The best influencer marketing advice never comes from a report
Think about the last thing that got you fired up about your job this year. Odds are it didn't come from a report. It came from someone who does the job, saying what happened when they tried it.
So we've filled a day with people like that.
Maya Shaff, ŌURA. Leah Walker, Adobe. Georgia Humphries, Stanley 1913. Enara Roy, Halfday. Tyler Vaught, Edelman. Josh Rangel, Ogilvy. Sarah Whittle, ex-Crocs and Duolingo.
And that's just the first wave of speakers.
Return on Influence Festival '26 is a free, virtual day dedicated to influencer marketing. Like any good festival, there are stages. But no mud, no queues, and absolutely no porta-potties.
Sign up today, and the rest of the lineup lands in your inbox first.
An AI system can give the right answer and still make things worse.
Imagine a hospital introduces a tool that identifies more patients who need specialist review. The predictions are accurate. The team celebrates better detection.
But the specialist service cannot absorb the additional referrals. The waiting list grows. Patients who need timely care face longer delays.
The model did its job. The organization still has a problem.
That is an illustrative scenario, but it captures the question at the center of our second peer-reviewed paper, published in Frontiers in Artificial Intelligence on September 14:
What happens after an AI output enters the real world?
Our paper, From model failure to system harm, focuses on healthcare. The question has stayed with me because it also applies to how we use AI at work.
The signal
We often evaluate AI at the moment it produces an answer.
Was the prediction accurate? Did the summary capture the facts? Did the recommendation make sense?
Those are necessary questions. They leave out the people, decisions, and constraints that come afterward.
In the paper, we propose a five-stage pathway for tracing risk: an upstream vulnerability, the AI's behavior, the human response within a workflow, the resulting decision or system change, and any downstream harm.
In everyday language, follow the chain:
What went in → What AI produced → How someone responded → What changed → Who was affected.
Risk can be caught or amplified along the way. A wrong answer that someone catches may never cause harm. A correct answer that triggers an overloaded process may still create problems.
One study we discuss makes the human part especially clear. In a randomized study involving 457 clinicians reviewing clinical vignettes, systematically biased AI without explanations reduced diagnostic accuracy by 11.3 percentage points compared with the unaided baseline. Adding image-based explanations did not significantly mitigate the negative effect. Read the JAMA study.
That was a controlled decision study, not a measurement of patient injuries. It shows why having a professional review an output does not, by itself, establish that the review will work.
The reviewer needs relevant knowledge, enough time, access to supporting information, and authority to intervene.
Our paper is a Perspective that brings existing evidence together and proposes an operational framework. It does not establish how often AI causes patient harm, and the framework still needs prospective validation.
My read
The lesson I take beyond healthcare is that we need to design the handoff as carefully as we choose the model.
Consider an AI meeting assistant that invents a deadline.
In a draft summary, someone might catch it. If that summary automatically updates a project tracker, assigns work, and triggers messages to a client, the same error now travels through several systems.
The useful question becomes: where could someone catch it, and what would they need to do so?
This changes how I would evaluate an AI workflow. Alongside output quality, I would look at what gets acted on, how much checking it creates, how quickly problems are noticed, and whether the next person has the capacity to handle the work.
An approval button tells you that someone clicked. You still need to understand what they were able to verify.
Put it to work
Choose one recurring AI task this week, such as a meeting summary, customer reply, or research brief. Follow one output all the way to its consequence.
Ask four questions:
What action can this output trigger? A message, a commitment, a payment, or a change to a shared record?
Who checks it, and against what? Name the person and the original source they need.
Can that person realistically intervene? Check their time, access, workload, and authority.
What would make us pause? Define a specific warning sign and who handles it.
Then improve one weak handoff. For a meeting assistant, that might mean confirming owners and deadlines against the transcript before they enter the project tracker.
The human side
This paper brings together two parts of our lives: my work in AI adoption and Irem's work in medicine.
Our first paper explored digital twins for adaptive cancer care. This one focuses on what happens when AI enters clinical workflows.
Healthcare makes the stakes especially visible. Behind an output is someone waiting for care, someone making a decision, and someone responsible for what follows.
I want that perspective to stay close as we bring AI into more of our work. A useful system should give people better information and the conditions to act on it well.
The question
Where in your work does an AI answer become a real-world commitment, and who owns that moment?
Reply and tell me.
Until next week,
Burhan


