DEV Community: Daniel Wishnia

Your AI Rollout Didn't Fail Because of the Technology

Daniel Wishnia — Mon, 22 Jun 2026 11:21:48 +0000

By Daniel Wishnia · Founder, Wish On Line · wishol.com June 2026 · 7 min read

Enterprise AI adoption almost never fails the way the post-mortem claims. The slide deck blames the technology, or "change resistance," or users who weren't ready. I've run these programs inside a listed real estate group, a top-five national insurer, hospitality operators across several countries, and teams small enough to fit in one room, and the real story is duller and far more fixable. The technology worked. The way it was deployed is what broke.

There's a moment in every rollout when the dashboard tells the truth. Usually week six. Licenses assigned: hundreds. Weekly active users: a thin stripe along the bottom of the chart, mostly the same fifteen enthusiasts who'd have found the tool on their own. The kickoff was loud, the demo got applause, and then the organization quietly went back to working the way it did in March.

That's only half the picture, though. And the half nobody puts on a slide is more interesting.

While the official rollout sits empty, AI has already walked in the back door

Here's what I find underneath the flat usage chart. The AI didn't fail to arrive. It arrived without anyone deciding it should.

One team is pasting client documents into a public chatbot to summarize them. Marketing is drafting in tools procurement has never heard of. Someone in operations automated a weekly report with a model and told no one. People use assistants every day without knowing which data they're allowed to enter and which they aren't. The strategy committee thinks adoption is at 4%. The real number, if you could see it all, is much higher and completely ungoverned.

This is Shadow AI, and it isn't really a story about employees breaking rules. It's a signal about the organization itself: the company is moving more slowly than its own people. When leadership gives no direction, motivated teams route around the silence. You don't get zero adoption. You get invisible adoption, which is worse, because invisible adoption carries every risk and captures none of the value.

In regulated environments, the stakes are blunt. Working with an insurer, the questions that mattered weren't about prompt quality. They were about whether sensitive data was leaving controlled systems, whether a decision made with AI assistance could be traced afterward, and whether two departments were quietly solving the same problem twice. Prohibiting everything doesn't work; the teams just hide it better. Looking the other way doesn't work either. You don't slow AI down by banning it. You bring it into the open with leadership, clear rules, and named ownership.

So the real failure has two faces at once. A top-down rollout that's all license and no behavior, and a bottom-up reality that's all behavior and no governance. Most companies have both, and they're the same root cause wearing two costumes.

What enterprise AI adoption actually is

Enterprise AI adoption is the process of turning licensed AI capabilities into changes in daily work, measured by whether specific roles complete specific tasks differently than before. It is not license distribution. It is not a training event. It is not a usage statistic, and it is definitely not the count of people who logged in once.

A rollout has succeeded when the controller closes the month using the tool, the underwriter assesses a case with it, the marketer drafts in it by default, and the organization cannot quietly revert without someone noticing the loss. That definition sounds obvious. Almost no program I've audited was actually managed against it. They were managed against deployment milestones, which is exactly how you produce the proud announcement and the empty dashboard in the same quarter.

The three failures, in the order they happen

The generic training failure comes first. Most programs teach the tool, not the job. Here's the interface, here's what a prompt is, here are ten use cases that apply to no one in particular. It works as a theater and fails as a transfer. A financial controller doesn't need "how to use Copilot." She needs to see her actual monthly close completed in a third of the time using her real data, plus the three places where the model will mislead her. When I built the AI-and-Excel workshop for a group's business controllers, the whole thing was built from their real files, even down to an awkward constraint about where the assistant could and couldn't sit inside the spreadsheet. People adopt when they recognize their own Tuesday in the training. They never adopt a feature tour.

The missing-constraints failure comes second, and it's the one with no budget plans for. Every organization has plumbing realities that silently cap what the AI can do. Files living on legacy on-premises servers that the tool can't index. Sensitivity labels half-deployed. Meeting recording switched off across most departments. Data sitting in a system with no connector. Roll out without mapping these, and your pilot users hit invisible walls in week one, decide "it doesn't work here," and that verdict travels faster than any comms plan. In the insurance program I ran, we documented those constraints up front and tested eighteen real scenarios against that honest map. The point of a pilot isn't to show off the tool. It's to find the walls before five hundred people find them for you.

