"Which of you, desiring to build a tower, does not first sit down and count the cost?" — Luke 14:28
There is a kind of software demo that should make you suspicious precisely because it is so impressive. The lights are perfect. The data is clean. The presenter types a question and the AI returns something that looks like magic. Everyone in the room nods. A pilot is approved. Budget is allocated. And then, months later, the thing quietly never ships — or ships in a hollow form that no one quite wants to talk about. If this has happened to you, you are not an outlier. You are the base case.
A widely discussed 2025 study of enterprise generative-AI initiatives found that the overwhelming majority — by its count, around 95 percent — were failing to deliver measurable impact on the bottom line. Not failing dramatically. Failing quietly: pilots that impressed in the demo and then stalled somewhere between the slide deck and the income statement. The number is debated at the edges, as all such numbers are. The pattern it points to is not.
The Gap Between the Demo and the Deployment
To understand why so many pilots die, you have to understand the distance between a demo and a deployment, because the entire sales motion is built on hiding it.
A demo is a performance in controlled conditions. The inputs are curated, the questions are anticipated, the edge cases are absent, and the failure modes are kept offstage. A deployment is the opposite: messy real data, users who do unexpected things, edge cases that arrive hourly, and a hard requirement that the system be reliable enough to trust with actual decisions. The capability that looks effortless in the demo turns out to be the easy 80 percent. The deployment lives in the remaining 20 percent — the part the demo was specifically designed to avoid showing you.
This is where the present-day reality of the technology asserts itself. These systems still hallucinate — they produce confident, fluent, completely wrong output, and they do it unpredictably enough that you can't fully automate around it. In a demo, you never see it. In production, where being wrong has consequences, it's the thing that quietly kills the project. The pilot doesn't fail because the team was incompetent. It fails because the gap between "looks amazing" and "is reliable enough to depend on" turned out to be much wider than the sale implied.
Why Vendors Prefer the Future Tense
Here is the hypocrisy, and it's a structural one rather than a matter of individual bad actors. The largest vendors have a strong incentive to keep selling you the future of AI specifically because it lets them avoid being judged on the present of AI.
The present is awkward. The present is hallucinations, integration pain, unclear ROI, and pilots that don't convert. If you evaluated these products purely on what they reliably deliver in production today, a lot of purchase decisions would look very different. So the conversation is kept relentlessly forward-looking. It's about what's coming, what's almost here, what next quarter's model will unlock, the roadmap, the trajectory, the inevitability. As long as the value is always arriving rather than arrived, the vendor never has to stand behind a present-tense claim that could be falsified by your actual results.
This is why the marketing leans so heavily on promise over performance — why one major platform was described in industry coverage as touting AI's promise over its reality. Promise is unfalsifiable. Reality has a P&L. A vendor who can keep you focused on the promise has freed himself from accountability for the reality, and collected your pilot budget in the meantime. The future tense is not enthusiasm. It is a liability shield.
Who Pays for the Pilot That Never Ships
Follow the cost, because it doesn't fall evenly. When a pilot stalls, the vendor has still won: they booked the revenue, generated a logo for their case studies, and kept you inside their ecosystem for another budget cycle. The cost lands entirely on your side — the spend, the staff time, the opportunity cost of the projects you didn't do instead, and the organizational fatigue of having chased something that didn't pay off.
And the cost is frequently larger than the sticker, because of how these deals are priced. Pilots are often seeded with generous credits that make the economics look benign. Then production usage arrives and the real bill lands — teams have reported scaling costs running five to ten times their estimates. The pilot that "didn't cost much" was a loss leader designed to commit you before the meter spun up. By the time the true cost is visible, you've already built around the thing.
How to Evaluate an AI Vendor Without Getting Sold
None of this means enterprise AI is a mirage. Real value is being delivered — by teams who evaluated rigorously and refused to buy the future tense. The discipline is straightforward, and it's mostly a matter of insisting on the present.
Make them show production, not a demo. Ask to see the system running on messy, real, comparable data — ideally yours — with the failure modes visible. Ask specifically what happens when it's wrong, how often that is, and who catches it. A vendor confident in the present tense will show you. A vendor who pivots back to the roadmap is answering the question.
Define the result before you start, in numbers. Decide what measurable outcome the pilot must hit to be judged a success — a real metric tied to your P&L, agreed up front — and hold the line on it afterward. Vague success criteria are how pilots that delivered nothing get quietly relabeled as "learnings."
Count the cost at full scale. Don't model the pilot's subsidized economics. Model what the thing costs at the volume you'd actually run it, with the overages and the per-use charges included. Build the tower's budget before you pour the foundation, not after.
Discount the future tense to zero. Evaluate every product strictly on what it reliably does today. Treat the roadmap as marketing, because that is what it is. If the present-tense product doesn't justify the purchase on its own, the promise of a better one next year shouldn't either.
The Point
The most expensive words in an enterprise software pitch are "imagine what you'll be able to do." They are expensive because they relocate the entire value of the purchase into a future the vendor is never obligated to deliver, while the cost arrives, fully present-tense, on your invoice. The pilot that never ships is not usually a story about a team that failed. It's a story about a sale that succeeded — at selling tomorrow to avoid being measured today.
Count the cost first, the old line says — before you build, not after. In the age of the magic demo, counting the cost mostly means one stubborn move: refusing to be impressed by the future until you've audited the present.
Sources: widely reported 2025 study (MIT-affiliated) finding ~95% of enterprise generative-AI pilots delivered no measurable P&L impact; reporting on AI model hallucination in production settings; "Salesforce Touts AI Promise Over Reality" and similar coverage of promise-over-performance marketing; MindStudio and procurement reporting on pilot-to-production cost overruns (500–1,000%).
