Individual speed is not organizational capability. The next challenge is designing AI systems the organization can trust, govern, maintain, and actually own.
Eight in ten people now say artificial intelligence has made them personally more productive at work. Ask their employers whether any of that shows up in operating profit and the number collapses to 37 percent, according to McKinsey’s 2026 global survey on AI, essentially unchanged from the year before even as more organizations scaled AI across the enterprise.
Anca Platon Trifan works with organizations to examine how work actually moves through the business and to audit where AI is already being used. Then she designs the intelligence architecture that comes after, the systems meant to support decision-making without losing the context, controls, and human judgment the organization still needs.
She walks into companies certain they already have an AI strategy. What she finds underneath is almost never a shortage of AI. She knows the other side of it too, running 49 active AI agents inside her own business and starting most mornings at 5:30 keeping them alive.
What follows is her account of the cost nobody budgets for, the test she uses to find it, and what she wants decided before any system is allowed to act on its own.
We Already Use AI
An experiential agency of roughly 100 people brought her in. They were paying for company-wide access to Gemini, and individual employees were also using Claude and ChatGPT. From the outside, Trifan says, it would have been completely reasonable for leadership to say they already used AI.
Then she looked at how the work was actually being done.
There was no common operating model underneath any of it. Employees had built their own prompts, their own habits, their own ways of checking a result. Two people could perform the same business task with the same technology and produce completely different processes. The reasoning behind a prompt that worked was rarely written down, and a correction one person made never became an instruction for anyone else.
That creates a form of organizational fragility. The company may own the software license, but much of the useful capability lives inside individual behavior. When the employee leaves, changes roles, or simply forgets how they got a particularly good result six months ago, that knowledge goes with them.
This is not a small-company problem. Verizon’s most recent Data Breach Investigations Report found that frequent AI use by employees on corporate devices jumped from 15 percent to 45 percent in a single year, and that unapproved AI now ranks third among the non-malicious activities through which company data leaves the building.
For Trifan, shadow AI is both a governance problem and an organizational signal. Employees using unapproved tools may expose company information or create inconsistent processes, but their behavior also tells leadership where the sanctioned systems are failing the work. Her response is not to prohibit the tools. It is to understand why people reached for them, what information they are putting into them, what business problem they are solving, and whether that capability should be brought into an approved and governed workflow.
“Simply banning those tools does not solve the underlying workflow problem that caused people to reach for them,” she says.
Show Me How the Work Is Actually Done
Her audits begin with a request, not a questionnaire. “I start by asking people to show me how the work is actually done rather than how the process is described in a meeting or written in an SOP,” she says.
Her simplest test costs nothing. Ask two people who perform the same task to demonstrate it side by side. The differences appear within minutes.
The instinct is older than her AI work. She finished a computer science degree at Babes-Bolyai University in Cluj in 2003, then spent the next two decades in live production, where nobody cares what the run of show says if the room is doing something else.
In one agency review she mapped work across sales, RFPs, proposals, registration, speaker management, sponsorship, reporting, finance, production, and post-event follow-up. Every individual tool was doing what it had been bought to do. The failures were in the spaces between them: information re-entered by hand, reports rebuilt because no system could answer the question leadership had asked, decisions that lived in one person’s memory. “A lot of experienced employees had also become the connective tissue between systems,” she says.
So she ranks workflows instead of automating everything, then starts with documentation rather than software: the approved source, the rules the experienced employee has been applying, the exceptions the system should flag, the point where a human must intervene, who owns the workflow, and the outcome being measured.
The audit is the diagnostic stage, not the product. Once the organization knows where information originates, how decisions are made, which rules are repeatable, where exceptions occur, and who owns the outcome, it becomes possible to design AI systems around the real operation rather than around an idealized process diagram.
The question shifts from where AI can be added to, in her words, “where is work breaking down, what would a better process look like, and which part of that process can AI perform reliably?”
What Happens When the Model Is Wrong
Live production taught her to trace consequences before approving anything. Move a single session and you can affect catering, labor, speaker timing, AV, room resets, sponsor commitments, attendee movement, contracts, and everything that follows. A five-minute change rarely stays a five-minute change.
It also taught her that reliability is something you design in advance. Rehearsals, backups, confirmed sources, a named person with authority to call the cue, a decision already made about what happens when the microphone fails or the network disappears ten minutes before doors open.
“You do not get to tell 700 people that the system worked beautifully during testing,” she says.
She brings the same assumptions to AI.
I assume a model will eventually misunderstand something. I assume an integration will fail. I assume a source will contain stale or incomplete information. I assume someone will enter something we did not anticipate.
What follows is a set of questions she wants leadership answering before an agent is allowed to act. What information is the system allowed to access. What decisions can it influence. What actions can it take without approval. How far can an error travel before a human sees it. Can the action be reversed. What evidence does the reviewer receive. Who owns the outcome when the system is wrong.
Those questions matter more every quarter, as organizations move from people asking AI for answers to agents and automated systems taking actions on their own. A polished recommendation handed over without the evidence, the uncertainty, the sources, or the ability to intervene is not human oversight. It is a signature.
She has been on the receiving end of a plan that stopped matching reality. She trained eight months for an IRONMAN, and by the time she reached the run the strategy she had trained for was useless, which left her with the question she now takes into system design: “Given what is true right now, what is the next intelligent thing I can do?”
The Bill That Starts After the License
Licensing is the easiest AI cost to see. Trifan’s argument is that the expensive part begins the day people start using it.
