AI Readiness Isn't About AI

Over the past year I have spent quite a lot of time talking to companies about AI. Not so much about which model is best or what the latest version of ChatGPT can do, but about the slightly less exciting question of how companies are actually going to make all of this work.

I have had versions of this conversation with people at H&M, Volkswagen, SEB and Accenture, among others, and earlier this year I organized a small lunch in New York around AI readiness and company knowledge. We had people from organizations including the UN, Meta, Rockefeller Foundation and The Weather Channel around the table. It was supposed to be a lunch, but people ended up staying until around four in the afternoon. Clearly there was quite a bit to talk about.

What I find interesting is how quickly these conversations stop being about AI. Most larger companies already understand that AI is important. Management wants it, boards are asking about it and employees are already using it, whether there is an official company strategy or not. The difficult part starts when you take this incredibly powerful new technology and connect it to an organization that has spent the last twenty years accumulating documents, systems, processes and data that were never created with AI in mind.

There is some interesting research starting to show the same thing. McKinsey recently found that 70 percent of employees say they personally feel ready to adopt and use AI, while only 27 percent of leaders believe their organizations are ready to make the changes required for an agentic future. That is a pretty extraordinary gap. People are learning faster than the organizations around them are changing.

That, to me, is where AI readiness actually starts.

Putting AI on top of twenty years of information

Take a fairly normal company with a few thousand employees. There will be information in Microsoft 365 or Google Workspace, documents in SharePoint and Drive, conversations in Teams and Slack, an intranet, a CRM, HR systems and probably hundreds of other SaaS applications. Some of the information will be excellent and some will be completely outdated. There might be five different versions of the same document and nobody is entirely sure which one is correct. Other information technically exists somewhere, but finding it requires knowing who created it or which colleague to ask.

People who have worked in a company for years learn how to navigate this. They know that the document called “Final Strategy” probably isn't the final strategy. They know Sarah in finance has the spreadsheet everyone actually uses and that a certain section of the intranet hasn't been updated for three years. Humans compensate for messy information with experience, relationships and common sense. AI doesn't automatically have any of that context.

I had a discussion with Accenture where we ended up describing the situation in fairly simple terms. Companies basically have three choices. They can ignore AI, which I don't think is going to be a very successful strategy. They can put AI on top of everything they already have and potentially accelerate the mess. Or they can use this moment to clean up some of the underlying problems and actually prepare the company for AI.

The more companies I speak to, the more I think that third option is where a lot of the interesting work is going to happen.

Again, the external research seems to be heading in the same direction. McKinsey reported this year that more than two-thirds of high-performing companies see data as the primary obstacle to enabling AI. Their point is an important one: making corporate information searchable isn't enough. AI needs to understand which information is current, what it means, how it relates to other information and which version should actually be trusted.

This sounds obvious when you say it out loud. Yet a surprising amount of the current enterprise AI discussion seems to assume that connecting AI to more information automatically makes it smarter.

It doesn't.

AI readiness starts with some fairly boring questions

The term AI readiness makes you think about technology. Which models should we use? Should we build something ourselves? Should we buy Copilot? How should employees use ChatGPT? Do we need agents? These are obviously questions companies need to answer, but I'm not convinced they are the most important ones to start with.

I would start with something much more basic: what does the company actually know and where does that knowledge live? Which information can be trusted? Who owns it? What is outdated? What should employees have access to and what shouldn't they have access to? If there are three versions of a policy, which one should an AI use? If nobody inside the company can answer that question, it is a bit optimistic to expect the AI to figure it out.

None of this is particularly new. Companies have been talking about knowledge management, information governance and breaking down silos for as long as I have worked in software. The difference is that poor information used to be mainly an employee productivity problem. People spent too much time searching for things, asked colleagues questions they should have been able to answer themselves and occasionally made decisions using the wrong information. Annoying and expensive, yes, but companies managed to live with it.

AI changes the scale of that problem. A person might use the wrong document once. An AI system can potentially use the wrong information thousands of times. The same technology that makes good information dramatically more valuable also makes bad information more dangerous.

McKinsey makes an interesting distinction here between searchability and usability. A company might have digitized everything and made it searchable and still not have information that is genuinely ready for AI. Versioning, context, ownership and traceability suddenly matter much more because an AI system is constantly taking information apart, combining it with other information and using the result somewhere else.

You cannot simply connect an AI to twenty years of corporate information and assume that twenty years of corporate information suddenly becomes good.

From AI that answers to AI that acts

This gets considerably more interesting when we move from assistants to agents. Most enterprise AI usage so far has been fairly harmless. We ask AI to summarize something, draft an email, search for information or help create a presentation. If it gets something wrong, hopefully a human notices before anything terrible happens.

Agents are different because they can potentially act. An agent could update a CRM, create a support ticket, communicate with a customer, move information between systems, initiate an approval or perform a series of tasks without somebody clicking every button along the way. I think this is where AI becomes really powerful inside companies, but it is also where some fairly boring parts of enterprise software suddenly become extremely important.

