A company’s first AI project is as much a learning experience as a technology project. Companies that have been through it report strikingly similar surprises and lessons, regardless of industry or size.
This article gathers those lessons — the ones you learn on real projects, not the ones in carefully edited success stories.
The data surprise
The first big surprise in almost every first AI project is the state of the data. The company was sure it had good data. In most cases it does — but in very different shape from what the project needs.
The data lives in four different systems with keys that don’t match. The history that looked complete has significant gaps in the most important variables. The system that supposedly logged customer returns did so differently depending on who entered them. And the data point the model depends on turns out not to be captured in any digital system — it lives on paper or in the department head’s memory.
None of this is unusual. It’s the norm. The best preparation for a first AI project is to expect the data to need work, not to expect it to be ready to use.
Defining success
The second common pattern is that nobody defined success properly at the start. “Improve forecast accuracy” is ambiguous. “Cut stockouts on A-class items at the central warehouse from 15% to under 5%” is a concrete definition of success.
Without a clear definition, three things happen: the technical team optimizes metrics that aren’t the most relevant to the business, leadership has no clear way to judge whether the project worked, and the project can end with a technically sound model nobody adopts because it doesn’t solve the problem that hurt most.
Time spent defining measurable success before development starts is always time well spent.
Underestimating integration
Building the model is the visible part of the project. Connecting it to the company’s real systems — the ERP, the CRM, the operations tools — is what turns the model into business value, and it’s also the part most often underestimated in time and complexity.
Many companies’ systems have decades of history, custom modifications, and often incomplete or outdated integration documentation. An integration that looked simple on paper can require significant analysis before anyone can build it.
The practical advice: even on the simplest project, set aside explicit time and budget for integration, and don’t estimate that time until you’ve audited the existing infrastructure.
Adoption takes longer than expected
A model in production that the team doesn’t use delivers no value. And adoption has its own process, which is rarely instant.
The first days and weeks after launch are critical. The team has questions that didn’t come up in training. The system produces an unexpected result that breeds distrust. Working with the system is a little different from working without it, and people need time to adjust.
Companies that handle this period well have active support available during the first days of real use — someone the team can ask when something doesn’t work as expected. Companies that don’t end up with teams that fall back on the manual process the moment the system gives a result they don’t understand.
What nobody measures but matters most
AI projects produce two kinds of learning: system learning (the model improves with more data and feedback) and organizational learning (the team understands how to work with AI, spots new opportunities, and builds judgment about what AI can and can’t do).
The second kind is the most valuable in the long run and the least measured. The companies making the most progress with AI are the ones whose team, after the first project, has developed an intuition for the problems AI can solve — and is already identifying the second and third projects before the first has fully stabilized.
Decisions that change the outcome
Some decisions, made before the project starts, have an outsized impact on the result:
Start with the right problem. Not the most ambitious or the most technically interesting — the one with enough data, measurable business impact, and a realistic chance of adoption.
Appoint an internal project owner. Someone inside the company with the authority to make decisions, the time to dedicate, and a stake in making it work. Without that owner, projects lose momentum at the first complication.
Set business success metrics before you start. And review them with leadership before the project moves forward, so they don’t change halfway through.
Plan post-launch from the beginning. The stabilization period after launch is part of the project, not an extra. Building it into the plan keeps the project from being “forgotten” once the external team considers its work done.
Frequently asked questions
What’s the minimum company size for an AI project to make sense? There’s no size threshold. There’s a process-volume threshold: if the process you want to automate or improve is frequent enough and has enough impact, AI can make sense regardless of company size. Companies of every size have found use cases where generative AI or automation pays off.
Is it better to start with a big, ambitious project or a small one? Experience consistently shows that a limited but well-executed project — one that delivers real, visible value in a reasonable time — has more long-term impact than a big project that takes a long time to show results. The small project builds trust in the technology, organizational learning, and leadership support for more ambitious work.
What sets apart companies that succeed with their first AI project? Three things: the problem was well defined before they started, an internal sponsor had the authority and time to drive the project, and the team was involved in the design from the start — not handed a finished system and told to use it.