Moving AI from Experimentation to Practical Business Adoption

Context

Like most large organizations, the firm had no shortage of AI enthusiasm and no shortage of tools — but a real gap between experimenting with generative AI and actually embedding it into day-to-day work in a way people trusted enough to rely on. Several promising ideas had stalled at the pilot stage.

Challenge

The hard part was never proving a large language model could do the task in a demo — it was designing workflows where the AI's output was reliable enough, reviewable enough, and integrated well enough into existing tools that professionals would actually use it instead of quietly reverting to the old manual way. That meant solving for data readiness, human oversight, and change management at the same time, not sequentially.

My Role

I led use-case identification and solution design for several applied AI initiatives, working with technical teams to move specific use cases from experimentation into supervised production use.

Approach

We started by being deliberately selective about use cases — prioritizing tasks that were high-volume, well-defined, and where an error was cheap to catch, rather than chasing the most impressive-looking demo. Examples included classifying fixed-asset transactions for general ledger coding, supporting SR&ED and R&D tax-credit eligibility review, and summarizing vendor and client email threads to speed up response times.

For each use case, we built in human-in-the-loop review at the points where errors mattered most, rather than treating oversight as an afterthought bolted onto a finished tool. We also invested in prompt design and light training for the professionals who'd be using these tools day to day — including direct coaching for tax professionals adapting to AI-assisted workflows — because adoption depended as much on their comfort and trust as it did on the tool's accuracy.

Governance mattered here in a very practical sense: clear rules about what the AI could and couldn't decide on its own, and a way to track where it was being used and how well it was performing, so leadership had real visibility rather than anecdotal confidence.

Outcome

Several of these use cases moved from one-off experiments into repeatable, supervised parts of day-to-day work, freeing up professional time for judgment-heavy tasks rather than repetitive review. Just as importantly, the use-case selection and governance approach became a pattern the team could reuse for the next AI idea, instead of starting the trust-building process over each time.

Capabilities Demonstrated