Subscribe to our newsletter
The CIO of a Health company stood on the worn carpet outside the main boardroom, the hum of the HVAC a low thrum in her ears. She wasn’t nervous about the technology—her team had built a brilliant proof-of-concept. She was anxious about the CFO.
In her mind, she could already hear his polite but pointed questions: “Another AI pilot? What’s the hard ROI on this one? How is this different from the last three?”
She took a deep breath. This time, it was truly different, because they hadn’t started with the technology. They had started by asking the right question.
Sound familiar? In dozens of conversations over coffee with technology leaders across the country, I’ve heard a version of her story again and again. The initial, breathless hype around AI has given way to a pragmatic—and sometimes painful—reality check. The C-suite is tired of expensive science projects; they want results.
Brutal reality: the variable that most reliably separates success from failure isn’t budget size or the PhDs on payroll; it’s matching the AI model to the mission on Day One.
Why Model Choice Still Trips Up the C-Suite
Let’s be honest: as an industry, we got ahead of ourselves. The arrival of powerful, general-purpose models felt like a magic bullet. But as many have learned, once you own a shiny new hammer, every problem starts to look like a nail.
The result? Roughly 54% of AI initiatives never graduate from pilot to production (Gartner, 2022); they languish in what one director of innovation calls “the pilot graveyard.”
“We spent six months and nearly a quarter-million dollars trying to make a giant language model predict sepsis,” a hospital COO confessed. “It was like asking a brilliant history professor to perform heart surgery—wrong tool, wrong training, wrong result.”
The issue isn’t a technology shortfall; it’s a strategy mismatch. We get dazzled by what a model can do and forget to ask what it should do for our business. Leaders like her succeed because they bridge that gap—starting with a straightforward, plain-English menu of options.
The Four AI Model Families
| Model Family | Handy Metaphor | Best For | A Word of Caution |
| Foundation Models | The Swiss-Army Knife | Drafting emails, summarising reports, powering chatbots | Prone to “hallucinating” facts; not ideal when accuracy must hit 95 %+ |
| Predictive-Analytics Models | The Seasoned Detective | Forecasting churn, demand, fraud, or maintenance events | Needs large volumes of clean, labelled historical data |
| Computer-Vision Systems | The Eagle-Eyed Inspector | Defect detection, medical imaging, planogram compliance | Training can be costly—often thousands of meticulously labelled images |
| Domain-Specific Models | The Michelin-Starred Chef | High-stakes expertise: legal discovery, genomic diagnostics | High upfront CAPEX and longer time-to-value; invest only for core advantage |
Match the Model to the Mission: A Three-Question Litmus Test
Question 1 – What’s the real problem?
- Forecast or classify a numeric/business outcome ➜ Predictive model
- Read, write, or summarise language ➜ Foundation model
- Recognise something visual ➜ Computer vision
- Replicate specialist judgement ➜ Domain-specific model
Question 2 – What’s the cost of being wrong?
- Low stakes (awkward email copy) ➜ Foundation model is fine
- High stakes (missed cancer, botched contract) ➜ Domain-specific or tightly governed predictive/vision model
Question 3 – What’s the reality of your data?
- Years of structured, labelled records ➜ Predictive paradise
- Mostly unstructured text ➜ Foundation models thrive
- Thousands of accurately labelled images? Great—computer vision is viable. If not, budget for the significant data labeling lift first.
Data Readiness: The Awkward Conversation Nobody Wants to Have
“Everyone thinks they have good data until they actually have to use it,” She laughed. She’s right—it’s the skeleton in every enterprise closet.
Before green-lighting any AI project, leadership must get brutally candid about data:
- Source of truth—Is there one, and can we access it?
- Quality—Is it clean, consistent, and de-duplicated?
- Governance—Who owns it, and under what rules can we use it?
It isn’t glamorous; it’s data archaeology. But without it, even the sexiest model will flop.
Cost vs. Performance: A CFO-Friendly Breakdown
Your CFO doesn’t care about GPUs or token limits; they care about CAPEX, OPEX, and payback.
Example—Automating first-pass legal review (100k pages/month)
- Path A – API Route (OPEX-heavy): $0.02/page → $24 k / year. Near-zero setup, cost scales linearly.
- Path B – Custom Model (CAPEX-heavy): $350 k build over nine months; post-launch cost ≈ $200 / month. Breakeven ≈ 18 months, plus 10-20 pp accuracy lift.
The right path hinges on volume, timeline, and strategic ambition—not hype.
Industry Vignettes: The Right Model in the Wild
Finance – Taming the Call Centre
A national bank fine-tuned a foundation model on two years of transcripts—result: 35 % drop in handle time and early detection of emerging issues.
Healthcare – Averting Crises in the ICU
A custom predictive model flags sepsis risk up to 24 hours earlier, reducing escalations by 27% (JAMA 2023). High upfront cost, but the price of failure was higher.
Retail – The End of “I Can’t Find It”
Computer-vision search lets shoppers upload a photo and see matching inventory, driving a 28 % conversion-rate lift.
Risk, Regulation & Responsible AI
Deploying AI means becoming a steward of ethics and compliance:
- Explainability—Can you show an auditor how decisions were made?
- Bias auditing—Continuously test for disparate impact.
- Reg-tracking—Map GDPR, the EU AI Act, and FDA guidance to your roadmap.
Innovative governance isn’t a checkbox; it’s ongoing risk insurance.
A Phased Playbook That Actually Works
| Phase | Duration | Goal & Key Question |
| Discovery | 4-6 wks | What business problem is worth solving? |
| Proof of Concept | 6-10 wks | Does the tech work on our data? |
| Pilot | 3-4 mo | Will real users adopt it, and will KPIs move? |
| Scale | Ongoing | How do we industrialise, monitor & iterate? |
Closing Reflection
When she finally stepped into the boardroom, she wasn’t pitching an AI model; she was presenting a solution—backed by a clear business case, a map of alternatives, and data-driven answers for every CFO curveball.
The journey to AI value isn’t about wielding the most potent model; it’s about wielding the right one for your mission, budget, and data reality.
Feeling stuck between a foundation model and a custom build?
Book a 30-minute, no-pressure strategy call to bring clarity and direction.
Let’s map your goals to the tech that will actually deliver.