Move slow and test things: the key to trusted AI in business

Artificial intelligence is the last digital transformation of our lifetimes, but the gap between AI in experiments and AI in production feels like a widening chasm.

It’s worth saying that the high failure rate identified in the S&P data isn’t stopping companies from experimenting. Most are still investing in generative AI, but the stumbling block seems to be firmly wedged between proof-of-concept phase and actual production.

In my experience, many experiments usually get as far as concluding “this kind of works, 85% of the time”. Business users are very interested in the promise of AI but is it any surprise that many are reluctant to take the next step and commit to more extensive projects when the risk of error or hallucination is still high?

Rapid evolution = risk

I have a theory about why this is happening: the technology is evolving so fast that it can feel like we’re in constant catch-up mode. It’s one thing to successfully launch something that works in experiment mode. It’s another to put it to work in the business – and to trust it with your customers’ and your own data.

Because fundamentally, that’s what it comes down to: a question of trust.

And for that, you still need people. With AI, you can insert automations within a process to speed it up, but you soon reach a point where you need to verify that it’s producing the right output. So how do we test that?

Eliminating the humans?

As it happens, Phased AI and Each&Other are working together on exactly this challenge: we’ve been finding the human pain points our customers have, and figuring out where to automate them, effectively and accurately.

What we’ve found is that most of the work is not in building a prompt or an automation to start the process: it’s in evaluating it robustly at the far side.

We work with our customers to ensure they can go to a really trustworthy application at the end of the experimentation phase, that works with human workflows. Together with those businesses, we sketch out the best ideas for how they can use AI with high impact but – and this is the crucial part – low risk.

Typically, some customers will have played around with ChatGPT and think: “could it do this task I have in mind?” “Can I make it work with the API and make it safe?” Some have a specific use case, and they think generative AI can solve this, or can make it faster, but don’t know how.

Workshopping trustworthy AI

In a workshop, we establish if this is a low or high risk use case, and we map out their current human process: is it a spreadsheet or is it another application? Then we identify the hypothesis: can AI do this, and typically, we build a very quick prototype for them, taking in as near-to-real data as we can.

Next we test the AI output at the other end to see if it’s accurate. In parallel, Each&Other tests how this will work from a UX point of view, evaluating how the user will experience this.

After the hypothesis, and the rapid proof of concept, we then build out a bigger data set, scale the test, build out a quick UI, and the users check for flaws. That can be a simple checkbox exercise for the test cases: did the expected output match the real output yes or no. Or you can give it a score on a sliding scale.

This part is so important: even at the proof of concept stage, we’re always testing if the AI is consistently giving the output that we think it should. And at the end of the proof of concept, when we hand over the tool to the customer to try in real case scenarios, it comes with a report from our test cases as to what we think.

After that, the next stage is how do we make this production level, discovering the allowable level of accuracy that can be with customers, and we establish what guardrails need to be in place if this is interacting with customers.

For years, the tech industry motto was “move fast and break things”. Right now, I believe we’re at the very beginning of the transformation that AI can bring and companies understandably want to be a part of it. The key to getting it right is to move slowly, intentionally, and to make sure that you are working to properly evaluate at the other end.

AI has the potential to make us all quicker and more efficient, but the current state of the art is not replacing humans. None of this can be just handed over to a machine: there’s a job to understand where the human needs to step in.