Skip to main content
Help Center

Why teamspace relies on rules for AI

Invoices that are 99.9% correct are not correct invoices. Why in teamspace rules do the calculating and AI at most reviews – and what that means for you.

You could build an ERP today in which an AI writes the invoices. It would work impressively well. We decided against it – and this article explains why, because the reasoning also explains why the AI features in teamspace look the way they do.

The invoice example

Imagine two systems that generate an invoice from project times, contracts and price lists.

One is rule-based. Someone has set up the rules: How is rounding handled? What is included in the flat price? When does the framework-contract discount apply? These rules can be simple or arbitrarily complex – but they are written down, traceable and run the same way every time. The result is 100% correct. Not “mostly”, not “very nearly always”. Correct.

The other lets an AI do the calculating. It has seen thousands of invoices and does its job well. Let’s say: 99.9% correct.

99.9% sounds excellent. For an invoice it is a disaster.

With a thousand invoices a month, one is wrong. Nobody knows which. It looks like the others – neatly formatted, plausible amounts, correct layout. Maybe the customer notices, maybe the tax adviser, maybe the tax auditor in three years’ time. And even if it is noticed, nobody can explain why this particular one was wrong, because there is no rule to look into. There was only a probability, and this time it came out wrong.

The difference is not the error rate. It is explainability. A rule-based system that calculates wrongly has a wrong rule – you find it, correct it, done, and from then on it is correct for all time. A system that is statistically off has nothing you could correct.

Our approach: the AI reviews the rules, the system stays rule-compliant

From this follows the division of labour that runs through all of teamspace:

Rules do the calculating. The AI may review.

It is not the AI that writes the invoice – but it may take a look and say: “This project has been billing without the framework-contract discount for three months, even though the customer has one. Are you sure?”

That is exactly the right role. It plays out where AI is strong – spotting anomalies in large volumes, seeing patterns, asking questions. And it keeps AI out of where it is weak: where the result has to be exact, repeatable and justifiable.

The result stays rule-compliant. The rule decides. The AI only noticed earlier that the rule might be wrong – and a human decides whether it is right.

What that means in practice

You’ll recognise this attitude everywhere once you start to notice it:

  • Summarising yes, deciding no. Summarising a ticket is harmless – you read the summary with the history right beside it. Closing a ticket on its own authority would be something else.
  • Flagging yes, posting no. Showing anomalies in a report is helpful. Turning an anomaly into a correction posting on its own would not be.
  • Suggesting yes, fixing no. Where an AI feature produces something, you can see it, change it and discard it before it lands anywhere.
  • You can always see that it was the AI. An AI summary is recognisable as such and sits next to the original – not in place of it.

In short: you won’t find a feature in teamspace that takes off your hands a decision you have to answer for yourself.

”Isn’t that too cautious?”

A fair question, and the answer is: it’s not about caution, it’s about responsibility.

There are plenty of tasks in a company where 99.9% is great. If a summary misses a nuance, you read two sentences back in the original – no harm done. That is exactly where we use AI, and gladly and increasingly so.

But invoices, billing, figures shared with third parties: these are matters someone is accountable for. There, “almost always correct” is not a category. A tool that puts its user in a position they can no longer keep track of is not a good tool – however impressive it may be.

And in all honesty: a rule-based system can handle the rules your company really has. Framework contracts, special agreements, that one arrangement from 2019 that only sales still remembers. An AI knows the statistical average of all companies – it doesn’t know your exceptions.

Where the AI is strong nonetheless

So as not to give the wrong impression – teamspace uses AI where it is better than any rule:

  • Condensing. A ticket with forty entries down to five sentences. No rule can do that.
  • Finding anomalies. Spotting the thing in an Invoicing analysis that stands out – without anyone having defined in advance what “standing out” means.
  • Explaining connections. Why this month’s figures look different, in full sentences instead of a column of numbers.
  • Answering questions nobody anticipated. For that there is also MCP.

These are all tasks where a human can immediately place the result – and where a mistake stands out instead of hiding. That is exactly the criterion.