If an AI is acting on my behalf, it needs to know not only what information exists but what I am allowed to see. It also needs to understand what I am allowed to do. If something goes wrong, the company needs to know what happened, which information was used and why an action was taken. Permissions, identity, ownership, audit trails and APIs don't sound nearly as exciting as autonomous agents, but without them I have a hard time seeing how companies can give agents meaningful freedom.

Deloitte's latest research puts some numbers behind this. Only 5 percent of organizations they surveyed believe their business processes are highly prepared for AI agents. At the same time, 74 percent of leaders expect nearly half of their business processes to be redesigned or rebuilt around agents within four years.

That is quite a gap between ambition and reality.

There is also an interesting contradiction here. The more autonomous we want AI to become, the better the underlying structure probably needs to be. Giving a human employee access to a slightly messy system is one thing. Giving thousands of automated actions access to the same mess is another.

Buying AI is easy. Changing the company is harder.

There is another part of AI readiness that I think gets too little attention. Giving employees access to AI is not the same thing as changing how a company works. If you give 5,000 people an AI assistant, some will become much more productive, some will find clever new ways to use it and some will barely touch it. That is still useful, but it is not necessarily transformation.

The more interesting exercise is to take existing work and ask whether you would design it the same way today. Why does somebody manually move information from one system into another? Why does a report take half a day to create every Friday? Why do four people approve something that perhaps only one person needs to look at? Why are five departments maintaining slightly different versions of essentially the same information?

There are thousands of these small processes inside large companies, often not because they are good processes but because they were the only practical way to do the work when they were created.

This is where I think the real productivity gains from AI will eventually come from. Making somebody 20 percent faster at producing the same report is useful. Realizing that the report doesn't need to be produced manually at all is much more interesting.

The problem, of course, is that the second option requires you to change workflows, responsibilities and sometimes the organization itself. Installing software is considerably easier.

I don't think this means companies should stop experimenting with AI until all their information is perfectly organized. That would probably take ten years and by then we will have entirely different problems to worry about. Companies should experiment aggressively, let employees learn and find the areas where AI creates real value. But at the same time they need to work on the foundation: which knowledge matters, where it lives, who owns it, which sources can be trusted and what an AI should actually be allowed to do with it.

What happens to SaaS?

Having spent much of the last 15 years building and working with SaaS companies, this is probably the part I find most interesting.

For most of the SaaS era we have built software around destinations. You open Salesforce to do one thing, go into the HR system for another, search the intranet when you need information and open the project management tool to see what your team is doing. Employees have effectively become the integration layer between all these systems. We learn where things live, switch between applications and assemble the context ourselves.

I think AI will change a lot of that.

Increasingly the starting point will be what I want to accomplish rather than which application I need to open. If I want to understand why a customer is unhappy, I don't really care whether the relevant information happens to sit in Salesforce, Slack, Zendesk, an email thread or a meeting transcript. I want the answer. Eventually I may also want the AI to suggest what should happen next and, with the right permissions, do some of it for me.

That doesn't mean SaaS disappears. Those systems still contain the workflows, business logic and data companies depend on. But it could change where much of the value sits. We have spent years obsessing over interfaces and user experience because humans had to navigate every application themselves. In a world where software is increasingly also being used by machines acting on behalf of humans, APIs, permissions, structured data and the reliability of the underlying system become much more important.

I suspect we are still very early in understanding what this means. A lot of today's software is essentially old software with an AI button added to it. That is natural; every technology shift starts by putting the new thing into the old thing. The more interesting products will probably appear when we stop doing that and start designing software around what is now possible.

So what does an AI-ready company actually look like?

I don't think anyone has the complete answer yet, certainly not me. AI is moving far too quickly for that. But after all these conversations, I am increasingly convinced that the winners won't simply be the companies that buy the best AI tools first.

Models are getting better incredibly quickly and the cost of intelligence keeps falling. Capabilities that seemed almost magical two years ago are already becoming standard features. Eventually most companies will have access to roughly the same models and many of the same tools. Having access to AI itself therefore seems unlikely to be much of a competitive advantage.

The difference will be what the AI has to work with once it enters the company.

One company might connect it to fragmented information, outdated documents, unclear permissions and twenty years of accumulated processes. Another might have clear ownership, trusted sources of knowledge, sensible access controls and workflows that have been redesigned around humans and machines working together. They could be using exactly the same underlying AI and get completely different results.

That is why I increasingly think AI readiness isn't really about AI.

AI is forcing companies to confront a lot of things they probably should have fixed anyway. The companies that use this moment to do that will not only be better prepared for today's copilots and chatbots. They will have a much stronger foundation for whatever comes next.

Further reading

McKinsey & Company — “From adoption to impact: Three horizons of AI transformation”
https://www.mckinsey.com/capabilities/people-and-organization/our-insights/from-adoption-to-impact-three-horizons-of-ai-transformation

McKinsey & Company — “AI data readiness: The key to scaling impact”
https://www.mckinsey.com/capabilities/mckinsey-technology/our-insights/ai-data-readiness-the-key-to-scaling-impact

Deloitte — “The path to agentic transformation”
https://www.deloitte.com/us/en/insights/industry/technology/path-to-agentic-transformation.html