Five levels of AI, from answering questions to making decisions.

Quotes run as one process your business owns, and you can check every step.

Instead of each person’s assistant doing it their own way, the quote becomes one process, built on the systems you already have: your ERP, your price lists, your inbox. Plain rules handle what rules can, and AI handles the parts that need judgment, like reading a messy request.

Every quote is checked against your rules before it goes out. Is this today’s price? Is the margin above your floor? Is freight included?

What it takes: Most companies’ data is messier than they’d like. Part of the work is fixing the gaps the quote depends on, like item codes that don’t match or prices that live in three different places.

A record you can check: You can trace a quote back to the prices it used, the checks it passed, and the person who approved it. Real examples, including the awkward ones, become tests for future changes.

It earns the right to send routine quotes.

Could the system approve some quotes itself? First, it has to show which ones it can handle and when it should ask Maria. We test ordinary cases, difficult cases, and past mistakes against outcomes your team has agreed are acceptable. These repeatable tests are called evaluations, or evals.

Responsibility, with limits: Only after it meets agreed standards in testing and shadow mode do you authorize it to send a defined class of quotes. AI can judge when an unusual request needs Maria, but it cannot waive your hard limits on price or margin. We test that decision to ask for help, as well as the quote it produces. Saying it is confident is not enough.

See what is happening: You can see what it handled, what it referred, where work stalled, and what happened afterward. We monitor errors and changes in performance; affected work can return to human review.

Calls that used to wait on someone senior move to the system, one tested decision at a time.

At Level 4, the system earned one job: sending routine quotes. At Level 5, earning trust is how your business adds work to the system. Each decision gets its own tests, its own run in shadow mode, and its own monitoring once it’s live. So calls that used to wait on someone more senior, like which substitute is acceptable or when a thinner margin is worth it, can move over as the system proves it can make them.

The people who used to make those calls set the limits instead, and handle what’s genuinely new. Every decision is visible, none depends on one person remembering to check it, and you can take any of it back.

One decision at a time: Doing one job well doesn’t earn the next. Each decision is tested against past calls that the person who owns it agrees were right, including when to hand off. Permission comes from those results, not from how sure the system sounds, and it covers that decision only.

When better AI arrives: Because every decision already has its tests and is watched as it runs, a better model doesn’t mean starting over. It runs against the same tests before it replaces what’s running, and if results slip after a change, that decision can go back to a person while it’s fixed. You get the gains as AI improves, without adding a part of the business nobody watches.