Insights/Decision science

What we learned shipping AI into 140 back offices

Priya RaghunathanPrincipal, Decision Science
March 18, 2026 · 12 min read
Copy linkXLinkedIn

On this page

The problem with undocumented callsWhat a decision record actually containsWhere AI genuinely helpsStarting this quarter

Want the decision-record template we use with clients?

Get the template →

The technical work was the short part. In every rollout that stalled, the cause was a person who had not been given a reason to trust the output.

Trust is earned per user, not per org

Pilot approval does not transfer. Each team re-runs the same scepticism, and each team needs to see the tool get a familiar case right before it gets a novel one.

Adoption is a cultural problem wearing a technical costume.

Show the working

Systems that surface their reasoning get adopted faster than more accurate systems that do not. This is not irrational: an explanation is what lets a reviewer catch the error you have not anticipated.

Keep a visible off-switch

Teams adopt faster when they know they can stop. The off-switch is almost never used, and it is almost always the reason they started.

Measure the second quarter

First-quarter usage measures novelty. Whether the tool is still open in month seven is the only number that has predicted a durable rollout for us.