Resolon Start a conversation

InsightsAI strategy

Published Updated 7 min read

A Claude power user is not an enterprise AI strategy

Protect your best people’s time. Encourage experimentation. But know when experimentation needs to become engineering.

Chad Neptune, CEO and Chief AI Architect

Contents

One of the most expensive AI mistakes I see in small and mid-sized businesses right now? Confusing a Claude power user with an enterprise AI strategy.

It usually starts with someone good

I see it almost every week. A talented employee discovers Claude. Three months later, they’re staying up until 2 a.m. building dashboards, promising to solve the company’s biggest bottlenecks, and slowly neglecting the job they were hired to do.

The early results are real. A report that took a day now takes ten minutes. A spreadsheet nobody trusted has a clean front end. People start asking for access.

Eventually, the CEO or COO calls us because it’s time to “scale” what they’ve built.

For leaders When an experiment starts asking for users, access or a budget, that is the moment to decide what it is.

Six colleagues work around a table, with a laptop and tablet showing a dashboard, notebooks and printed charts

What I find when I sit down with the team

When I sit with a team and ask which AI tools their people use, it is always personal accounts, and a lot of the time the business is paying for them. Most of it is generic prompting that nobody trusts for more than basic things, and the shadow AI essay covers how to find it. Now and then, though, someone has built one or two tools to automate a task they used to do by hand, and I have seen some of those make an impact in the business.

Across four or five engagements, not a single tool we have come across as a team has scaled without being completely rebuilt. The data underneath is usually an even bigger problem.

This isn’t a criticism of Claude. I use it extensively, I hold two of Anthropic’s Claude certifications, and it is a remarkable tool. Nor is it a criticism of the people building with it, who are often the ones who understand the work best. It’s a leadership problem: nobody decided when the experiment stopped being an experiment.

For leaders Name who decides when an experiment becomes engineering; if nobody owns that call, it gets made by default.

Having Claude is like having a workshop in your backyard

Having Claude is like having a high-performance workshop in your backyard. You can build a custom, one-off car with custom parts, and it does custom things. But if you took that car to Ford or GM, they would put it through crash-rating tests and fit seat belts and everything else a road vehicle needs.

You would never send your 16-year-old down the highway at 75 miles an hour in that car without a seat belt.

And you would never build 100 of them and sell them, because you would get sued. That is essentially what is happening right now.

In business after business, I see people experimenting with Claude, ChatGPT, Codex or Cursor who do not have the depth to put guardrails around what they build, or to bring it into production in a way anyone could replicate and understand.

For leaders Before anyone relies on a tool, ask who has crash-tested it and who answers for it.

What a prototype answers and an application has to

A prototype answers one question: can this be done? An enterprise application has to answer many more. Who is allowed to see which records? Where does the data come from, and who owns it? What happens when the AI gets an answer wrong? Who fixes it when the person who built it is on vacation, or leaves?

PrototypeEnterprise application
Who can see which recordsWhoever has the builder’s loginAccess set by role, with an owner who reviews it
Where the data comes from, and who owns itSpreadsheets exported by hand, with no named ownerThe systems of record, with agreed definitions and a named owner
What happens when the AI is wrongSomeone notices, or nobody doesA person approves the decisions that carry risk, and every action is logged
Who maintains it when the builder leavesNobody; the logic lives in their promptsA team, with documentation, tests and a backup owner
What it costs to runA personal card, or a guessA known monthly cost, with limits
Whether it passed a security reviewIt never went through oneReviewed before it touched live data
Figure 1. Prototype vs enterprise application

A tool built by one person, for one person, usually has none of those answers written down. It runs on that person’s account, reads spreadsheets they export by hand, and holds its logic in prompts only they understand. That is fine for a test. It is a risk once the business depends on it.

A working prototype isn’t an enterprise application. A prompt isn’t an architecture.

And another model release isn’t going to magically fix broken processes, poor data governance or missing infrastructure. Too many business owners are waiting for AI to become capable enough to deliver on promises that were never technically grounded in the first place.

Seven signs a prototype needs engineering

If any of these is true of a tool someone in your company built, it has outgrown the experiment.

What you will see
Other people depend on itIf it breaks on a Monday morning, work stops for someone besides the builder
It touches sensitive dataCustomer, employee, health or financial records flow through a personal account or a tool nobody has reviewed
It acts, not just answersIt sends emails, changes records or moves money, without a person reading first
Only one person knows how it worksThe logic lives in their prompts and their head, with no documentation and no backup owner
Its numbers reach decisionsFigures from it show up in a board pack, a forecast or a price, and nobody can trace how a figure was produced
Nobody knows what it costsThe bill is on a personal card, or the monthly spend on AI usage is a guess
It runs on hand-exported spreadsheetsSomeone downloads, cleans and uploads data to keep it alive
Figure 2. Seven signs a prototype needs engineering

For leaders If any sign is true, give the tool an owner, a budget and engineers.

What engineering adds

Engineering does not mean throwing the prototype away. It means putting four things under it before the business relies on it. One source of truth for the data, with agreed definitions and a named owner, which is the work of data foundations. People in control: access by role, a person approving the decisions that carry risk, and a log of who or what acted and when, which is AI governance and human control. Something that keeps running, with tests, monitoring, error handling, cost limits and a way to roll back a bad release, which is AI and automation engineering. And a plan that fits the rest of the business: what to rebuild, what to buy, what to retire, and how it connects to the systems you already run, which is AI architecture and advisory.

For leaders Budget to keep what the prototype taught you and rebuild what sits under it, not to start over.

What to do this quarter

Protect your best people’s time. Encourage experimentation. I think it is great for people to experiment in the right sandboxes, in the right environments, to prove out and pilot use cases, with approved tools, company accounts and no live customer data. Agree in advance where the line is, using the seven signs, and when a tool crosses it, make it a project with an owner, a budget and engineers. Keep the builder in the room: their prototype is often the clearest statement of what the business needs that you will ever get, written by someone who does the job. And bring the right people in before three months of productivity and thousands of dollars disappear into a dashboard that was never designed to scale.

AI should be removing your organization’s bottlenecks, not creating new ones.

Frequently asked questions

What is the difference between an AI prototype and a production application?

A prototype shows that something can be done, usually for one person on their own account. A production application is built for the business to rely on: it uses governed data, controls who can see and do what, keeps a record of what it did, is tested and monitored, and does not depend on one person to keep it running.

Is it risky when employees build their own AI tools?

Experimenting is healthy. The risk comes when a tool built outside any review starts handling sensitive data, acting on live records, or feeding decisions. This is often called shadow AI. The answer is not to ban it, but to give people approved tools and a clear point at which their work gets engineering support.

Should we ban employees from building with AI?

No. A ban pushes the work out of sight and loses the people who understand your processes best. Set rules instead: which tools are approved, which data is off limits, and when a tool has to move from experiment to engineering.

When should an AI experiment become an engineering project?

When other people depend on it, when it touches sensitive data, when it acts rather than only answers, or when its numbers reach decisions. Any one of those is enough. The earlier the move is made, the more of the prototype can be kept.

Does a mid-sized company need a formal AI strategy?

It needs clear decisions more than a long document: which problems AI should solve first, who owns the data, who approves what the AI does, and when experiments become engineering. Those decisions are an AI strategy, whatever you call them.

Share this article

New each month. Follow by RSS

Senior-led, start to finish.

A senior Resolon lead is accountable for every engagement, from the first conversation to handover, with a team built for the work. For an active engagement or a new requirement, start a conversation.