I started building GenAI apps for clients in mid-2023 as part of my consulting work.
The hype was high. Everyone—me included—was dazzled by early ChatGPT and those first LLM APIs. Businesses put money on the table, chasing 30–40% productivity gains “by next year” (thank you, McKinsey PDFs).
(It turns out 10% is already spectacular after a couple of years).
Honestly? I believed it, too.
We spun up a team, set governance, picked the tech, partnered with lines of business and product owners—and in three months we had 5**–6 apps demo and test-ready**.
That’s when the real story started—and when “GenAI will revolutionize everything” met reality (at least for me).
We weren’t naïve; we knew it wouldn’t be magic. We just couldn’t name the reasons why until we shipped.
In a nutshell, three bottlenecks block AI from delivering real value at scale:
- Adoption — no use, no value.
- Trust — would you stake your money or brand on the output (without verification)? Or…
- Verification time — does the win survive after you re-check the AI’s work?
A recent MIT study, part of the NANDA initiative, found ~95% of business AI projects never deliver real ROI. Companies poured $30–40B into AI hoping for miracles; most got all splash, no cash.
The study went viral, so I’m combining key insights from the paper with what I’ve learned about building AI—what works, and what will never work.
This is the guide I wish I had a couple of years ago. It keeps the full story, and turns it into a practical system.
Let’s go.
The Great AI Letdown (Why 95% Flop)
Myth: "It's going to work (take over the world). It's just the tech that isn't strong enough yet."
Reality: The tech is already fine. Workflows, adoption, goals and TRUST are the problem.
What consistently goes wrong:
- The Learning Gap
Most tools are one-and-done: they do a task, then forget everything. No persistent memory. But more importantly (especially for GenAI apps): NO improvement with feedback. Great on stage, useless in messy workflows.
Would you trust an assistant who wakes up amnesic every morning? Exactly.
- Human Factors Beat Model Accuracy
The #1 blocker wasn’t cost or accuracy—it was resistance to change. If the outputs feel off or the UX is clunky, people stop using it. No adoption = no ROI.
People are also usually afraid of seeing their jobs eliminated by AI. They're often happy with the status quo and reluctant to invest extra effort when AI fails to do the tasks they used to do (they feel important and useful).
- Shiny-Object Syndrome
"Everyone's doing AI—let's do AI." With no clear business problem, pilots wander, stall, and die. If you can't name the KPI you'll move, you don't have a project—you have a hobby.
This is unfortunately amplified by the whole "LinkedIn / social media vibe"…
- Data: Ferrari, Empty Tank
Siloed, messy, or scarce data = expensive nothing-burgers. Many teams discover late they lack a data foundation or that cloud costs nuke the budget.
For advanced people: lack of data foundation or an MCP middleware to consume data.
If you have none, there's no AI at scale… The company is just a collection of AI enthusiasts doing "stuff" with their Copilots or ChatGPT…
- Culture Clash
Job anxiety + weak training = quiet sabotage. People keep the old process and smile in meetings.
Stop with "AI is going to take over all human jobs." We'll need more trained and skilled humans to manage AIs and to do more meaningful work than before. Say this to your people and make it your culture. Otherwise you'll never benefit from humans nor AI.
- The Shadow AI Economy
At ~90% of companies, employees use tools like ChatGPT off-label while only ~40% have official deployments. Value is happening, eventually—just not where IT expects.
People keep the value for themselves. If they save 3 hours a day, they would rather use those hours for their own benefit than "giving them back" to the business ;)
That's fine and positive. AI should benefit humanity as a whole and kick us out of the "grind" culture. Fix the business/enterprise/team culture. Otherwise (again) you'll never benefit from humans nor AI.
- Uneven Impact by Industry
Real structural change so far? Mostly Technology and Media/Telecom. Elsewhere it’s lots of experiments, little transformation. Information-native work is AI’s home turf; physical/regulated workflows take more redesign.
Key takeaway: Most failures aren’t about weak tech. They’re about humans, culture, connection, trust, memory, workflow designs, adoption, or no KPI.
If you're a business owner or team manager, you know what you have to do.
If you're in a team, share these insights with your team and/or organization.
What Successful Teams Do Differently
This letter is not just about the why and what of AI failure. The recent AI tech leap is a major one in human history, and I've worked on numerous projects that are an immense success.
The winners ran a better playbook.
1. Put business first (not model first).
Tie AI to a hard metric: “Reduce claim cycle time from 4 hours to 15 minutes by Q4.” Budget and training align to that.
2. Redesign the workflow.
Don’t bolt AI onto yesterday’s process. Rebuild the process so AI is the default path and humans handle exceptions.
(If you do this properly, you’re in the top 0.1% I would say).
3. Close the Learning Gap.
Choose / design tools with persistent memory and feedback loops. The system should get better as people correct it—goldfish ➝ intern ➝ dependable teammate.Don’t think complex systems. Simple choices matter:
- Use (or train people to use) longer conversations
- Use (or train people to use) “Assistant” features and update their prompts regularly
- Use (or train people to use) “Project” features and update their context regularly
4. Recruit power users (and one skeptic).
Let the folks already tinkering with AI co-design prompts, SOPs, and quality checks.
Golden rule: Adoption follows peers—not memos.
