The Signal
A recent study found that 95% of enterprise generative AI pilots fail to deliver measurable profit-and-loss impact. The reason is rarely the technology — it's integration gaps, data problems, and unclear ownership. Meanwhile, McKinsey reports that 79% of organisations saw real ROI from at least one AI initiative. That gap tells you everything about how the winners approach this differently.
The Story: Anatomy of an AI project that actually worked
The most expensive sentence in business right now is "We need AI."
Not because AI doesn't work — it does, remarkably well for the right problems. But because starting with the technology instead of the problem is how you end up in the 95% that fail. Every successful AI project I've seen started the same way: not with "let's use AI," but with "what takes too long?"
Start with friction, not fascination.
Here's a pattern that actually works. A mid-sized logistics company — about 120 employees — was drowning in manual invoice processing. Their accounts payable team spent roughly 30 hours per week extracting data from invoices, matching them to purchase orders, and flagging discrepancies. It was repetitive, error-prone, and the team hated it.
They didn't start by shopping for AI tools. They started by asking three questions:
The three-question filter:
Is the task repetitive? If it follows the same pattern hundreds of times a month, AI can likely handle it. Invoice processing? Absolutely — same format, same data fields, same matching logic.
Is the data available? AI needs something to work with. If the invoices are sitting in email attachments, scanned PDFs, and a shared drive — that's messy, but it's available. If the process lives entirely in someone's head with no documentation, you've got a bigger problem to solve first.
Is the outcome measurable? You need to know what "worked" looks like before you build anything. For this company, the metrics were clear: hours spent per week on processing, error rate, and time-to-payment.
If a task passes all three, it's a strong candidate. If it fails any one of them, fix that gap first or pick a different process.
The 90-day pattern
The company followed a structured timeline — and this is the framework LF Labs genuinely uses because it consistently works:
Weeks 1-2: Assess. Map the current process end-to-end. Quantify the baseline (30 hours/week, 8% error rate, average 12 days to payment). Identify the data sources and integration points. Set the target: reduce manual processing by 40% within 90 days.
Weeks 3-6: Build. Develop the solution — in this case, an AI-powered document extraction system connected to their existing accounting software. The key here was not building the fanciest system, but building the one that plugs into what they already use. The best AI solution that nobody adopts is worthless.
Weeks 7-10: Train. This is where most pilots die. The technology worked in testing, but the AP team hadn't been brought along for the ride. Training wasn't just "here's how to use the new tool." It was "here's why this exists, here's what it does well, here's what it doesn't do, and here's when to override it." People adopt tools they trust and understand.
Day 90: Measure. The result: manual processing time dropped from 30 hours per week to 11. Error rate fell from 8% to under 2%. Time-to-payment shortened from 12 days to 4. The team didn't lose anyone — they redirected those hours to supplier relationship management and early payment discount capture, which generated an additional $40K in annual savings.
The real lesson? The technology was maybe 30% of the success. The other 70% was picking the right problem, getting clean data, building around existing workflows, and — most critically — investing in training and adoption.
According to McKinsey's latest research, organisations that invest in change management alongside AI deployment are 6x more likely to achieve their target ROI. That's not a marginal advantage. That's the difference between the 5% that succeed and the 95% that don't.
The Operator's Toolkit: The AI readiness scorecard
Before you invest in any AI initiative, score yourself honestly. Give each question a score of 1 (not at all), 2 (partially), or 3 (yes, clearly).
Data & Process
- Can you identify at least one process that is repetitive, manual, and time-consuming?
- Is the data for that process available in digital form (spreadsheets, databases, documents)?
- Is the data reasonably clean and consistent — or do you know what it would take to clean it?
Team & Culture
- Does your team have someone willing to own the AI initiative day-to-day?
- Is leadership actively supportive (not just curious, but willing to allocate time and resources)?
- Has your team shown openness to adopting new tools in the past?
Strategy & Measurement
- Can you define what success looks like in specific, measurable terms?
- Do you have a realistic budget for implementation, including training — not just software?
- Is there a clear timeline and milestone structure for the first 90 days?
Integration
- Can the target process be improved without replacing your core systems entirely?
Your score:
24-30: Move now. You have the foundations in place. Start scoping a specific project.
15-23: Prepare first. The intent is there, but close the gaps — especially in data readiness and ownership — before committing budget.
Below 15: Not yet. Focus on the fundamentals: clean up your data, document your processes, and build internal alignment. The technology will be there when you're ready.
The Radar: Three things worth knowing this week
PwC's 2026 AI predictions call it "the disciplined march to value." The era of AI experimentation for experimentation's sake is ending. Companies are shifting from pilot culture to ROI discipline — and the ones that made that shift early are pulling ahead.
Customer service automation delivers 340% average ROI within six months, according to Zendesk data. It's one of the clearest quick-win use cases, and a strong first project for businesses scoring in the "prepare first" range above.
IDC projects that 80% of enterprise applications will embed AI agents by end of 2026. This isn't about standalone AI tools anymore — it's about AI becoming a layer inside the software you already use. The question isn't whether your tools will have AI, but whether you'll be ready to use it well.
From the Field
I'm curious — where does your organisation land on the readiness scorecard? If you ran through those ten questions, what surprised you? Hit reply and let me know. I read every response.
— Francois