Every engagement below follows the same discipline: a baseline captured before the work began, and results measured against it. Client identities are anonymised; the method is the point.
A manufacturer with multiple plants closed its books through a nine-day scramble of spreadsheet consolidation, inter-unit reconciliations by email, and manual journal preparation. Finance leadership spent close week firefighting instead of reviewing.
We baselined every close task — owner, hours, error incidence — before touching anything. Python automation took over data consolidation and recurring journals; reconciliation matching ran nightly through the close window with exceptions queued for the team each morning. Controls were designed in with the automation, not retrofitted, so the auditors accepted the new process at first review.
Close reached four days within three cycles. The controller's team redirected the recovered week into variance analysis and business partnering — the work the close had always crowded out. Automation payback landed inside the first year against the pre-project baseline.
The internal audit team tested controls annually on small samples. High-risk anomalies were statistically unlikely to appear in any sample, and findings surfaced months after the transactions they related to.
We deployed an internal audit agent with a written, approved mandate: re-perform defined control tests across the full transaction population daily, read-only access, all activity logged. Exceptions were documented with triggering evidence and queued for auditor judgement. The agent's own test logic was version-controlled and change-managed like any control — it had to pass the audit function's own standards.
Coverage inverted: full-population testing surfaced roughly four times the exceptions annual sampling had found, within days of occurrence. The audit plan was rebalanced toward judgement-heavy areas. The audit committee's question changed from "what did we sample?" to "what did we miss?" — answer: measurably less.
Previous AI training had produced enthusiasm and nothing else: high feedback scores, zero changed workflows. Leadership wanted capability that showed up in how work was actually done.
A multi-week programme covering the five-tool stack and vibe coding with Python, with one structural difference: between-session assignments required each participant to automate something from their real job, reviewed individually. The organisation's data-handling rules were embedded in every exercise. Success was defined upfront as artefacts in use at day 30 — not satisfaction scores.
Six months on, the function was running 60+ self-built automations, from bank-statement processing to report generation. The 30-day usage check — our standard honesty mechanism — confirmed the majority of built artefacts still in active use. The training budget line moved from "L&D" to "productivity investment" in the following year's plan.
Leadership believed AI "hadn't really started" in the organisation. A confidential shadow-AI discovery — anonymised survey plus tool-signal review — found seventy percent of knowledge workers using consumer AI tools weekly, including on client-confidential documents.
Amnesty first: nothing surfaced in discovery was punished, which made the data honest. An interim acceptable-use note shipped in week one. Sanctioned enterprise-grade tools replaced the shadow ones in week two — better tools than the banned ones, or shadow use continues. Weeks three to six built the durable layer: a two-page policy, a tool-approval workflow, a DPDP obligation map and a board oversight briefing.
Board approval in six weeks. More importantly, visibility: AI usage moved from invisible and ungoverned to registered and rule-bound, without the productivity loss a ban would have caused. The AI-use register now sits in the internal audit universe like any other controlled process.
Site engineers spent large parts of each week manually verifying subcontractor bills against BOQ line items and site records. Verification was inconsistent across sites, and the same document chaos weakened the group's position whenever contract claims arose.
Document intelligence extracted line items from incoming bills and matched them against BOQ and measurement records, flagging deviations for engineer review rather than auto-approving anything. The same extraction pipeline began structuring the project document flow — correspondence, instructions, delays — into a claim-ready evidence trail.
Every bill now gets checked — previously a practical impossibility — with engineers reviewing exceptions instead of everything. Around thirty percent of engineer administrative time returned to site work against the pre-project time baseline. The structured evidence trail proved its worth the first time a variation claim was assembled in days instead of weeks.
Store and back-office staff asked head office the same questions endlessly — returns policy edge cases, SOP steps, product specifications — because the answers lived across hundreds of documents in multiple versions. Wrong answers from outdated documents caused real customer-facing errors.
A retrieval-augmented (RAG) assistant grounded exclusively in the group's own current documents, with every answer citing its source paragraph. Document governance came first: a single controlled repository with version ownership, because a knowledge assistant is only as trustworthy as its sources. Confidence thresholds routed uncertain questions to humans rather than guessing.
Roughly half of routine queries to head office stopped arriving — answered at source, with citations staff could verify. The forced document clean-up delivered a second, unplanned benefit: one authoritative version of every policy, for the first time.
ENGAGEMENT PROFILES ARE ANONYMISED AND FIGURES ROUNDED. REFERENCES AVAILABLE ON REQUEST FOR QUALIFIED ENQUIRIES.
Every result above was measurable because measurement came first. Let's capture yours.