There Is No AI Strategy

The vendors are selling tools. The real work is deciding what to stop doing.

In 2010, the company I worked for hired a Chief Mobility Officer.

An actual title. On an actual org chart.

We talked about “going mobile” like it was a place you could drive to. Eventually the useful conversation showed up: what could the business do differently now that the technology existed? I couldn’t tell you what was in our mobile strategy. I can tell you which teams changed how they worked.

I thought about that title a lot walking the floor at the MBA Annual Conference. Booth after booth. Screens running demos, another demo a few feet away, then another beyond that. Established vendors next to new ones, all competing for a few minutes of your attention. And on almost every booth, the same words: AI. Agents. Agentic.

I’ve been the vendor. I know how much work goes into building something useful and getting someone to stop long enough to see it. Walking that floor, I could also see how hard we’ve made the buyer’s job. If you run a lending business, where do you even start? Which products solve a problem you actually have? Which ones will work with the systems you’ve already paid for? Which ones can you trust?

“What’s our AI strategy?” is a fair question to bring home. You’re probably hearing it from multiple angles, and consistently from your board. It just gets harder to answer when you start with everything there is to buy (or arguably even build).

If you’ve run a mortgage business for any length of time, you’ve already bought several versions of the future. You paid for the implementations, asked your people to change, and waited for the savings. Sometimes they arrived. Sometimes the new system became one more thing somebody had to manage. And pay for.

AI gives us real reason to be ambitious. It can read documents, prepare analysis, write software, and carry out whole sequences of tasks. In many ways, it sounds like the holy grail that paper-heavy industries like mortgage have been waiting decades for. Which is why buying a tool and leaving the business unchanged is such a small response, and it often blinds us to the real opportunity at hand.

What’s Underneath the Software?

I’ve built two technology businesses. A good product can solve a real problem. What a product can’t do is make the management decisions around it. Which work should stop? Who owns the outcome? What should a role become when a big chunk of its current work isn’t needed anymore?

Call those decisions an “AI strategy” and it becomes surprisingly easy to hand them to someone else. Maybe it’s your IT department, maybe it’s a newly minted “Chief AI Officer.” There’s a pattern we’ve followed for decades.  For leaders who delegate AI elsewhere in the organization, the conversation turns to platforms, pilots, and whose budget it comes from. The existing business becomes a container, and we go looking for places to put AI inside it.

What doesn’t change is the worrying part. The job descriptions stay. The handoffs stay. The workarounds you accumulated over the last 20 years are buried deep within your workflows. The manager who approves everything still approves everything. We give people tools to move faster through a process nobody has questioned.

Saving someone an hour of tedious work is worth doing. But look at what MBA’s own numbers say about the business. In the second quarter of 2026, independent mortgage banks spent $10,936 to produce each loan and kept $973 of profit from it. When a loan costs more than 11x what it earns, an hour here and there doesn’t change the economics. You have to look at the work that made the hour necessary.

A file came back with a defect, so someone added another review. A branch needed an exception, so someone built a workaround. The LOS couldn’t talk to another system, so a person became the connection between them.

Each decision made sense at the time, but workarounds rarely come with an expiry date.  They last for generations.

Years later, the people, the systems, even the original problem may have changed. The workaround keeps going. People get hired to carry it out. Eventually, questioning the step sounds like questioning the person whose job it is.

I wrote about this in The 200-Email Job: an inbox full of chasing starts to look like the job itself. I get the temptation to make the emails faster. It’s a much easier conversation than asking why there are so many.

What Does a Closing Actually Wait On?

I run a company that handles real estate closings. For years I treated the back-and-forth between lawyers, lenders, agents, and clients as part of the job. Documents moving. Details being confirmed. Everyone waiting on somebody else.

Follow any closing step by step and you see how much of the work exists to get information into a form the next person can act on. AI can help with that. But writing the next email faster leaves the reason for the email intact. The better question is whether the next person could get a complete, usable picture without the chasing.

Once you ask that, you’re redesigning a process. Sooner or later, you’re redesigning a role.

What Happens When You Take an Underwriter’s Job Apart?

An underwriter’s title bundles very different work: organizing documents, calculating income, checking guidelines, writing conditions, weighing compensating factors, making the credit decision. Hand that person an AI license and the bundle stays intact.

When we took the bundle apart, document indexing was an easy candidate for automation. Income analysis and the credit memo could start from AI-prepared work that the underwriter checks against the source documents. Weighing compensating factors, calling the loan officer about a borderline file, owning the credit decision: those stayed human.

Then came the part most business cases skip. AI adds work.

Someone has to check the analysis, chase down discrepancies, and explain the work to quality control. That takes time, and it needs a place in the job description. Leave it out and you’ve overstated the savings before you’ve started.

A business case that leaves out the review is a sales deck.

