it, is work that can’t be outsourced to a model vendor, and it won’t go out of date when the next model ships.
