A Clean Demo for a Dirty Organisation

Walk any enterprise AI event right now and the room starts to fold in on itself.

Different booths. Different logos. Different badges. The same diagram. First connect your data. Then add agents. Then automate the workflow. Then show the executive dashboard. The colours change, the nouns change, the promise holds.

Google says it. Salesforce says it. ServiceNow says it. The vertical SaaS vendor says it. The consultancy stitching APIs together says it. The twelve-person startup with a very confident founder says it too.

At one booth I asked a plain question: what happens when the CRM has duplicate customer records, owners who left in 2019, and three private spreadsheets nobody admits to? The demo went quiet. That quiet is the whole article.

One script, many machines

The convergence is real. Nearly every vendor describes the same journey: capture the data, govern it, let the model query it, build an agent, show the result on a dashboard. Google has a clean name for the centre of gravity, the Agentic Data Cloud, ^[1] and Reuters reported earlier this year that Google is putting agents at the heart of its enterprise push, ^[2] so this is the official direction, not a side bet.

The architectures underneath stay genuinely different. A hyperscaler building infrastructure, a foundation-model lab, a workflow SaaS company, and a consultancy wiring APIs together play different games, with different security models, data access, pricing power, and failure modes. The part that fused is the pitch.

That is the tell. The language has stopped sorting the infrastructure from the theatre. And the moment every vendor sells the same certainty, the buyer’s danger moves from picking the wrong vendor to believing the demo is a picture of their own business. The window to tell the real builders from the confident talkers is widest now, while the pitch is still new and the failures are still quiet.

Every demo is a clean company

In a demo the data is clean, the permissions behave, the record exists, and the workflow is obvious. The agent retrieves the right customer, summarises the account, drafts the follow-up, updates the CRM, and produces a confident dashboard. The room nods.

Most companies run on something messier. Duplicate customer records from a 2019 migration. Salespeople keeping private spreadsheets because the CRM is too slow. Contacts who left years ago still marked active. Three versions of revenue depending on whether finance, sales, or the founder is talking. Documents spread across email threads, Slack messages, SharePoint folders, and a desktop file named final_final_v3.

Point an agent at that and watch the magic thin out. In production the same agent finds three customer profiles with different names, two stale contacts, a missing contract, and a sales note that contradicts the finance record. It answers anyway, because answering is what these systems are built to do. The dashboard stays confident. The workflow still runs.

“Bad enterprise AI rarely explodes. It quietly automates the confusion that was already there — and the dashboard is the easiest part of the illusion to keep clean.”

Someone pays for that confidence. The frontline rep who sends the wrong follow-up. The customer who receives it. The buyer whose credibility erodes one plausible-but-wrong answer at a time. The failure arrives as a slow leak rather than a crash, which is exactly what makes it expensive: it reads as success right up to the quarter it stops.

Twenty years of master data management, semantic layers, warehouses, and governance programs left most companies without one clean operating truth. Agents inherit that debt. As AI builds fast lanes inside company red lights, the organisation still sets arrival time — and a confident agent running on broken data just reaches the wrong destination faster.

What the sameness hides

Only the visible layer commoditises: the language, the demos, the decks. The real contest runs underneath, over context, accountability, integration depth, and behaviour change.

Underneath the identical pitch, vendors sell genuinely different things. Some sell infrastructure. Some sell workflow lock-in. Some sell implementation labour. Some sell executive confidence. Some wrap a thin model over an old product. A few are building a real operating layer for work. The buyer’s job is to tell them apart.

This is the enterprise cut of a wider shift. When everyone can build, the durable edge moves to the human layer that decides what the market notices and whom it believes. When intelligence itself meters down toward a utility, the scarce thing becomes the judgement to aim it at the right problem and stay accountable for the result. The edge belongs to whoever stays in the room when the clean demo meets the dirty organisation.

Johnny’s verdict

Do not ask a vendor whether they have agents. Everyone has agents now. Ask the uglier questions, the ones that separate a partner from a pitch:

• Show me a data-readiness audit before the demo.
• Show me the failure mode, the same way you showed me the happy path.
• Show me the source lineage behind every answer.
• Show me who fixes duplicate, stale, and contradictory records.
• Show me adoption ninety days after the demo team leaves.
• Show me which part of the fee depends on a measurable workflow change.

That last one is where the sameness breaks. Anyone can run the happy path; only a real partner ties their money to your messy middle.

AI reveals a confused organisation faster than anything before it, and scales the confusion if you let it. The booth that went quiet when I asked about duplicate records was telling the truth the slide left out. The vendors worth choosing say it out loud — and they are the ones you can hold answerable for the result.