Eight mistakes to avoid before you start

The numbers on AI adoption are brutal and worth sitting with. The MIT NANDA study found in 2025 that 95 percent of enterprise AI pilots showed no measurable P&L impact. Not 95 percent of bad ideas. Ninety-five percent of pilots that were funded, staffed, and backed by executives.

The technology is not the variable. The same models are available to the 5 percent who win and the 95 percent who stall. What separates them is a set of decisions made early, and the eight below are where it goes wrong most reliably.

1. Buying tools before doing the groundwork

This is the big one, and it is why the landing page of this site leads with it. A company buys nine AI subscriptions, wires none of them to its actual knowledge, and is surprised when the agents invent answers. Tools are the easy part. The procedures in people's heads, the documents scattered across inboxes, the brand voice nobody wrote down: that is the part that makes AI work, and it is the part that gets skipped because it is unglamorous and there is nothing to purchase.

The fix: do the free, boring work first. Write down the procedures, organise the documents, get your knowledge into a shape an agent can read. We go deep on exactly how in Data as the groundwork.

2. Believing the low rungs are cheap

The second mistake is a cost model built on the consumer price you saw first. A single ChatGPT seat is about $20 a month, so "give everyone AI" feels almost free. Then the bill arrives: Microsoft 365 Copilot is a $30 per user add-on on top of a licence you already pay for, and for a team of 40 that is $1,200 a month before anything is connected. A plan that budgeted "around 100 a month" was never real.

The fix: budget for the real thing, seats plus the orchestration and groundwork that the higher modes need. The honest cost is knowable up front. We map it across the four modes, where the spend actually lives.

3. Chasing the highest mode for its own sake

More autonomy looks like more progress, so teams reach for "agents run everything" before the work is ready for it, and before they have asked whether it should. In 2025 Klarna replaced about 700 customer service roles with an AI assistant, then walked it back, with its CEO admitting the company "went too far" as quality and customer trust dropped. They rebuilt a hybrid model: AI on the routine, people on the complex and the emotional. The lesson is not that the AI failed. It is that some work should not run unattended, and pushing it there is a choice that costs you.

The fix: match the autonomy to the work. Let AI run the routine on its own, keep a person on the calls that carry judgment, nuance, or real risk, and decide where that line sits on purpose rather than by default.

4. Automating everything you technically can

The previous mistake is about how far to let AI run; this one is about how much to hand it at all. Once a team sees what AI can do, the instinct is to point it at everything. But "can" and "should" are different questions, and the gap between them is where money quietly leaks. Every automation is now a thing you maintain: it breaks on the edge cases, it drifts when the inputs change, and a process that runs twice a month rarely earns back the weeks spent automating it. Automate a broken process and all you have done is make the mess run faster.

The fix: treat each automation as an investment that has to pay back, not a box to tick. Wire up the high-volume, stable, well-understood work first, fix the process before you automate it, and leave the rare or messy tasks to people, who handle exceptions for nothing. A handful of automations that hold beats a wall of them that need babysitting.

5. Locking tools and processes in too early

AI moves faster than any procurement cycle. The best model this quarter is rarely the best next quarter, and a stack hard-wired to one vendor in early 2025 looked dated by the end of it. Teams that bet everything on a single platform, pour all their data into its format, and rebuild every workflow around its quirks pay twice: once to build, and again to escape when something better arrives or the price moves. The early commitment that felt efficient becomes the thing you cannot afford to undo.

The fix: plan the sequence, but keep your options open inside it. Hold your data and your prompts in a portable shape, prefer tools you can swap without a rebuild, and make the first steps reversible. Commit to the direction, not to a vendor you picked before you understood the work.

6. Treating AI as a project, not a capability

A pilot launches, a deadline is hit, a box is ticked, and then nothing compounds, because no one owns it and no one is measuring it. The MIT study's central finding was that the failing systems did not retain feedback, adapt to context, or improve over time. A project ends. A capability gets better. Most companies fund the first and hope for the second.

The fix: give it an owner, measure the gain in real numbers, and build the organisational memory that lets the system learn. If you cannot say what a tool improved this quarter, it has already joined the 95 percent.

7. Building it in a silo

A team builds a clever pilot off to the side, disconnected from the systems where the real work happens: the CRM, the order data, the documents, the tools people work in every day. It demos well, then stalls. The pilot cannot read the live data it needs or write its results back into those tools, so using it means copying things by hand between an AI in one window and the actual job in another. AI that runs beside your company instead of inside it stays a clever demo and never becomes part of how the work gets done.

The fix: design for the whole company, not one corner of it. Before you build, map where the data lives and where the output has to land, then pick tools that connect to what you already run and let the new system read from and write back to the rest of your stack. Get the systems talking to each other first, and a pilot becomes something the whole company can actually use.

8. Forgetting the team has to be ready and willing

Choosing a tool with people in the room is not the same as preparing them to use it. A tool nobody uses returns nothing, however good it is. Roll an assistant out to people who were not trained on it, given no time to learn it, and quietly afraid it was bought to replace them, and it will sit idle while everyone keeps working the old way. The hardest part of AI is rarely the model; it is the people who have to change how they work, and whether they believe the change is for them.

The fix: treat readiness as part of the rollout, not an afterthought. Say plainly how roles will change, back the tools with real training and real time to practise, and prove the value on a task people feel in their own day. People take up what makes their work better and resist what is done to them.

The thread through all eight

Notice what these mistakes share. None of them is a technology problem, and none of them is fixed by a better model. They are decisions about sequence, cost, restraint, timing, ownership, and people, all made before you buy anything. That is the good news. The cheapest moment to avoid every one of them is right now, before the money is spent, by being honest about where you stand and what the work actually requires.