The measurement failure comes third and finishes the job. "Adoption" gets reported as logins, which is like rating a gym by counting door swipes. What works instead is a per-role scenario matrix: for each function, the five tasks the tool should transform, each with a pass/fail test and an owner. Now the rollout has a scoreboard that managers actually understand, the gaps have names, and the conversation moves from "people aren't engaged" to "scenario twelve fails because the data source isn't connected," which is something a human can go fix on Thursday.

The six gaps Shadow AI leaves behind

When adoption stays invisible, the cost shows up as a predictable set of holes. I see the same six in nearly every organization that skipped the governance step:

Teams using AI with no oversight.
Sensitive data leaving controlled systems.
The same process was rebuilt in parallel two or three times.
Decisions no one can trace back later.
Productivity gains that never get measured or scaled.
Governance that exists on paper and nowhere else.

None of these gets solved by a better model. They get solved by a decision to govern, which is a leadership act, not an IT ticket.

The shape of a rollout that sticks

Strip away the vendor methodology decks, and the working pattern is short. Start narrower than feels ambitious: two or three roles, their highest-friction recurring tasks, real data in every exercise. Build role-specific playbooks instead of tool documentation, including the failure modes, because teaching people where the model is wrong earns more trust than pretending it never is. Run the constraints audit before the pilot, not as its autopsy. Assign a named owner to the scenario matrix and review it monthly, like any other operating metric. And bring Shadow AI into the light early, not with a ban but with a simple, usable policy that tells people what they can put where, so the energy already in the building gets pointed somewhere safe.

The honest timeline is quarters, not weeks. The honest budget is mostly people-time, not licenses. Any plan whose cost is 90% software is a plan to fail with excellent procurement.

Why is this now a board topic, not an IT one

For a while, a stalled rollout was an embarrassing line in the IT review. Two things changed that. The first is compounding skill. The organizations where AI use has become the default work are accumulating prompt fluency, workflow redesign, and hard-won judgment about the tools at a rate latecomers can't buy back later, because it lives in people, not contracts. The second is regulation. The EU AI Act and its relatives quietly reward exactly the disciplined pattern described here, documented use cases, named owners, tested constraints, traceable decisions, and they punish the laissez-faire version where nobody can say how, where, or with what data the company is using AI.

Which is the real question for any executive now? Not whether your company is using AI; it is, with or without you. The harder one is whether you actually know how, where, for what, and with whose data. Adoption was never the soft part of the AI story. It was always the whole story. The technology was the easy bit.

FAQ

What is Shadow AI? Shadow AI is the unsanctioned use of AI tools by employees without organizational oversight, governance, or data controls. It usually signals that the company is moving more slowly than its own teams, and it carries the risks of AI use while capturing little of the measurable value because nobody is tracking or scaling what works.

Why do most enterprise AI rollouts fail? They fail on deployment, not technology. The three recurring causes are generic training that teaches the tool rather than the job, unmapped technical constraints that block pilot users in week one, and measuring adoption by logins rather than by whether real tasks have changed. All three are fixable without changing the underlying tool.

How do you properly measure enterprise AI adoption? With a per-role scenario matrix: for each function, list the specific tasks the tool should transform, give each a concrete pass/fail test, a status, and an owner. This replaces vanity login counts with a scoreboard that turns vague "low engagement" into named, fixable blockers.

Daniel Wishnia is the founder of Wish On Line (https://www.wishol.com). He has led GenAI adoption and Copilot 365 enablement across a listed real estate group, a top-five national insurer, and hospitality operators, following his tenure as Chief Digital Transformation Officer. Contact: daniel.wishnia@wishol.com

The Citation Economy: What Replaces Traffic When the Click Dies

Daniel Wishnia — Mon, 15 Jun 2026 04:18:47 +0000

By Daniel Wishnia · Founder, Wish On Line · wishol.com
June 2026 · 8 min read · Field: AI Strategy / Citation Economy

Every business model on the open web rested on one transaction: attention arrived as a click, and you converted the click into money. Media sold the attention to advertisers. Retail converted it to purchases. B2B converted it to leads. Twenty-five years of marketing practice, an entire industry's worth of tooling, and most of your current KPIs all assume the click keeps arriving.

It is not arriving. Over 60% of searches now end without one, AI Overviews answer questions for more than a billion people a month without sending most of them anywhere, and assistants increasingly complete the whole task inside the conversation. The funnel didn't get less efficient. The funnel's raw material is being shut off.

I started calling what comes next the citation economy, and the name has stuck because it describes the actual replacement mechanism, not just the loss.

