Safer rollout
Building great tech is only half the battle; getting it into users’ hands successfully is the other half. This is why applied AI is far harder than publishing impressive offline results1. But in the real world, rolling out new AI features comes with pitfalls – outrageous model behavior, system outages, eroding painstakingly-earned trust in your product/brand.
Not even the promise of major impact or FoMO can overcome the fear of those consequences. The status quo wins. In mid-sized companies, it’s often the fear of bad PR that holds teams back (small companies can shrug off mistakes; very large ones are PR battle-hardened). In the biggest corporation, individual careerism can be the blocker.
Encouraging boldness, blameless postmortems, and celebrating risk-takers are all great steps to take. We can also be pragmatic by lowering the risks of deploying a new future:
Gradual roll-out: Start with internal dogfooding, then beta testing or live experiments with a small user segment. Catch issues before they escalate.
Communicate value-added: Don’t just announce a change. Tell a story. Explain why you’re building this feature, and show users the benefit. If you can, personalize the impact: “this feature reduces your workload by 28%”.
Easy reversibility: Offer opt-outs (or opt-ins). Make it possible/easy for users to reverse the update: “Would you like to disable this?” Coupled with a reminder of the value-added.
Keep improving the product, but without burning trust along the way.
To give credit where it’s due, publishing in a peer-reviewed journal is no cakewalk – and often a humbling experience.