IMPACTME.AI

← All postsAI · Business · Strategy

The AI Onboarding Gap

By Adam Murphy — impactme.ai

Do we expect too much from AI? Or is our expectation right about the destination and wrong about the distance?

I've been working on a couple of healthcare AI projects recently, and I keep seeing a tension that I think is present across the industry. The people using these systems aren't wrong about what AI should be able to do. They can see that AI understands language. It can listen to a clinician describe a patient visit, organize the key information, write a clean note, and help place it in the correct form. In one workflow, we've reduced what used to take roughly 25 minutes to around 8.

So the natural question is: if AI can do all of that, why can't it just do the whole thing?

I understand the expectation because I share it. I believe AI is going to systematize, automate, and help make decisions in parts of our businesses that we haven't even considered yet. Greater intelligence will help, but many of the problems I am encountering today come from the distance between having access to intelligence and turning it into a system a business can depend on.

That distance can be hard to see. I once had a job that was only six miles from my home, but the trip could easily take more than 30 minutes. The mileage told you almost nothing about the route or the traffic. Our expectations for AI are often similar. We may be looking in the right direction while badly underestimating the road between here and there.

A Forest Is Not a House

Imagine I give you a forest, a box of nails, and every tool in Home Depot.

I can truthfully tell you that what you need to build a house is sitting in front of you. The wood is there. The tools are there. The fasteners are there. Almost any kind of house you can imagine is possible.

I still cannot tell you that your house is almost finished. The trees need to become lumber. Someone needs to decide what is being built. We need measurements, plans, plumbing, wiring, permits, inspections, and hundreds of small decisions that nobody notices when they walk into a completed home. Everything required may technically be present, while most of the work that turns those materials into a usable house is still ahead of us.

AI is a little like that forest. Maybe it is the forest, the tools, and a very capable builder all wrapped together. It changes what is possible and can dramatically reduce the labor required. The builder still needs to understand what you are trying to build, why you are building it, and what will make the finished result acceptable to you. Many AI projects begin to struggle inside those unanswered questions.

Your Business Is More Complicated Than You Think

Most businesses believe they have trained their employees and that their processes are clear and repeatable. The procedures are written down, and many of them have been reviewed or audited.

In the day-to-day environment we call business, they usually have the outline of a process and several people who know how to make it work.

The actual process lives in hundreds of little judgments:

  • If this field is empty, check that screen.
  • If the patient says this, phrase the note differently.
  • If these two numbers disagree, trust the second one unless the visit happened before Tuesday.
  • This person can approve it, but that person can only review it.
  • We normally do it this way, except for these three customers. They are too large to follow that process, so we handle them differently.
  • Nobody wrote this down, but Susan knows exactly what to do when this happens.

We have all known Susan in the workplace. Maybe she built the process. Maybe she knows the customers better than anyone else. Maybe she is the person who makes the final call, so everyone has learned to do it the way she expects. When something unusual happens, you let Susan dial it in.

Then AI arrives, and we expect it to understand the process because it understands the words. It can see the screen and move the mouse, so it appears to have all the pieces. What it lacks is the history that gives those pieces meaning: where the reliable information lives, which exceptions matter, what the business considers successful, and which mistake would be mildly annoying versus legally catastrophic.

Your AI may be brilliant, but it is also the newest employee in the company. Throwing every document at it is not much better than doing that to a new hire. The documents contain data and formal processes. They rarely contain all the reasoning, expectations, preferences, relationships, or criteria that determine what people actually do.

We Would Never Train a Person This Way

Imagine hiring the smartest person you have ever met. On their first day, you sit them in front of a computer and say, "You are extremely intelligent. Run this department for us."

They ask where the files are, and we tell them to check SharePoint. They ask which customers require special handling, and we tell them to look through the customer history. Then they ask who approves a refund, what happens when information is missing, and whether the spreadsheet or the CRM should be treated as correct.

Eventually, we would have to acknowledge that intelligence and institutional knowledge are different things. An intelligent employee can learn the business, ask better questions, and perhaps see improvements that everyone else missed. They still need access to the history and relationships that explain why the company works the way it does.

We often skip that distinction with AI. We show people a demonstration where the model reads a clean document, receives clear instructions, and produces a beautiful result. Then we place it inside a real business where the document is sideways, the customer changed their mind twice, the data conflicts, the password expired, and the person who understands the exception is on vacation.

The demonstration showed the base path, where the inputs and rules made sense. In real life, the SharePoint site is actually several sites. Each contains data and timestamps that do not quite agree. The customer exception list is a database log showing that employee number 15, Susan, changed the standard price from $49 to $44 for customer number 3 on August 1, with a note that says, "I adjusted this for Tim."

The AI has no idea who Tim is. Tim might be the customer, the owner, an employee, or a friend who said the customer would buy more if the price came down.

Many, and maybe most, businesses live in exceptions that have no clean path from the written rule to the expected behavior.

The Last 20 Percent Is Where the System Lives

There is an old saying in software development, and in many other fields, that the final 20 percent of a project takes 80 percent of the time. You can probably think of examples in your own area of expertise. With AI, we do not always see that final stretch clearly because getting the first result can happen so quickly.

Getting AI to perform a task once is often surprisingly easy. Getting it to perform that task reliably across different users, incomplete information, changing conditions, security restrictions, and real consequences requires a different level of work.

That final stretch can include:

  • Connecting the correct data in the right order.
  • Defining permissions, approval points, and when a person should take over.
  • Teaching the system what good work looks like.
  • Handling missing information, contradictions, edge cases, and preferences.
  • Measuring whether the output is actually improving.
  • Recording what happened and providing a way to recover.
  • Updating the system as the business changes.