Someone has to validate outputs. Someone maintains the instructions, decides which sources are trusted, what happens when they conflict, who has access, and what breaks when the model or the software changes. Add retraining, troubleshooting, exception handling, security, duplicated experimentation, and the time spent correcting work that looked finished and was not.
She counts it in her own operation. Those 5:30 mornings produce gateway errors, authentication failures, broken approvals, stale setup documents, agent testing, validation, fixes, and retesting, on systems built to give time back. “Time saved at the point of execution does not automatically become organizational capacity,” she says.
The industry is arriving at the same arithmetic. Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, and attributes the failures to escalating costs, unclear business value, and inadequate risk controls rather than to the models themselves.
Her test for whether an automation counts is deliberately unforgiving: “If AI saves ten minutes in drafting and creates twenty minutes of additional checking, I do not count that as an improvement.”
Responsible AI Is a Leadership Decision
Trifan wants responsible AI moved out of the margins of the conversation. Bias, privacy, security, shadow AI, environmental impact, data provenance, permissions, autonomy, and accountability are not issues to hand to legal or IT after deployment. They are architecture decisions.
If AI is influencing hiring, customer communication, strategy, operations, creative work, or access to information, leadership has to understand what information the system is using, what assumptions it may reproduce, what resources it consumes, what authority it has, and who remains accountable for the result.
A system can be technically functional and still reinforce poor assumptions, incomplete data, or historical inequities, which is why she pushes back on review that exists only on paper. “Human review cannot simply mean that a person clicks approve at the end,” she says. People need enough context, evidence, and authority to challenge what the system produces.
She is equally interested in restraint.
I do not believe every process should be automated simply because it can be. Some decisions deserve friction. Some work should remain human because judgment, trust, authorship, discretion, or accountability are part of the value being created. A mature AI strategy has to include decisions about where not to automate, where to slow the system down, and where the cost of efficiency may be higher than the time saved.
Ninety Minutes, Then Five
Her authority on this comes from a combination rather than a single track. More than twenty years operating technical systems under pressure, from front of house and monitors on shows featuring Snoop Dogg and Linkin Park to technical direction and show calling for C-suite broadcasts and stadium-scale hybrid events. More than 100 AI workshops taught. Seven years of the Events: Demystified podcast. Dozens of AI agents built and maintained inside her own business, and organizational workflows audited inside other people’s. The recognitions followed, among them Top Voice in Events and a place on the Top 13 LinkedIn Influencers in Events list, and she now advises the University of San Francisco School of Management’s AI Driven Marketing Leadership Program. One client’s summary of her workshops is short: “Anca customized my team’s session on AI and facilitated in a way we could understand. It addressed our specific guardrails and use cases.”
The proof she reaches for first is a scheduling question at a live conference for roughly 700 attendees. A client asked whether six thirty-minute sessions could be restructured into four forty-five-minute ones. It sounds like calendar math. In live production it moves room resets, staffing windows, lunch timing, sponsor commitments, program flow, and several documents that each hold a different piece of the same event.
Working the consequences by hand took her about 90 minutes. After she defined the rules and built the workflow around them, the same restructuring took about five minutes. Roughly 85 minutes came back on one change.
The system compared dependencies and surfaced what needed attention: lunch compressed, a staffing window moved, sponsor exposure at risk. “It still did not make the final decision,” she says. She knows which tradeoffs that particular client can absorb, and when losing five minutes turns into a collision involving 700 people, catering, production, sponsors, and the next block of programming.
Where Anca Is Now
Trifan works from Meridian, Idaho, across both halves of her practice. Through Tree-Fan Events Productions she produces corporate events, SKOs, conferences, executive programs, and live experiences, and alongside that she takes organizations through workflow audits and the design of AI intelligence architecture.
The consulting work increasingly comes down to helping leadership teams decide how AI should operate inside the organization: what information systems should access, what knowledge should be retained, what decisions can be delegated, where human judgment must remain, how the systems should be governed, and how to build capability that survives the employee who first introduced it. Her keynotes and workshops are moving toward the same questions, and she is developing Season 12 of Events: Demystified with founders, builders, executives, and researchers willing to discuss what is failing as openly as what is working. The work she wants more of runs to keynotes and workshops, consulting and workflow-audit clients, intelligence architecture projects, and executive-level advisory partnerships.
Outside it she is a five-time bodybuilding champion who has since finished an IRONMAN, and she moved her next competition from September to November because forcing another punishing timeline was no longer the intelligent choice. In October she and David T. Stevens run the #Fit4Events push-up challenge at IMEX America for the fifth year.
Her governing rule for her own AI use is one line: “I don’t need AI to make me sound impressive.” The one she wants leadership to sit with is longer. “Responsible adoption requires understanding not only what AI makes possible, but also what it makes easier to overlook, abandon, outsource, or normalize.”
Five Checks Before You Automate Anything
- Ask two people to demonstrate. Have two employees who perform the same task show you how they do it, and watch how fast the differences surface.
- Document the rule before automating it. Name the approved source, the exceptions, the review point, and the owner, because that is the capability the company actually keeps.
- Count the supervision, not the license. Budget for validation, maintenance, exception handling, and the hours someone will spend keeping the system trustworthy.
- Decide where you will not automate. Protect the decisions where judgment, authorship, discretion, or accountability are part of the value being created.
- Check that a manual step disappeared. If the work moved from drafting into checking, the workflow did not improve, it relocated.
Anca Platon Trifan produces live events and designs AI intelligence architecture at ancaplatontrifan.me. Connect with her on LinkedIn.
