Challenges of AI Adoption

AI adoption is one thing, transformation is something else entirely.    Few companies have been bold enough to fundamentally change how they are organized and how decisions are made. At Pfizer, we are determined to break through that inertia and be the most AI-forward company in our industry because of what this technology will do for patients.    In R&D, as I said on our earnings call this week, our ambition is to build an AI-native organization where every insight, from target discovery through medical evidence, continuously informs the next decision.    We are making great progress, so I'd like to share more about how we think about it overall: why our 177 years of data is our alpha, why I chose a federated model over a central team, and why we are certifying every eligible colleague, including me, in AI fluency.

The uncomfortable math of AI transformation: 10% is the model. 90% is work your organization hasn't budgeted for. And that 90% is where almost every AI strategy quietly dies. It almost always starts ambitious at the top: sized opportunities, an approved budget, transformation targets, the competitive-advantage language everyone nods at in the boardroom. Then it has to fall through the reality gap into the actual work. Siloed and poor data. Legacy systems that won't expose what you need. Compliance and risk. Change resistance. And the silent killer: no clear owner. The strategy doesn't fail loudly. It quietly stops being anyone's job. This is the gap nobody budgets for, the distance between ambition and execution. The model was never the constraint. The operating model was. What survives the fall has a spine the boardroom version skips: strategy, then operating model, then workflow redesign, then integration and data, then adoption and enablement, and only then business value. Each step is a handoff where strategy can die, and the two that get cut first are workflow redesign and ownership. Which is exactly why pilots inflate and nothing has moved six weeks later. As far as I can tell, AI success is not a technology challenge. It is an execution problem, and the execution side does not improve on its own. PS: Every week in Human in the Loop, I break down where AI actually creates value and where it just creates demos: https://lnkd.in/dbf74Y9E

Of course you need to use open-source models if you’re an enterprise leader. Close model providers, that are now forcing data retention, are gaining immense leverage on your business if you don’t. As you connect models to your business context, they see it and learn from it, and have a track record of going after their most successful customers thanks to this information. But that’s not enough, you also need to store your data and records in open systems, or your software vendors might block you from building AI systems outside of the walled garden they have set up for you. If you can’t convince them to give you complete access to the data they manage for you, AI fortunately allows you to migrate quite fast. Once you’ve got hold of your data, you’ll need to manage how AI systems can access this data on behalf of human users, because you don’t always want Bob to see what Alice is doing in your company. That’s hard and merciless, since AI models are great at finding need-to-know errors. It takes systems that check hard access rules and models that check soft access rules. Now comes the most important part. You need to set up your own continuous training flywheel, so that you can improve your AI systems based on their interaction with your employees and your users. This is how you turn the edges of your business into AI systems your vendors and competitors cannot replicate. It’s also how you reduce deployment cost as well, as you can shrink models according to model input distribution. Those bills are getting substantial, we need to collectively become efficient if we want AI development to continue, so that matters. All of these efforts might seem daunting – they are. This is both a complete replatforming of your IT, and a complete change in the way you’re developing software, and operating your business. AI lifecycle management requires understanding human behavior and gradient descent, that’s a stretch. At Mistral, we facilitate that work by providing all primitives that you need in a single control plane, Studio, and a training platform, Forge. With our applied AI engineers and scientists working hand-in-hand with our customers, we ensure that we transfer knowledge, and that we can disappear once the systems are up and running. We deploy on our customers' infrastructure, or through our zero-data-retention hosted services, so that your edges remain your edges, and the switch button can be fully in your hand. Frontier AI can accelerate the growth of your business, but if it’s not in your hands, it’s not going to be your growth.