Looking down that list, it becomes easier to understand the mismatch. From the outside, it looks like the AI knows how to do the work. It can explain the rules and may understand every individual step. What it has not learned is how those details came together to become your process.

This onboarding gap shows up in the numbers. In a 2026 BCG survey, 64% of CEOs said their companies pursued AI pilots, but only 26% had embedded AI in a broader business transformation. Access to capable AI and measurable organizational value remain farther apart than most headlines suggest. (BCG, 2026)

When 95 Percent Still Saves Nothing

The gap becomes especially important when a business tries to measure the payoff. Here is a short version of a conversation I had with a CFO several years ago:

Me: I can save Todd one or two hours every day by automating this process. This is a huge return on investment.

I was very excited. ROI is how finance people understand decisions, so I thought I had made the case.

CFO: What is Todd going to do with those two hours? Can we reassign him? Does he have critical work in his backlog? Is there a position we will no longer need?

Me: Um, well, I guess...

CFO: If we cannot put those hours into something else or remove a line-item cost, let us focus on something else.

I learned that time savings and efficiency are not always measured the way I naturally think about them. The business captures that value only if the time can be redirected toward something important or converted into a measurable financial result.

A later conversation with an inside sales manager added another part to that lesson.

Inside sales really did have a large backlog of valuable work, so a similar automation should have been a slam dunk.

Sales manager: You can gather all the information on this account, and it will be accurate and current?

Me: It is nearly perfect. About 95 percent of our testing has been correct. The problem is that information is missing in roughly 5 percent of the cases, and those can still cause problems.

Sales manager: You understand that if five out of every hundred proposals contain that type of mistake, it could be devastating. I still need someone to review every proposal to make sure those five never reach a customer.

The automation was 95 percent accurate, but it had not released the two hours. The risk contained in the remaining 5 percent meant that someone still had to review 100 percent of the work.

I think about this in the same way I think about a total eclipse. If you have only seen an eclipse at 99 percent totality, you have seen something impressive, but you have not experienced what happens when the final 1 percent disappears.

At totality, the sky goes dark. A sunset-like glow can surround the horizon. The temperature drops, and the crickets begin chirping. Looking through eclipse glasses as the sun becomes a thin glowing sliver is fascinating, but the final percentage changes the nature of the experience.

Some business processes have a similar threshold. Ninety-five percent may be a remarkable technical achievement, while 100 percent review is still required. The value appears only when the system becomes reliable enough to change what the person must do next.

AI Implementation Is Organizational Archaeology

Building AI systems has shown me that many of these are familiar business challenges in a new form. Companies have always struggled to get new employees up to speed inside environments that are complicated, inconsistent, and changing.

You begin with what everyone thinks the process is, then you watch the work happen. You ask questions, map decisions, find the data, and build the first version. Someone says, "That looks great, but sometimes we do this instead." You find the exception, revise the system, and test again.

Over time, you uncover a system that already existed but had never been completely visible. The project may reveal that two employees have performed the same process differently for years, that a critical decision has no clear owner, or that the official workflow stopped matching reality three software versions ago. AI gave the company a reason to examine the work closely enough for those differences to become visible.

This can make training AI feel slower than doing the task yourself. When you explain an exception to a person, you can watch them connect it to what they already know. When AI says, "I understand. I will not use copy and paste for this type of document," you are left wondering whether it learned a transferable principle or a sentence that applies only to the example in front of it.

The AI Industry Is Looking for the Missing Experience

Model builders increasingly recognize this problem. OpenAI's GDPval evaluation uses 1,320 tasks from 44 occupations, created and vetted by professionals averaging more than 14 years of experience — and OpenAI acknowledges that one-shot tasks cannot capture the full complexity of real work. (OpenAI GDPval)

The same shift shows up in how training data is collected. Mercor pays experienced users to record and narrate real screen workflows, explaining where AI missed context or used its tools incorrectly. Scale AI reports that nearly half of its new data-training projects now involve environments that simulate tool use and professional workflows — and distinguishes information visible in documents from institutional knowledge stored inside subject-matter experts. (Mercor software workflows; Scale AI on enterprise knowledge)

The models already contain extraordinary general knowledge. What they increasingly need is experience with the interconnected, exception-filled way people accomplish real work. Like an employee with ten years of experience, they will still need onboarding, but their questions should be better, familiar patterns should be easier to recognize, and early mistakes should become less costly.

You Only Need to Solve It Once

One of my favorite sayings when dealing with AI is, "You only need to solve this once if you do it right."

The system will still change as the business, customers, regulations, and software change. But once the workflow, intentions, exceptions, and success criteria are understood, they can become part of a reusable system instead of remaining scattered across documents and people's memories.

That captured knowledge can run at 2:00 in the morning, apply the same standard across a thousand cases, learn from corrections, and return unusual cases to a person. Building it can also give the company a clearer understanding of its own work. The leverage appears when the system becomes reliable enough for people to change what they do next.

The Questions Businesses Should Be Asking

Businesses will continue asking, "Can AI do this?" It is a reasonable place to begin, and the answer is increasingly some version of, "Yes, probably." That answer should begin the implementation conversation rather than end it.

We also need to ask whether AI understands how the business performs the work, can access the right information, knows when to ask for help, and can be measured when reality does not follow the expected path.

Like a promising new employee, AI needs context, access, expectations, examples, feedback, and time to find the unusual exceptions. The business has a responsibility to examine its processes carefully enough to teach them.

I think the future the AI industry is promising could be possible for us. Reaching it will require better models, but also something more familiar and human. Closing the onboarding gap is a team effort between the technology, the people who understand the work, and the business asking them to produce something dependable.