Guide
Making it work, Part 2 of 2
Share this
Email me a PDF copy
Thank you!
The ‘Making it Work?’ guide has been sent to your email.
In this guide we cover
No single layer is sufficient on its own.
Deterministic tooling is exact and auditable but cannot write a narrative. Generative AI is fast but will invent things if left ungrounded. Human judgment carries context and accountability but does not scale to a 200 page solicitation under deadline. Error risk is a function of task structure, and task structure is something you design.
Grounding is four properties, not a setting.
Scope, attribution, freshness, and refusal. A system that never returns “no source found” is not grounded, it is just confident. The vendor supplies the mechanism, your content library decides whether it has anything to hold onto.
A pilot buys you the right to go big.
One intact team, one deliverable type, a fixed 60 to 90 day window, a baseline captured before anyone logs in, and written stop criteria. Governance, measurement and a retraining path belong on day one, not phase two.
Adoption drifts when the reinforcement stops.
In an anonymized VisibleThread pilot with a large defense contractor, usage fell close to 40% the moment the cadence ended, and recovered to 90% in four weeks only because measurement and ownership were already in place. Track outcomes rather than logins, and pair every participation metric with an outcome metric.
