How Moldman rebuilt operations with a custom AI system after fraud

Executive overview

Hyper-growth without process redesign creates control gaps — Moldman lost $1.4M to fraud after tripling in size during Covid. The crisis forced a full operational rebuild, and AI became the foundation for doing more with fewer people.

Gerard Murtagh's five-step framework — identify, cut, automate, offshore, optimise onshore — resequences how businesses should approach AI investment. Start with process elimination, not automation of existing waste.

Building a proprietary data lake before deploying any AI tools is the prerequisite most businesses skip.

The fraud that triggered transformation

• 300% growth in under a year meant 80–100 new hires joining a business built on personal trust
• Same credit card and approval processes scaled to 160 people — no new controls
• $600K directly misappropriated; $1.4M total hole including unpaid obligations
• First AI use: uploaded all transactions as CSV into a ChatGPT wrapper; surfaced anomalies instantly — $9K Uber Eats in one month, 19 Spotify subscriptions, flights booked under the founder's name
• Lesson: don't add a new process or a new hire to fix a control failure — find the technology that removes the failure entirely

The five-step process framework

1. Identify — define the exact process and the problem it creates
2. Cut — remove steps ruthlessly; if you don't need to add 10% back, you didn't cut hard enough (Musk rule)
3. Automate — only after cutting; automating a broken process builds expensive, useless systems
4. Offshore — if it can't be automated, evaluate Philippines or similar before hiring onshore
5. Optimise onshore — reserve human capacity for high-value judgment tasks

Building the data lake (Ty)

• Moldman built a proprietary LLM on Azure, named "Ty," fed with 2M+ emails, all phone calls, scripts, podcasts, and every interview
• The data lake is the "car" — the engine (GPT, Grok, Llama, Claude) can be swapped without losing the asset
• Ty writes marketing copy in Moldman's brand voice; answers mold questions; will handle inbound calls as an AI agent within weeks of filming
• Security point: never put identifying company data into public ChatGPT — you're training their model, not yours
• If Azure falls behind, the underlying engine swaps out; the data lake stays intact

Job intake: email to job management in 15 seconds

• Property managers send work orders via their own job systems — API links break whenever vendors update, so Moldman abandoned that approach
• Ty reads incoming emails, classifies them, extracts job data from attached PDFs, and creates the job record automatically
• If information is missing, Ty sends follow-up questions to the client before a human sees it
• Human in the loop reviews a prepped, complete job record — confirms or corrects, then feeds the outcome back to Ty as training signal
• Target: reduce human checking from 100% of jobs to 5% as accuracy improves

Route scheduling: 57% better optimisation than humans

• Moldman has tracked vehicles for years — that historical data trains Ty's scheduling model
• Google Maps API feeds live traffic; model avoids schools at 3pm, known bottlenecks at peak hours
• When a new job comes in, Ty checks whether a technician is already nearby and slots it in — turning a two-week wait into same-day
• A full day's run planning now takes under a minute vs. a human scheduler's full day
• Future state: live rerouting if accidents occur, automatic job reshuffling without human input

Customer experience and FAQ standardisation

• Frequently asked questions now answered uniformly by Ty — no over-promising, no scope creep
• Weather event triggers: when a flood hits a city, Ty identifies affected clients and sends proactive outreach
• Within the filming window: quote, job booking, and credit card capture all completed before the customer hangs up
• Invoicing and collections handled offshore (Philippines team), double-checked onshore — an 8-minute task reduced to a 15-second check

Managing internal resistance to AI

• Staff build processes that validate their own roles — they resist automation because it threatens their perceived value
• The message that lands: "I'm not removing you, I'm giving you an 8-minute task as a 15-second check"
• Human in the loop is the framing — AI handles volume, humans handle judgment and edge cases
• Reality: staff who won't engage will eventually face the same automation under a different CEO
• Moldman went from 36 admin staff to 15–16 while doing more revenue than at peak headcount

The competitive urgency argument

• At a 500-person event, 80% said they'd wait for their software vendors to build AI for them
• Off-the-shelf AI from HubSpot or any job management platform gives everyone the same capability — no differentiation
• The differentiator is proprietary data plus process — companies investing now are building a moat
• Large competitors are spending millions; small operators waiting will be left with low-margin, undesirable clients
• The near-term endgame: a customer calls, gets an instant quote and books the job before hanging up; the competitor says "we'll come inspect next week"

Tools and staying current

• Maintain a separate Instagram or X feed following only AI accounts — consume it in a fixed window, then exit
• Use multiple LLMs: different models suit different tasks; treat the big LLMs as interchangeable engines
• Data lake built on Azure; applications layer plugs in via API from job management system, CRM, phone system
• Credit card control solved with Airwallex — virtual cards on mobile, instant freeze/cancel, no new headcount needed