Most companies already have AI wins. A sales lead preps for calls faster. A project manager builds a tool over a weekend. The problem is that those wins tend to stay with the person who found them.

Our co-founder Mike Gadsby opened 1682, our annual business of innovation conference, with a way to think about closing that gap. AI capability tends to grow in three stages: personal experimentation, then team process, then internal products the whole company relies on. Most organizations have plenty of the first stage. Moving into the second and third is where things stall.

Speakers from Wawa, HubSpot, ServiceNow, Guru, and others spent the rest of the day showing what it takes to make that jump. Here are five moves that came out of it.

1. Fix the process before you pick the tool

“AI doesn’t need a use case, it needs a process.” That’s how Danny Fisher of West Shore Home put it. His company grew from an $18 million business to a run rate of about $1 billion by getting its operations clean and consistent first. AI came after.

Wawa took a similar path. Its customer experience team mapped more than 550 friction points, then found that most traced back to four themes. That pattern now guides an AI tool that helps the team decide what to fix first.

Marjón Dean of ServiceNow called the alternative “organizational debt.” Years of layered processes surface fast once AI touches them. If your team can’t describe how a workflow runs today, automating it will only make the mess move faster.

2. Let the people closest to the work build

The best ideas for fixing a broken experience usually come from the people who live with it every day. That’s long been a core principle of good innovation and experience design. What’s changed is that AI-enabled tools now let those same people help build the fix too.

Brett Norton, president of Buffalo Construction, saw this firsthand. When his team took on project closeouts, the field teams who handled them every day helped shape the solution, and closeouts dropped from as long as 120 days to as few as 25. Buffalo’s AI Council keeps that going, bringing problems up from the field and sending solutions back. Brett’s rule of thumb is a good one to borrow: if you’ve done a task two to four times, automate it.

We’ve seen the same thing inside O3XO. Our project managers built and maintain their own planning tool, and updates that used to take months now happen during the meeting where someone asks for them.

3. Give AI the one thing it doesn’t have: your company

Today’s models can handle nearly any task. What they don’t know is your business. Rick Nucci of Guru made the case that the lasting problem in enterprise AI is “accuracy. Not access.” With new models shipping every few weeks, building around a model’s current weaknesses is a short-lived bet. Building around trusted company knowledge lasts.

That’s part of a broader shift. More companies are moving toward models they own and train on their own data for specific tasks, rather than relying only on general-purpose tools. West Shore Home is a good example. Its design consultants take thousands of bathroom photos a day, and the company built its own smaller image models to read them, down to the location of the drain. Danny Fisher explained that the big frontier models didn’t understand why a drain’s location changes the cost and complexity of an install. A model trained on West Shore’s own work did.

The same logic applies outside your walls. Christina Clark of HubSpot shared at 1682 that 42% of HubSpot’s prospects used AI search to evaluate the company. By making its content the answer AI tools cite, HubSpot saw up to 1,850% more AI-referred qualified leads.

4. Expect sprawl, then consolidate

When building gets cheap, the new risk is too much of it. Lazar Jovanovic of Lovable shared the story of a 250-person company that gave every employee a license, built 295 apps in six months, and found it needed just 9.

Experimentation is healthy. Getting value from it means setting up the organization to mature it. Ethan Mollick of Wharton described three pieces that work together: leadership that sets direction and incentives, a crowd of employees experimenting with good tools, and an internal lab that tests what works and feeds it back.

Lazar added the role that ties it together: a designated builder who owns what gets made and acts as an AI transformation lead. At Buffalo Construction, an AI Council fills a similar role, keeping a roadmap to consolidate tools so data doesn’t end up in silos.

If you’re just starting, Rick’s advice is to keep it small: pick one workflow, one outcome, and one team.

5. Manage the loop, and bring your people with you

As agents take on more work, no one can review every step. Ethan put it plainly: “I don’t think you could be in the loop anymore. I think you could manage the loop.” People define what good looks like, judge the results, and own them.

That shift needs governance that keeps up. In his keynote, Mike named three priorities: an AI policy that works as a living system, a clear record of what AI can access and do, and testing that grows as solutions get more complex.

It also needs patience with the people doing the work. Marjón described the whiplash employees feel when new tools, new processes, and new roles all land at once. ServiceNow now shares a three-month outlook every quarter so change doesn’t arrive as a surprise. As she put it, trust is the only currency in change.

Where to start

Individual AI wins are proof that the tools work. Getting from personal experimentation to team process and internal products is a different job, and most of it has little to do with the tools. Pick one workflow where your team already sees results, and ask what it would take for the whole company to work that way.

That’s the work we do at O3XO. We help teams find where AI can make the biggest difference, then build systems they actually adopt. If you’re somewhere between individual wins and company-wide adoption, let’s talk.

For the full rundown of every session, read O3's 1682 recap.