- Key Insight: Discover how AI-trained performance systems can systematically disadvantage neurodivergent employees.
- Expert Quote: Ramakrishnan: AI trained on neurotypical inputs risks misclassifying neurodivergent performance.
- Forward Look: Prioritize inclusive training data ahead of wider rollout.
Source: Bullets generated by AI with editorial review
As more adults are being diagnosed with autism, and 2.21% of adults in the U.S. are reported to have autism spectrum disorder according to the most recent CDC data, employers are increasingly addressing the needs of this population — though they may be overlooking AI's role in their inclusion efforts.
A study published in the Journal of the American Medical Association Psychiatry found that between 2011 and 2019, the overall rate of adults diagnosed with autism more than doubled, though a 2025 Gallup report, Neurodiversity in the Workplace, found that 37% of employed neurodiverse individuals do not to share their condition with coworkers, fearing stigma.
Meanwhile, 91% of organizations are increasing their investments in AI, according to the 2026 Data and AI Leadership survey, with 99% deeming AI investments a top organizational priority, and that includes in performance management systems.
Rita Ramakrishnan, founder of Iksana Consulting, executive coach and current master's degree candidate researching coaching modalities for neurodivergent leaders, warns that these AI-powered performance tools are structurally biased.
Read more:
Ramakrishnan also has familiarity with performance review processes, having recently served as chief people officer at startup Cadre before it was acquired and continuing to partner with growth-stage startups as a fractional chief people officer.
She advised companies to evaluate their performance systems and where they could be falling short in accounting for neurodiverse employees — and a critical starting point is how the AI is being trained.
"The number of AI platforms or centered around performance management or performance management platforms that have now integrated AI capabilities is astounding," she observed, noting that the technology is often measuring against a narrow behavioral baseline that might not take into account traits commonly associated with neurodivergence, and thus read as underperformance.
Read more:
"The other piece is more and more people are getting diagnosed, and they're trying to figure out what this means for me, and the community is not a monolith, so you're seeing various traits show up. If we start to vilify these traits, we stop ourselves from being able to design for them," she continued. "In my mind, there's a massive opportunity here around performance management, where we can set new standards around what impact means within the context of an organization, and focus on outcomes versus just outputs, and I think one of the biggest issues here is the oversimplification of that definition of performance."
Her recommendations for companies in evaluating where their performance systems fall short is to first assess the training methodology.
Applying the right training
"The challenge is AI models have to be trained, so you have to give them data points on what to look for," Ramakrishnan explained. "When I was engaging in thoughtful conversations with many of these providers, one in particular stands out, where… she [said] the inputs are largely neurotypical, so the expectations around what good looks like, what good performance looks like, is very neurotypical, and when we think about the inputs that are going into the process and the outputs that are coming from it, the processing that's going through it is inconsistent with the way a neurodivergent brain works."
One example Ramakrishnan shared was eye contact, which "can be really challenging for those with autism."
While this is not a trait for everyone with autism, those who do exhibit it are "still able to thoughtfully identify needs, ideas, iterate, ideate, and they are in their roles for a specific reason," she continued, "but what we're doing is saying we're making the threshold to become an executive a lot harder by establishing neuro normative standards for what good looks like, and so the question is, 'What are you training your models on?'"
A diversity of data
In speaking to one benefits professional about their AI training methods, Ramakrishnan discovered that they input the company's values and handbook as the sole resource.
"That's not exactly a rich data source," she commented.
Contrary to that example, Ramakrishnan shared an experiment she conducted while designing performance management at Square, now known as Block, "where we parse through several years worth of performance data, and we identified a number of high performers who…have been with the organization for a while, and then we took a control sample, and we looked through the test to identify specific qualities or skills or attributes that kept coming up in those performance reviews that were differentiators between those two samples," she explained.
"So I think that's a great way of identifying, hey, these are some of the differentiators that really drive high-quality performance for the organization," she continued. "So, being able to identify rich qualitative data sources that you can offer to the model, so that they're working off of real information that's not neuronormative and also not overly simplified, allows for that thinking of another important data point ins the business strategy."
Read more:
For this data to be helpful, there must be alignment, Ramakrishnan stressed, around an organization's key objectives. She shared that when prompted during her coaching sessions, five members of an executive team could often recite five different objectives.
Once objectives and goals are effectively communicated, the correct skill sets can be identified and used to train AI.
"Putting additional research around [skill sets] is also really helpful, I think," she said. "You just have a richness of available information in this day and age that you can feed to a model to make it more effective, in not just this process, but in a lot of different processes. I just don't know that people are putting in that legwork."
The business case
That legwork could be an impediment for companies viewing any additional investment beyond financial into their AI systems as too time-consuming. But Ramakrishnan cautioned against that thinking.
"You can use AI as a very expensive tchotchke, or you can use it to drive extraordinary value for your business," she said. "If you're not going to put the legwork into setting this up right — making sure your human capital is actually getting human answers and putting in the right parameters — then you paid a whole lot of money for very little value, because the data it's going to give you is unlikely to be sound, and it's unlikely to be useful in the way that you think it's going to be useful."
Start small, and be specific
Getting the most value out of AI in performance management systems requires employers to be as inclusive and accurate as possible in designating the AI's purpose.
"Let's actually identify the jobs to be done that you are hiring this AI for, this agent for this product for, and narrow the scope there," Ramakrishnan said, "because an AI agent or an AI model is going to send whatever it wants at you if you're not specific about the parameters and the prompts that you give it, so be specific, thoughtful, and targeted about the prompts that you give it, and make sure it's solving for a job to be done, that's a critical job."
Incremental changes are also helpful.
Ramakrishnan recommended identifying "small use cases first for ways that [AI] can make you do your work better, or make you do your work faster, or both. Teeny tiny use cases, take a look at your business processes and see what things might be automated. See what things shouldn't be automated… Start small, then ask the bigger questions, then ask more. Put some legwork into thinking about where human thinking and judgment is necessary, and empathy is critical, and identify [where] human judgment is less important."
Make your people part of the change
As employers consider how to more effectively equip performance management systems to be inclusive of neurodiverse employees, those same lessons apply in evaluating the AI all companies are implementing across myriad departments and purposes.
"If you do one thing well, whether you're dealing with a neurodivergent employee population, whether you're doing anything with neurotypical populations or mixed populations, one is acknowledge that AI is not infallible, it is a deeply flawed system, but it can add value," Ramakrishnan advised. "And two, I've been a change practitioner for 17 years, and we've known for a long time, you should involve people in the change as early as possible. You should involve the people being impacted and support their needs through it, and yet with AI, there's so many organizations that are just, I'm going to throw this at you, you're going to figure it out. It is a massive change, and we need to be thoughtful."