5. Partner for speed.
External expertise often 2× the odds of reaching production vs. pure DIY. Ship value now, insource later.
This is why my clients work with me. They just want to go 5–10× faster and learn from others' mistakes. Simple as that.
6. Hunt unsexy money.
Biggest wins are back office: invoices, call summaries, eligibility checks, scheduling, doc extraction. Quietly saves millions. Zero fireworks. Maximum ROI.
7. Measure outcomes, not vibes.
Contract on hours saved, errors reduced, dollars saved/earned, CSAT. If it doesn’t move, fix it or kill it.
Two Quick Examples (Composite, but realistic)
A) The Flop: “One Bot to Rule Them All”
Retail launches a generic chatbot for everything. It forgets context (or connected to a very large context), can’t look up orders, and punts edge cases to humans.
- Result: Support volume unchanged, agent workload higher (clean-up duty).
- Why: No data access, no memory, no workflow redesign, no KPI.
B) The Win: Invoice Triage + Reconciliation
Ops targets one pain. AI reads PDFs, extracts line items, matches POs, flags mismatches. Human reviews exceptions. Corrections are collected, synthesized, new prompts and parameters are generated and tested to validate improvement.
- Result: 68% cycle-time reduction, 30% fewer errors, ~$1.2M/year saved.
- Why: Narrow scope, integrated data, feedback loop, ops-led rollout, measured outcomes.
The 5% Playbook (Use This, As-Is)
1) Pick One Boring, Costly Process
- Claims summarization, KYC verification, loan pack assembly, L1 IT tickets.
- Litmus test: high volume, repeatable, rules with exceptions, measurable pain.
- Write the KPI: “Cut [X] by [Y%] by [DATE].”
2) Map the Real Workflow (20–30 min)
- Steps ➝ failure points ➝ rework loops ➝ data sources/owners ➝ decisions.
- Mark what must be human vs. what could be AI.
- Estimate if the outputs must be checked or not by humans (depending on trust) and verification-time by humans/owners.
3) Choose Tools That Remember (Sounds basic, I know)
- Must-haves: memory, feedback capture, audit trail, RBAC, APIs.
- Nice-to-haves: in-app prompts help, adoption nudges, usage analytics.
4) Recruit/Train 3 Power Users + 1 Skeptic
- Make them authors of the SOP, reviewers of outputs, and trainers of the model.
- Incentivize with recognition and small rewards (yes, gift cards work).
5) Ship a 30-60-90
- 0–30 days: Integrate data, ship V1, baseline metrics.
- 31–60: Close feedback gaps, handle edge cases, expand to adjacent tasks.
- 61–90: Lock KPIs, automate reporting, train next cohort.
6) Contract on Business Outcomes
- Vendors: tie fees/bonuses to KPIs.
- Internal: publish a weekly AI Scorecard to leadership and the frontline.
Cheat-Sheet (print this):
- Clear problem + KPI.
- Tools that learn + remember.
- Partner > perfection.
- Measure on your data.
- Design for adoption.
What Changes (and When)
Right now, people trust AI for first drafts and quick analysis. Roughly 70% are fine with AI doing the first pass on emails or basic numbers. For high-stakes work? ~90% still choose a human. Fair.
The shift happens when your system earns trust:
- It remembers yesterday.
- It improves with corrections.
- It removes annoying work without adding new annoying work.
- It performs well most of the time at scale (that's not easy to do and measure).
You don’t force adoption; you win it by making the right path the easy path.
Quick, Tangible Use-Case Ideas to Steal
- Banking: Auto-assemble loan packs (IDs, income docs, policy checks). Human signs off. Hours ➝ minutes.
- Healthcare: AI summarizes patient calls to EHR with structured fields. 7 minutes ➝ 30 seconds review.
- Insurance: First-notice-of-loss triage with fraud flags; adjusters start with context, not chaos.
- HR: Parse resumes, score against must-haves, draft outreach with rationale. Recruiters spend time interviewing, not sorting. (Not very compatible with EU regulation, thought…)
- Manufacturing: Extract specs from vendor PDFs to ERP, flag inconsistencies before line-down events.
Your One-Week Action Plan
This is not only for businesses and teams. Try to do this for your own workflows.
By next-week :
- Pick one process and write the KPI in one sentence.
- Map the workflow (people, steps, decisions, data / data owners).
- Draft the 30-60-90, pick tools with memory, recruit 3 power users.
Save / Send a one-pager to your sponsor with the subject: “Our 5% Plan.” If they don’t reply, walk it over. Real change starts with this boring page.
Steal This Script (for your kickoff)
You:
[process]
[X%]
[date]
The system remembers context and learns from corrections.
We redesigned the workflow—this isn’t a bolt-on bot.
Your role: use it, correct it, tell us what’s annoying; we’ll fix the annoying.
If it doesn’t move the KPI, we kill it. Deal?”
Short. Clear. Adult-to-adult.
If you made it this far, you now have a simple, gutsy plan to join the 5%. Not with a moonshot—with a narrow, learning system that makes someone’s Tuesday better.
Hit reply and tell me the one process you’ll target. I’ll do my best (no hard feelings if I don’t) to send back a prompt pack + workflow sketch you can copy.
Have a great day :)
— Charafeddine
Sources: MIT “GenAI Divide” research & analysis, business press coverage of AI pilot ROI, market commentary from 2025.