Why most AI projects fail — and what to do instead

Last issue I showed you my personal AI setup — the tools I’m running outside of work, the experiments, the things that surprised me. A few of you replied saying it felt exciting but also slightly alien to your day job. Fair point. Most of us aren’t running personal AI labs. We’re sitting inside large organisations with Microsoft Copilot licences, cautious IT departments, and boards that want to see ROI before approving anything.

So this week, I want to talk about corporate AI — and specifically, why so many projects in this space quietly die.

I should be honest: I’m on this journey too. I’m not reporting from the other side. But I’ve been close enough to enough AI initiatives — as a director, as someone building my own AI infrastructure in parallel — to recognise the failure patterns when I see them. And they’re remarkably consistent.

Failure mode one: confusing automation with intelligence. Most corporate AI projects are still thinking like it’s 2018. They’re looking for repetitive, structured tasks to automate — and yes, AI can do that. But that’s the floor, not the ceiling. The real capability of modern LLMs isn’t task automation. It’s synthesis, reasoning, and generation at a level that used to require expensive human expertise. If your AI initiative is a glorified if-then rule engine, you’re not using the technology — you’re using the idea of the technology.

Failure mode two: no owner who owns the failure. Every AI project that works has one person who is personally accountable for it working. Not a committee. Not shared responsibility spread across IT and operations and a vendor account manager. One person who understands the technology well enough not to be misled by it, understands the process well enough to see where it actually fits, and has the authority to unblock things when they get stuck. They’re rarely the most senior person in the room. They’re the one who’s genuinely curious and won’t let it quietly die.

Failure mode three: treating integration as an afterthought. The demo works. The pilot works. Then it goes live — and nothing connects. The AI tool operates in isolation. Its outputs don’t flow into the systems people already use, so everyone reverts to the old way within a month. Integration isn’t a technical detail you hand to IT at the end. It’s the whole strategy. An LLM that doesn’t connect to your data, your workflows, and your outputs is an expensive chatbot. The organisations seeing real returns have wired AI into how work actually moves — not bolted it on the side.

What do the projects that work have in common? They’re targeting problems where the bottleneck isn’t execution — it’s thinking. Reading, synthesising, connecting dots across large amounts of information, turning raw data into something a human can act on. That’s where modern AI earns its keep. And it’s almost exactly where most corporate AI programmes aren’t looking.

More on that next week.

Next issue: where AI actually works brilliantly in corporate environments — and why the answer might not be where you’re looking.

When I first looked at AI avatar tools, my honest reaction was: interesting, probably gimmicky. Twelve months later I’ve revised that. Here’s where it’s genuinely useful and where it isn’t.

Where it works: Internal training content that needs a face — policy updates, process walkthroughs, onboarding. Social video where consistency of format matters more than raw production quality. Anything that needs updating regularly — re-record the script, same avatar, same production standard, in minutes.

Where it doesn’t: Anything where your audience knows you and expects authentic energy. A town hall. A client pitch. Anything emotionally charged. Avatars are scripted by definition — anything that needs genuine spontaneity doesn’t work here.

I use HeyGen for my YouTube content and some internal explainers. The avatar quality is now at the point where most viewers don’t flag it. Six months ago that wasn’t true.

Verdict: 7.5/10. A few hours to set up correctly. Once your avatar is trained, the speed advantage is real and the pricing is reasonable for what it does.

→ I use HeyGen for all my YouTube videos — try it here (affiliate link — I earn a commission if you sign up, at no extra cost to you)

If you haven’t found the channel yet — everything I write about here, I’m also building and testing on camera. Five videos are now live:

Video #001 — Stop Accepting the Default Answer
Why senior leaders accept vendor solutions without questioning whether a better alternative exists. The first video in the series — this is what the channel is about.

Video #002 — I Built a Full YouTube Automation System in 2 Hours
Using OpenClaw and AI, I built a complete content production workflow from scratch. Strategy layer, infrastructure, weekly ops. The whole thing documented in real time.

Video #003 — My AI Lost Its Mind (And How to Prevent It)
AI memory architecture explained practically — why most AI tools have no persistent context, and the checkpoint system I use to solve it. Directly applicable if you’re building any kind of AI workflow.

Video #004 — What Happens When Your Two AIs Talk to Each Other
I use Microsoft Copilot at work and a personal AI lab at home. In this video I explain how to use a personal AI to engineer better prompts for your corporate AI — and why it’s fully governance-compliant.

Video #005 — Don’t Lose Your Voice to AI
This is about using AI without sanding off your own judgement, tone and point of view. If Copilot, ChatGPT or Claude can make everyone sound competent, your actual voice becomes more important, not less.

👉 Subscribe, like, and hit the bell — new videos every 2 weeks.

If any of these are useful, sharing them with a colleague is the single biggest thing you can do to help the channel grow.

1. The organisations building internal AI governance frameworks now — acceptable use policies, data handling rules, model selection criteria — will be significantly better positioned for compliance requirements in 12 months. It’s the boring infrastructure work that pays off later. Even a one-page policy is better than nothing.

2. Google’s “Agents Companion” whitepaper on AI agent system architecture is worth your time if you’re thinking about where AI automation goes next in your organisation. Dense, but worth it. Free via Google’s developer docs.

3. The teams most likely to successfully adopt AI in your organisation aren’t always the most technically sophisticated. They’re the ones with the clearest, most repeatable processes. AI amplifies clarity. It doesn’t create it.

“We want to start an AI pilot but the board wants to see ROI before approving budget. How do you break the chicken-and-egg?”
— Operations Director

This one I’ve thought about a lot, because it’s the trap almost every organisation hits. The answer is: start smaller than you think is credible. Not a pilot — a proof of concept. Scoped to a single process, four weeks, one named person accountable, one pre-agreed metric you already measure. Hours saved, error rate, cost per unit — pick one number and agree it upfront. Most boards will approve a small, time-boxed test with defined success criteria even when they’d reject a broader “AI programme” proposal. Win the small test cleanly, then scale the ask. The mistake is trying to justify the whole vision before you have a single data point.

Two things:

Reply with your AI story. A win, a failure, or something still in progress. The best ones (anonymised where needed) will feature in a future issue. I read every reply.

I’ve also put together a practical beta toolkit. The AI Workflow Audit Pack is for assessing which workflows are actually ready for AI, which should be parked, and what a sensible first pilot could look like. It includes a workshop deck, scoring workbook, completed worked example and template documents. Founder price is £29; Gumroad may display checkout in USD depending on buyer/payment settings. 10% of gross sales supports Action for Children’s Boycott Your Bed campaign.

Share this newsletter. If someone in your network would find this useful, forward it or send them the subscribe link: theaidirective.co.uk/subscribe. Word of mouth is how early newsletters grow — and it’s genuinely appreciated.

See you next Tuesday.

— Toby

@TobySmithFirth on YouTube

These are tools I use personally. Affiliate links marked — I earn a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend tools I actually use.