Your Contract AI Is Only As Good As Your Contract Playbook

When one of our customers first switched on AI contract review inside Hevi, they did what everyone does.

They dropped in a big live contract, hit run, and waited for magic.

The result was… fine. Not terrible. Not amazing. The analysis was sensible, but it did not feel like having their best commercial lead sitting beside them. It felt more like a smart new grad who had read the contract but did not yet understand how this business really negotiates.

A few weeks later, the same team told us something that we wish every customer could see on day one.

As they put it, the more they refined their contracting guidance, the better the AI became. Their words were simple. The better our prompts are, the better the analysis will be. Over a handful of tenders, the quality of the output kept improving, not because we secretly changed models overnight, but because they sharpened how they told the AI what good looks like.

This is the part of AI adoption no one talks about.

The myth of the all knowing contract AI

The quiet assumption in many teams is simple.

If the AI is any good, it should just know what a good contract looks like for us.

We see a version of this in almost every rollout.

A team uploads a contract, presses run once, and if the first review is not perfect, the verdict is that the AI is not ready yet. In reality, nothing in your business works that way. You would not hire a new contracts engineer, throw them a 120 page EPCM, give them no playbook, and then fire them when they miss a few of your internal preferences.

Yet that is exactly how many teams treat AI.

What contracting guidance actually is

When we talk about prompts in Hevi, we are not talking about clever one liners for a chat window.

We are talking about your contracting guidance. The rules, preferences and red lines that live in three places right now.

1. In the heads of your senior commercial people
2. In old mark ups and departure tables that no one can find at 5 pm
3. In scattered Word documents and share drives that are out of date

When our customer first set up their guidance for AI review, they did what most of us would do. They tried to cram everything they could think of into one complex block of text.

It was technically correct, but not sharp. It made the AI try to do too much at once. Over time, they started to do something very simple and very powerful. They pruned.

They simplified the guidance. They pulled out the truly non negotiable items. They grouped similar risks together. They used clearer, more direct language. The result was not only better AI analysis, but clearer thinking for the team itself.

The moment the penny drops

After running a handful of contracts, the team noticed a pattern.

The more they ran, and the more they refined their prompts, the better the analysis became. Clauses that were originally missed started to be flagged every time. Nuances around liability caps, liquidated damages and notice periods were handled more consistently. The AI started sounding less like a general purpose lawyer and more like their own commercial team.

Nothing mystical happened inside the model.

What changed was the quality of the instructions they were giving it.

They moved from generic language like flag any onerous clauses to specific guidance such as highlight any clauses that cap liability below our standard, remove consequential loss exclusions in these scenarios, and call out any notice periods shorter than ten business days.

They stopped asking the AI to guess and started telling it exactly what matters in their world.

Turning your tribal knowledge into an AI ready playbook

If you are thinking about rolling out AI contract review, or you already have and felt underwhelmed by the first run, here is the practical lesson from this customer.

Your first job is not to perfect the model.

Your first job is to get your own playbook out of people’s heads and into the system.

A simple way to start.

1. Pick your last three big tenders or contracts
2. Ask your most experienced commercial person what actually mattered in those deals
3. Turn those answers into clear, concrete preferences and questions
4. Load that into your Hevi contracting guidance and run a live contract against it
5. After each review, adjust the guidance instead of abandoning the AI

You will quickly see the same pattern. Each iteration makes the AI a little more like your best team member on their best day.

Why this matters for heavy industry teams

In mining, manufacturing and construction, your contracts are not generic.

They reflect site realities, local labour markets, safety obligations, production targets and long term relationships with owners and suppliers. Off the shelf AI cannot know that.

It can read the words, but it cannot know which risks are politically sensitive on your board, which caps your insurer has already agreed to, or which carve outs you know the other side will concede if you push.

Only your playbook knows that.

The teams who get the most from Hevi are not the ones who chase every new AI feature. They are the ones who are willing to sit down for an hour, write down how they actually want to do business, and then let the system learn with them over time.

The uncomfortable truth and the opportunity

The uncomfortable truth is that most organisations do not really have a single, up to date contracting playbook today. They have fragments.

AI makes that visible very quickly.

The opportunity is that once you do the work to sharpen your guidance for AI, you also sharpen your entire contracting function. You reduce variance between projects. You onboard new staff faster. You spot drift between regions or business units. You stop rewriting the same mark ups every quarter.

And, like our customer, you stop asking whether the AI is good enough, and start asking a better question.

How do we make sure our guidance is sharp enough for the AI to show us what it can really do.

‍