🚨 Zero Trust for AI Agents Anthropic just released "Zero Trust for AI Agents." As we're thinking about agentic permissions, applying a Zero Trust discipline is critical to secure adoption. AI agents interpret goals, call tools, chain actions, delegate to other agents, and maintain context across sessions. The trust surface is different. The paper introduces "least agency" — a concept that OWASP has been promoting — and the distinction from least privilege is worth sitting with. 👉 "Least privilege" asks what an identity can access. 👉 "Least agency" asks what an agent can do, under what conditions, with which tools, and with what level of oversight. Autonomous agents introduce action risk alongside access risk — and the boundaries around behavior need to be architecturally enforced, not assumed. The paper includes a design test worth writing down: 🔥 Does the control make the attack impossible, or merely tedious?🔥 The practical controls follow directly from Zero Trust fundamentals — cryptographic agent identity, short-lived credentials, tool allow-listing, sandboxed execution, and full traceability from prompt to action to outcome. None of this is new doctrine. It's existing architecture applied to a harder problem. Full disclosure: the paper cites NIST SP 800-207 on Zero Trust Architecture and the CISA Zero Trust Maturity Model, both of which I co-authored during my time supporting Federal Zero Trust efforts at CISA. Zero Trust is built for a world where we have to remove implicit trust. Agentic AI is the next version of that same problem — valid identities, valid credentials, legitimate-looking actions, and still no basis for assumed trust. Access is earned. Actions are constrained. Agency must be governed. 👉🏼 Link to Anthropic's paper in the comments.

There is a lot being written about the stylistic tells of AI writing (em-dashes, "doing the heavy lifting," etc.) but this paper looks at AI narrative tells instead, analyzing over 50,000 pieces of AI story writing compared to human short stories. There are fascinating differences between AI & human narrative (including, interestingly, more sensory descriptions) and asking AI to write in different styles doesn't do much to change the themes. Paper: https://lnkd.in/erS5-Z5f

One of the things that has struck me in recent conversations with British businesses is how often AI projects stall for reasons that have nothing to do with the technology itself. A new report from the Stanford Digital Economy Lab reaches a similar conclusion. They looked at 51 companies that had deployed AI and, in most cases, the problems came from having to change internal processes, sort out data and generally just get the organisation to operate differently. It also explains why outcomes vary so much. Two companies can try to do the same thing with AI and end up in completely different places. One moves quickly, the other gets stuck. The difference tends to come down to leadership and how the organisation is set up to adopt it, which reflects what I’ve seen more broadly on the ground here in the UK. The businesses making the most progress are not necessarily the ones with access to better technology, but the ones prepared to adapt how they operate around it. That should give some confidence. This isn't about having access to something others don’t - it's simply about how you use it. If more businesses are able to make that shift, there's a real opportunity here for British companies to move faster and compete more effectively than many may expect.

My 6 year old son Rohan asked me a simple question. Before I responded, he opened his IPAD and said, "Hey ChatGPT, why is the sky blue?" He got a great answer in three seconds. Clear, accurate, age-appropriate. I then asked: "So why IS the sky blue?" He responded immediately with "I don't know. ChatGPT told me." He got the answer, but never learned it. This is what I call Cognitive Atrophy and it's not just an issue with our kids. Working with an executive last month, I noticed he had his team use AI to create the first draft of every board memo. The decks came up as flawless. During the board member, a question was asked about specific recommendation and whether it included an often underrepresented customer. Crickets were in the room. In that silence, they all realized that nobody in the room could defend the recommendation because nobody had actually thought through the reasoning. Here's what I believe: If you can't explain it without the tool, you don't know it. Period. The struggle IS the learning. The effort IS the point. We don't hand a toddler an electric scooter before they learned to walk. Their legs wouldn't learn how to move. They'd atrophy. So why are we handing our teams cognitive scooters before they've built thinking legs? What's one thing you've stopped thinking through because a tool does it for you? #DigitalBodyLanguage #AIThinking #Leadership #FutureOfWork #CognitiveAtrophy

