Understanding the challenge
By early 2026 the firm had done everything the AI playbook asked of it. It had the budget, a data-rich business, and a leadership team that wanted in. What it did not have was a single AI tool running in production. Over the previous year it had launched a string of pilots, a proposal-drafting assistant, a research bot, a reporting helper, and watched each one stall in the same place. "We kept building things that demoed well and died in a month," the managing partner recalls. The problem was never the model. No one had defined where the AI was allowed to act, where a human had to sign off, or how anyone was meant to trust the output. The tools lived outside the systems the firm actually ran on, chiefly Dynamics 365 and its internal knowledge base, so they worked off copy-pasted data. Partners stopped relying on them, consultants quietly went back to doing the work by hand, and staff started pasting client information into personal AI accounts just to keep up. The firm was spending real money on AI with nothing in production to show for it.
Our approach
The firm did not need another pilot. It needed a way to run AI safely, in production, on its own operations. Our AI Operations Design engagement gave it that, structured as a clear Audit, Design, and Deploy:
• A two-week operations audit that mapped the real bottlenecks and returned 12 opportunities ranked by ROI, with 4 quick wins the team could adopt that week.
• An operating model that fixed exactly where AI acts, what it controls, and the human approval and escalation points on anything client-facing.
• A data architecture wired into the firm's existing Dynamics 365, knowledge base, and reporting systems, so the AI worked on live engagements instead of copy-pasted snippets.
• An adoption plan that made the sanctioned tools faster to use than the personal-account workarounds, so shadow AI disappeared on its own.
Technical innovation
At the core is a governed operating layer where every AI action is bounded, logged, and reversible, with a human sign-off on anything that reaches a client. Self-calibrating models learn from real outcomes, studying which proposals win and which drafts get flagged, so the recommendations sharpen month over month. Because it runs inside Dynamics 365, it acts on the firm's actual work, and the complete audit trail is what finally gave partners the confidence to let it run without watching it.
Outcome
Within a single quarter, five pilots that had been stuck for a year were live in production. Consultants got back about six hours each per week that used to go to manual reporting and coordination, a 62% reduction, and the average proposal now goes out in two days instead of five. Every action is logged and reversible, which gave partners the audit trail they needed to sign off, and the personal-account workarounds vanished once the approved tools were the faster option. Because the models recalibrate against real outcomes, the system is measurably better each month than the one before it. "For the first time, our AI runs the work instead of sitting in a slide deck," the managing partner says.
