Proof is proof,
whether it says yes or no.

Garrison & Associates had built a strong reputation in commercial litigation over 14 years. Associates were spending an average of four hours reviewing each incoming contract before producing a summary memo for the supervising partner, handling 18 to 22 contracts per month. That was roughly 80 hours of non-billable associate time consumed by a single workflow.

The managing partner had reviewed a large AI agency proposal for $65,000 and declined. The ROI case was not clear enough to justify without evidence first.

The partners were not opposed to AI. They were opposed to AI that produced legally imprecise summaries. The firm had a specific memo format refined over a decade that any AI output had to match.

We agreed on a blind evaluation rubric before the PoC: five quality dimensions including legal accuracy, clause identification, and structural adherence. Both AI and associate outputs were scored by two senior partners who did not know which was which.

In early 2025, Hartwell Advisory Group decided to pursue an AI-driven automation of their audit preparation workflow, collating, categorising, and summarising client financial data ahead of annual audits. It consumed approximately 120 hours of senior staff time per audit cycle.

Rather than validating the concept first, the firm moved directly to implementation. They engaged a generalist software contractor through a referral, allocated an internal project manager, and began building a custom AI workflow tool designed to ingest client data and produce structured audit prep summaries.

Seven months in, the tool had produced inconsistent outputs across different client data structures, failed to handle the firm's non-standard chart of accounts formats, and generated summaries that required more time to correct than the original manual process.

The firm abandoned the project. The tool was never deployed. The contractor was terminated. The internal project manager had spent an estimated 30% of their working hours on the initiative across seven months, none of it recoverable.

Twelve months after the failed implementation, a director identified a new AI opportunity: automating the narrative commentary in their quarterly management accounts, the written analysis accompanying financial data packages for clients. It consumed approximately 180 senior hours per quarter.

This time, the firm did not move directly to implementation. They engaged Perceptrus to run a Proof of Concept first, designed to test one specific hypothesis: can AI produce management accounts commentary that meets Hartwell's quality standard across their diverse client base?

The Proof of Concept ran across 28 real management accounts packages, a genuine cross-section including SMEs, holding companies, and not-for-profits. AI and human commentary were scored blind by three senior directors on technical accuracy, client-appropriate tone, variance explanation quality, and actionability of insights.

The result was unambiguous: AI performed adequately on simple single-entity accounts but failed materially on consolidated group accounts, entities with non-standard revenue recognition, and clients requiring industry-specific regulatory language. On 11 of the 28 packages, 39% of the sample, AI outputs fell below the firm's minimum acceptable threshold.

Northside Family Dental's practice manager had identified a persistent problem: clinicians were completing post-appointment documentation and follow-up correspondence outside of clinical hours, adding 45 to 75 minutes to their daily schedule.

The workflows in question were high-volume and repetitive: post-treatment care instructions, insurance pre-authorisation letter drafts, and appointment re-engagement messages. None required clinical judgement to produce. All required clinical language and HIPAA-aware communication standards.

Patient communications carry both a compliance obligation and a trust obligation. The practice principal made clear that any AI output would need to pass his personal review before the playbook was accepted.

We structured the Proof of Concept around three discrete document types, each with its own quality rubric. All outputs were reviewed blind by both the practice principal and practice manager independently before scores were compared.

Vantage Talent Partners had carved out a niche in mid-level financial services and operations placements. Their competitive advantage was the quality of their candidate briefs: detailed documents presenting shortlisted candidates to hiring managers with enough context to make confident interview decisions.

A thorough brief took a senior recruiter between 75 and 110 minutes from interview notes and CV data. With 15 to 20 active roles at any time and three to five candidates per role, the team was spending 50 to 80 hours per month on brief writing alone.

Candidate briefs were client-facing and the primary evidence of the agency's value-add. A brief that read as generic would signal exactly the opposite of what Vantage was known for.

The Proof of Concept tested AI brief generation across 30 real candidate profiles. Evaluators were three hiring managers at existing client companies who scored both sets blind on presentation quality, insight depth, usefulness for interview preparation, and overall impression.