In the last eighteen months, "AI" has become the most abused word in enterprise software. Every consultancy has an "AI practice." Every staff-aug shop rebrands as an "AI transformation partner." Every SaaS tool bolts on a chatbot and calls itself an "AI-native platform." And most of it is AI washing: rebranded services, shiny demos, and vague promises that never survive contact with a real P&L.
The problem is not that AI transformation is impossible - it's that the market has learned that saying the word is enough. Real transformation looks fundamentally different from the pitch decks. This is how to tell them apart.
What AI washing looks like
AI washing has a house style. Once you know it, you can spot it in a fifteen-minute discovery call.
- Deliverables are demos, not decisions. The proof-of-concept dazzles in a Zoom call and never gets deployed to a workflow that has real users, real error rates, and a real owner on the customer side.
- Metrics are activity, not outcomes. "We processed 50,000 documents" instead of "we cut reconciliation time by 68% and reassigned two FTEs to higher-value work."
- The model is the product. The vendor talks about which foundation model they use, which vector database, which framework - anything except your workflow, your data, and your adoption plan.
- No mention of change management. The pitch skips the humans who have to change what they do on Monday morning. That's the part that actually fails.
- Pricing is a headcount table. "Two senior engineers and a solutions architect at $X/month." That's staff augmentation with an AI sticker.
What real AI transformation looks like
Real transformation is boring in the pitch and dramatic in the outcome. Look for these five signals.
1. It starts with a workflow, not a model
The first artefact of a real engagement is a workflow map: who does what, in what tool, with what data, and where the time and error rate actually sit. Only then does the conversation move to what a model could automate or augment. Vendors who lead with "we'll fine-tune a Llama variant" before they've looked at your ops are selling a hammer.
2. There is a named owner on your side
No AI system survives without an internal champion who owns adoption, error escalation, and iteration. Real programs identify that person on day one and design around their calendar. AI washing hands you a Slack channel and a monthly steering committee.
3. Adoption is measured, not assumed
Real transformation instruments the workflow before and after. It reports weekly on usage rate, override rate, and time-to-complete. It shows a graph that goes up. AI washing shows you the same demo screenshot in every QBR.
4. The team ships in weeks, not quarters
A first production workflow should be live in six to eight weeks. Not a POC - a real workflow, used by real humans, with real error handling. Anything longer usually means the vendor is doing exploratory R&D on your dime. That's a legitimate spend, but it's research, not transformation - price it accordingly.
5. IP and knowledge transfer are in the contract
After the engagement ends, your team should be able to run, monitor, and iterate on the system without the vendor. That means documented code, deployed infra you own, evaluation harnesses, and a handover plan. If the vendor's business model requires you to keep them on retainer to keep the lights on, that's a subscription, not transformation.
The five questions to ask on the discovery call
- Show me the last three engagements you deployed to production. What's the current usage rate?
- Who on our side owns adoption, and what's their weekly time commitment?
- What does the workflow look like today, and what does it look like in eight weeks?
- Who owns the model weights, the prompts, the eval harness, and the deployment infra when we're done?
- What's the smallest change we could make in the first four weeks that a real user would notice?
If the answers are vague, the engagement will be vague. If the answers are specific, boring, and slightly uncomfortable, you're probably talking to someone who's shipped this before.
The cost of getting this wrong
A stalled AI program costs more than the invoice. The typical mid-market company we speak to has spent $200K–$400K on AI initiatives over 12–18 months with nothing in production. That's the fee - but the real cost is the fourteen months of executive attention, the internal team that lost trust in AI as a category, and the competitor who quietly shipped while you ran a POC.
The good news: this is a solvable problem. The frameworks for shipping real AI systems - workflow-first design, embedded adoption, evaluation harnesses, clean handovers - are well understood. They're just rarely priced or scoped in a way that survives a procurement process built for staff augmentation.
The bottom line
AI transformation is a workflow discipline dressed up as a technology purchase. When you evaluate vendors, ignore the model names and the demo reels. Ask about workflow, ownership, adoption, timelines, and handover. The vendors who can answer those questions concretely are the ones worth a second call.
The ones who can't will happily wash your budget in AI-branded activity for the next twelve months. Your call.