What is the citation economy?

The citation economy is the system in which a brand's commercial value is determined by how often AI systems cite or name it as the answer, rather than by how much traffic its website receives. In the click economy, visibility meant ranking, ranking produced visits, and visits produced revenue. In the citation economy, visibility means being the entity a model trusts enough to name when a person asks a question in your category. The citation is the new impression, the recommendation is the new conversion path, and trust transferred from the AI to the user is the new currency.

This is not a metaphor. It changes what you measure, what you build, and who in your organization owns growth.

Why citations behave differently from clicks

Clicks were rented. You paid for them with ad budget or earned them with rankings, and the moment you stopped paying or slipped a position, the flow stopped. Citations compound. A model that has learned, from consistent evidence across the web, that your brand is the credible answer in a category keeps giving that answer, conversation after conversation, at zero marginal cost to you. Reputation became infrastructure.

Citations are also winner-concentrated in a way rankings never were. Page one of Google had ten spots and everyone skimmed several. An AI answer names two or three entities, sometimes one. The gap between being cited and not being cited is the gap between existing and not existing for that user, because there is no position eleven to limp along in. Categories will consolidate around their most-cited players faster than they ever consolidated around their best-ranked ones.

And citations are earned through a different mechanism. Rankings could be engineered with links and on-page tactics. Citations come from what I call the consensus loop: models cross-check what you claim about yourself against what the rest of the internet says, trade press, communities, reviews, Reddit, event programs, other people's content. Ahrefs' study of 75,000 brands found brand mentions correlate roughly three times more strongly with AI visibility than backlink authority. The strongest signal isn't what you publish. It's what others confirm.

The new scoreboard

If traffic is no longer the score, something has to be. Four metrics make up the citation economy dashboard I build with clients.

Share of Model comes first: across a fixed set of real buyer questions run through ChatGPT, Gemini, Perplexity, and AI Mode, what percentage of answers name you versus competitors? It's the successor to share of voice, and it's measurable today with a spreadsheet and discipline.

Citation quality comes second, because being named is not enough. Are you cited as the recommendation, the example, or the cautionary tale? With accurate facts or stale ones? The semantic association a model carries about your brand is auditable, and it's often wrong in ways no one inside the company has noticed.

Consensus coverage is third: the volume and consistency of third-party evidence about you across the sources models check. And machine legibility is fourth, the unglamorous plumbing: structured data, crawler access, content formatted as extractable answers. AI crawlers don't execute JavaScript, so a surprising share of corporate websites are simply unreadable to the systems now mediating their demand.

Notice what's missing. Sessions. Bounce rate. Keyword positions. Keep reporting them if you like, but they are lagging indicators of a game that's ending.

Who owns this? (The org-chart problem)

Here's where the citation economy stops being a marketing topic and becomes a management one. Citations are produced by the joint output of SEO, PR, content, product data, community, and customer experience. In most companies those report to different people who meet quarterly at best. The click economy tolerated those silos because each channel had its own funnel. The citation economy doesn't, because the model reads all of it as one body of evidence about one entity.

In practice the brands moving fastest have done something structurally simple: assigned a single senior owner for AI visibility, given them a baseline (the Share of Model measurement), and let them pull levers across departments. The ones moving slowest are running GEO as a line item inside the SEO budget, which is roughly like running e-commerce as a line item inside the print catalog budget in 1999. The category precedent is not encouraging for them.

The transition playbook

You don't abandon the click economy on a Tuesday. Search still sends traffic, ads still work, and the two systems will overlap for years. The mistake is sequencing, treating citations as something to address after the traffic decline becomes painful, when the entire dynamic of compounding means the positions are being locked in now, while attention is cheap.

The order I give boards: measure first, because a Share of Model baseline turns an abstract threat into a number with names attached. Fix legibility second, schema, crawler access, answer-formatted content, since nothing else lands while machines can't read you. Build consensus third, through trade press, original research, community presence, the slow assets. And re-measure quarterly, because models retrain and retrieve, and the scoreboard genuinely moves.

The brands that ran this loop through 2025 are already showing up as the default answers in their categories. Their competitors will eventually notice, run the same playbook, and discover the hard part: a model's learned consensus is much easier to establish than to displace.

The click is dying. The citation is already deciding who wins. The only open question is whether your company found out from this article or from its revenue line.