For a lender, the real question is whether the whole file moves through with less effort and acceptable quality. Minutes saved on prep can disappear into corrections, extra review, or another queue. And recovered time isn’t automatically lower cost. You still have to decide whether it becomes capacity, a smaller team, or better service.

The goal is to put more of an experienced underwriter’s attention on the files that need judgment. That means changing training, decision authority, and how performance gets measured to fit the role you’ve actually designed.

Buying the software doesn’t answer any of that. Neither does telling your team to “embrace AI.”

Where Do You Start?

Pick a recurring problem whose output your team already knows how to judge. Underwriting is one. Prepping a closing package, checking a file for missing documents, or drafting an internal status summary are others. Agree on the result before you buy anything, and measure the current process, waiting included, so you have something honest to compare against.

Bring compliance and risk in while the workflow is still on the whiteboard. A step that looks redundant to operations may be doing an important job. Keep its purpose, even if you change how it gets done.

And be specific about the human. A reviewer needs the source material, time to check it, and the authority to reject the output.

“Human in the loop” doesn’t tell you whether any of that is happening. Done badly, it creates additional risk.

The same goes for timelines. A draft someone checks before use is a very different build from a system that writes into the LOS or talks to a borrower. A working demo settles none of that.

Why Does an Executive Have to Get Hands-On with AI?

Business-first thinking gets sharper once you’ve seen what the tools can actually do. Otherwise you’re redesigning around yesterday’s limits, or somebody else’s demo.

I know that’s a hard ask of senior people. You’ve spent years becoming the person with answers. Now you’re supposed to be a beginner in front of the people who look to you for direction.

“The AI team handles that” is a comfortable sentence, which is why it’s so easy to fall into. A briefing won’t give you a feel for the work you’re asking people to change.

You can’t learn to swim from the pool deck. Start with a task you understand, in an approved environment, with data you’re allowed to use. Sit with someone who knows the tool. Stay long enough to see what it misses and how much context you have to feed it. You don’t need an audience for your first try.

It’s easier with people who aren’t on your payroll. That’s why I started the AI Lending Lab: executives from different lenders spend a day building with AI on a real lending problem, with a developer at the table so nobody has to become one. Being a beginner in front of a peer from another lender turns out to be very different from being one in front of your own team. Nobody has a reputation to protect, so people ask the basic question out loud and we learn as much from each other as about AI.

At one Lab, an executive brought her sales pipeline. By the end of the day she’d worked out who her top performers were, what they did differently, and a training plan to spread it across the rest of the team. Then she did the math out loud. That work, she said, would have taken 6 months and $300,000. She wasn’t leaving with a finished tool. She was leaving with a clear picture of how she’d actually do it.

That’s the shift. People walk in wondering what AI can do and walk out knowing what they’d do with it, fluent enough to lead it instead of approving it.  You don’t need to become a developer. You need enough fluency to lead the conversation about what your business could become.

What Happens After the Leadership Meeting?

Rich Weidel describes the management side in his recent blog post, Technology Is Not Culture, and I appreciate how unglamorous his account is: making costs visible, confronting differences in how work gets done, changing who’s accountable. In his telling, that work came before the decision to build their own software.

A common process means somebody gives up their preferred way of working. Removing an approval changes a manager’s authority. Sharing information shrinks the importance of the person everyone used to go through.  Those aren’t technology decisions, but they are core to getting traction with a powerful technology like AI.  You can agree to redesign in a leadership meeting and undo it with the first exception you approve afterward.

People also deserve an honest account of what this means for them. Some work will disappear. Some roles will change a lot. Vague promises about “higher-value work” don’t help the person wondering what they’ll be doing on Monday.  Instead, they push people to hoard knowledge and information as job protection.

So it’s not an AI conversation.  We need to be clear on what we want more of. The loan officer who hears hesitation in a borrower’s voice and knows it needs a conversation? The underwriter who knows which question hasn’t been asked? The manager who uses a difficult file to teach a junior colleague how to think through the next one?

One Loan File

Start with one process that matters. Map the work and the waiting with the people who actually do it. Ask what each step accomplishes and whether that reason still holds. Keep the controls that matter. Remove the steps whose reasons have expired.

Then rewrite the roles around what’s left. Decide who owns the result, what they’re allowed to decide, and how you’ll know the new way is better.

There is no AI strategy. There’s your business, and whether you’re willing to change it.

Before the next vendor demo, pick one loan file and follow it through the business. Where did it wait? Who had to chase? Which checks caught something that mattered?

Then ask which parts of that journey you’d design the same way today.

Reuven Gorsht is the co-founder of The Variable, an AI advisory firm for regulated industries, and CEO of Deeded, a virtual real estate closing platform. He helps leaders in lending and financial services build the AI fluency to lead change. He writes The Human Variable on Substack.

Source for cost/profit figures: MBA Quarterly Mortgage Bankers Performance Report, Q2 2026, via HousingWire (Aug 18, 2026): production expenses $10,936 per loan; net production profit $973 per loan.