The demo is engineered to hide the budget
A demo runs on clean sample data, one pre-wired integration, and a workflow the vendor rehearsed forty times. It is a controlled film set. The number on the final slide — the license, per seat per month — is the one cost the seller is comfortable showing you, because it is the smallest one. Say a 60-person services firm signs at that number, then discovers six months later it has spent triple on everything around the tool. That gap is not a surprise. It is the predictable difference between a staged demo and a live operating model.
Whether an implementation budget survives contact with your business has nothing to do with features. The RSM middle-market AI survey, the OECD report on AI adoption by small and medium-sized enterprises, and Deloitte State of AI in the Enterprise 2026 all point the same direction: programs stall not on model quality but on workflow ownership, data readiness, and the move from experiment to production. So before you watch a single feature, make the seller map one of your workflows on a whiteboard — name the source systems, the data owner, and who reviews the output before it reaches a customer. If they reach for the product instead of your operating reality, you are being sold a film set.
The six line items nobody quotes
Here is the part of the cost that lives off the pricing slide. Walk a quote against this list and watch the real number assemble itself:
Source cleanup. The demo's data was clean because someone groomed it. Yours is in three systems, two spreadsheets, and a shared drive with eleven versions of the same contract. Budget the hours to make it usable, because the tool will not.
Permission design. The model can read everything it is pointed at. That is the problem. Deciding what it may see, and proving a junior account can't surface a partner's comp data through a prompt, is design work — and the CISA AI Data Security Best Practices and the NIST AI Risk Management Framework exist precisely because this step is where unbudgeted programs leak.
Review capacity. Every AI output that touches a client needs a human who checks it until trust is earned. That person has a salary and a calendar. If the plan assumes zero review hours, the plan assumes the tool is never wrong.
Retained logs, training, and post-launch support round out the rest — the audit trail you keep, the sessions to get staff actually using it, and the help desk for week three when something breaks. Separately, verify the data controls you are actually buying: OpenAI Enterprise Privacy is a useful reference for the questions to ask any tool or partner — what you submit, how access is governed, what is retained, and who can approve production use. A vendor who can't answer those in plain language hasn't priced them either.
Ask the one question that ends the demo theater
There is a single question that separates a consultant who will deliver from one selling a slide deck: what will be true at the end of the first production workflow that is not true today? A strong partner answers without hesitating — they name the baseline metric, the workflow owner, the cadence at which adoption gets checked, the review process, and the decision rule for continuing or killing the project. A weak one answers with a roadmap, because a roadmap commits to nothing measurable.
For a business in the 10-to-250-employee range, the deliverable you are paying for is an implementation path your own leadership team can run after the consultant leaves — not a strategy deck, not a tool comparison. The first milestone should be one accountable workflow with a number attached to it. If you want to pressure-test what you're hearing, start with the AI readiness assessment buyer guide to confirm you're even ready to spend, sequence the work against the 90-day AI implementation plan, and keep the result honest with AI ROI measurement without fake savings. Price the hidden work first, and the demo becomes what it always should have been: the least interesting part of the decision.