FAQ

Is the citation economy just another name for GEO?
No. GEO (Generative Engine Optimization) is a set of tactics for getting cited. The citation economy is the economic shift that makes those tactics matter: the relocation of brand value from harvested traffic to AI-conferred trust. GEO is how you compete; the citation economy is what you're competing for.

Can citations actually be measured?
Yes. A fixed prompt set of 50 to 100 real buyer questions, run monthly across the major AI systems with every brand mention logged, produces a defensible Share of Model trend line. Tooling is emerging, but the manual version works today and costs an afternoon a month.

Does this kill SEO?
It demotes it from strategy to foundation. Google's AI features still draw heavily on well-ranked pages, so SEO remains the entry ticket for one platform. But ChatGPT and Perplexity weight entity authority and third-party consensus far more, and only a small minority of domains get cited by both ecosystems for the same query. One discipline became two.

Daniel Wishnia is the founder of Wish On Line. He coined the term "citation economy" and advises organizations on AI visibility strategy, drawing on a CDTO tenure at Aroundtown SA and two decades in digital distribution. Contact: daniel.wishnia@wishol.com

The Conversational Shift: Are We Ready to Talk to a Device?

Daniel Wishnia — Thu, 28 Aug 2025 07:41:09 +0000

Twenty years ago, I began discussing something that most people found far-fetched.
I called it Digital Sense, the idea that information, education, commerce, development, and creativity would no longer be about typing, clicking, or filling forms. Instead, they would become conversational tasks.

I spoke of Conversational Navigation, a future where we wouldn’t browse menus or search bars but say what we wanted, and the system would understand. I spoke about Conversational Commerce, where we wouldn’t “add to cart” but ask directly, “Find me the best option, under this budget, delivered tomorrow.” I even pointed toward Conversational Creativity, which involves generating images, videos, or music through a dialogue with technology.

Back then, these were just sketches on the wall.
Today, they are a reality.

From Clicking to Talking

Look around: voice assistants are no longer clunky gadgets. AI models perceive, hear, and respond in real-time. Search results are turning into conversations. Websites are starting to answer rather than display. Messaging apps are turning into shopping malls.

The entire interface of the digital world is shifting:
From clicks to questions.
From menus to meaning.
From navigation to conversation.
From developing to asking.

And yet, the biggest question isn’t about the technology. It’s about us.

Do We Know How to Converse?

We humans spent the last 30 years training ourselves to “speak computer.” We learned to type keywords into search engines. We filled forms with fields designed by someone else. We clicked through endless menus to find a simple option.

Now the machine speaks our language. But here’s the paradox:
Are we fluent enough in our own intent to ask well?

Talking to an intelligent device is not like asking Google a quick fact. It’s a relationship of intent, context, and trust. If you say, “Plan my trip to Rome,” the system will do it, but the outcome depends on how clearly you describe your needs, how much you reveal, and how well you check the answers.

Conversational AI forces us to develop a new literacy: the ability to express goals, constraints, and expectations in human language, not computer shortcuts. In other words, the art of good conversation is now a technological skill.

The Shift in Practice

Education: No more searching for “best math tutorial.” You ask, “Explain quadratic equations like I’m 12, but challenge me with one hard problem.”
Commerce: No need for endless scrolling. You ask, “Show me a jacket that works for Berlin winters, under €200, and ethically made.”
Creativity: No Photoshop tutorials required. You say, “Create a sketch of an underwater office — futuristic, minimal, glowing with soft light.”
Work: Forget forms and workflows. You state your task, and the assistant composes documents, schedules meetings, or even negotiates small steps with other agents.

We are stepping into a world where conversation is not a support channel; it is the interface.

The Human Side of the Shift

Here is the danger: if we treat machines as perfect, we will lose judgment. If we treat them as toys, we will miss the opportunity.

The right mindset is somewhere in between. Think of your conversational AI as an apprentice, capable, fast, creative, but needing guidance and review. We must stay human in the loop: checking, refining, deciding.

This is not about replacing human thinking. It’s about freeing it. When machines handle navigation, transactions, and repetitive tasks, humans can focus on creativity, strategy, empathy, and meaning.

So, Are We Ready?

We dreamed about this for decades. Now it’s here. The digital world is no longer something we click through. It’s something we talk to.

The question is no longer, “Will machines understand us?” They do.
The question is, “Will we learn to converse with them well enough to create the future we actually want?”

That is the conversational shift.
And it’s only just beginning.

ConversationalAI #AIShift #FutureOfWork #DigitalSense #HumanInTheLoop