AI security/securing the use of AI is going to kill me. I use Claude Code almost daily. It's a problem.... Here's what I have to change AGAIN this week. Security researcher Ari Marzuk disclosed 30+ vulnerabilities across AI coding tools. Cursor. GitHub Copilot. Windsurf. Claude Code. All of them. He called it IDEsaster. The attack chain includes prompt injection, hijacking LLM context, and auto-approved tool calls executing without permission. Then, legitimate IDE features are weaponized for data exfiltration and RCE. Your .env files. Your API keys. Your source code. Accessible through features you thought were safe. Most studies I read claim that around 85% of developers now use AI coding tools daily. Most have no idea their IDE treats its own features as inherently trusted. 𝗦𝗼... 𝗮𝗳𝘁𝗲𝗿 𝗿𝗲𝘃𝗶𝗲𝘄𝗶𝗻𝗴 𝗔𝗿𝗶'𝘀 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵, 𝗵𝗲𝗿𝗲'𝘀 𝗜 𝘄𝗶𝗹𝗹 𝗯𝗲 𝗱𝗼𝗶𝗻𝗴... Be warned: All this is SO much easier said than done! Audit every MCP server connection. Checked for tool poisoning vectors where legitimate tools might parse attacker-controlled input from GitHub PRs or web content. Removed servers I couldn't verify. Disabled auto-approve for file writes. The attack chains weaponize configuration files and project instructions like .claude/settings.json and CLAUDE.md. One malicious write to these files can alter agent behavior or achieve code execution without additional user interaction. Move all credentials to a secrets manager. No .gitignored .env files in agent-accessible directories. API keys live in 1Password CLI. Environment variables inject at runtime through a wrapper script the LLM never sees. Start running Claude Code in isolated containers. Mounted volumes limited to specific project directories. No access to ~/.ssh, ~/.aws, or ~/.config. If the agent gets compromised, blast radius stays contained. Enable all security warnings. Claude Code added explicit warnings for JSON schema exfiltration and settings file modifications. These exist because Anthropic knows the attack surface. Add pre-commit hooks for hidden characters. Prompt injections hide in pasted URLs, READMEs, and file names using invisible Unicode. Flag non-ASCII characters in any file the agent might ingest. The fix isn't to stop using AI coding tools. The fix is to stop trusting them implicitly. What controls do you have for AI tools with write access to your codebase? 👉 Follow for more AI and cybersecurity insights with the occasional rant #AISecurity #DevSecOps

Leadership across companies expects AI to magically cut time-to-ship in half. Yes, AI can help engineers code faster. But coding was never the bottleneck. The real drag is non-tech stuff - planning, periodic status updates, cross-team coordination, unclear requirements, stakeholder alignment, reviews, approvals, handoffs, deployment cycles, and on-call rotations. Cumulatively, that's ~80% of the work. Optimizing the remaining ~20% cannot produce a 50% reduction. If leaders really want to see gains, they should focus less on coding velocity and more on eliminating process friction.

My biggest takeaway from Chip Huyen: 1. Most AI product problems aren’t AI problems. When companies think they have an AI performance issue, it’s usually a user experience problem, an organizational communication gap, or a data quality issue. One company thought their AI lead scoring system was broken, but the real issue was that the marketing team wasn’t asking the right questions to get useful data. 2. Your best performers benefit most from AI tools. In a controlled experiment, the highest-performing engineers got the biggest productivity boost from AI coding assistants, not the lowest performers. Senior engineers who already knew how to solve problems used AI to work even faster, while low performers often just copied and pasted code they didn’t understand. 3. How you prepare your data matters more than which database you choose. Companies see their biggest AI performance gains from better organizing and preparing their information—breaking content into the right size chunks, adding summaries, converting content into question-and-answer format—rather than agonizing over which technical infrastructure to use. 4. The biggest improvements to your AI product come from talking to users and understanding their feedback, not from adopting the latest models or staying glued to AI news. Many companies waste time debating which technology to use, when the real wins come from better user experience and data preparation. 5. Fine-tuning should be your last resort. Before investing in fine-tuning a model, try simpler solutions first: improve your prompts, add basic post-processing scripts, or fix your data pipeline. One company caught 90% of its model’s mistakes with a simple script. Fine-tuning creates ongoing maintenance headaches and should only be used when everything else has been maxed out. 6. You don’t need to be perfect to win. Many successful companies choose “good enough” over perfect when implementing AI systems. They calculate whether investing two engineers to improve accuracy from 80% to 85% is better than using those same engineers to launch an entirely new feature. Often, the new feature provides more value. 7. AI productivity is nearly impossible to measure. Companies invest heavily in AI coding tools but can’t clearly prove they work. When forced to choose between expensive AI subscriptions for their team or hiring one additional person, many managers choose the person, not necessarily because AI doesn’t help but because headcount feels more tangible. 8. Many people don’t know what to build despite having powerful tools. Even with AI tools that can build almost anything, many employees face an “idea crisis”—they simply don’t know what to create. The best approach: spend a week noticing what frustrates you in your daily work, then build small tools to solve those specific pain points.