AI enters the lab - innovations and highlights at Co:Lab 2026

NEWS  - eHealthNews editor Rebecca McBeth

AI was a hot topic at Co:Lab this year with speakers highlighting patient concerns about its use in clinical care, the importance of selecting the right combination of tools and the need to upskill the workforce to take advantage of them.

"If we keep coming back to what matters to the patient, we will create a far better healthcare system than we currently have.”

National chief medical officer Helen Stokes-Lampard flagged growing patient anxiety around artificial intelligence in clinical care and spoke about genomics and precision medicine as significant opportunities for the health sector.

She described a case where a hospital patient tried to sign a consent form refusing any AI involvement in their care. 

She said task-specific AI is integrated into all levels of the health system and the patient's real concern was about sensitive health data being processed by large language model platforms such as ChatGPT.

Stokes-Lampard pointed to the ‘black box’ problem, where the reasoning behind AI outputs cannot be fully understood, as a key concern for regulators and health leaders globally.

She highlighted the opportunity of genomics and personalised medicine for New Zealand, saying the Māori and Pacific populations give the country a distinctive genomic profile. 

"We are uniquely placed to provide a massive contribution to that world genomic database of information," she said.

"This is not about AI being done to the organisation, it is about people in the organisation leaning into AI to help the tasks that they undertake."

Spark Health chief executive John Macaskill-Smith told attendees that the organisations making the most progress with AI are not buying off-the-shelf products, but training their own people to build and manage AI agents themselves. 

He said one of the Big Four firms has set a target of matching its 40,000 human consultants with an equivalent digital workforce, creating a one-to-one ratio of human to AI agents. 

"Rather than being threatened by the opportunity, why don't you reach out and have a go at it and start to explore what the opportunities might be?" he said.

For New Zealand's pathology sector Macaskill-Smith identified four workflow areas suited to agentic AI: pre-analytical errors that lead to incorrect test orders, patient uncertainty about where results are and what they mean, critical result follow-up, and the management of complex ‘send-away’ testing.

“5 years ago the microscope was my main tool… now I have ChatGPT Enterprise and openevidence.com open on the desk and use it all day, every day.”

David Roche anatomical pathologist at Awanui Labs gave examples of what AI can do in pathology labs right now:
- Detecting and grading tumours in prostate, breast, and colon cancer cases
- Identifying cancer spread into lymph nodes
- Flagging slide quality issues before they reach the pathologist

He told attendees that multimodal AI models are also beginning to combine imaging, genetics, clinical records, and patient data into integrated diagnostic outputs.

When it comes to challenges and risks, Roche said training data quality really matters, accountability frameworks need to be in place before AI makes a consequential error, deskilling is a genuine risk and the environmental cost of AI infrastructure is also significant.

"The order in which you use these tools, and how you actually implement this with people, is so incredibly important.”

Poor workflow design can neutralise even the most capable AI tools, but a well-structured combinations of models - including older, traditional ones -  consistently outperform single-model approaches, Colby Raley told co:lab 2026.

Raley is a strategist in Microsoft’s NZ public sector team and spoke about a diagnostic study using GPT-4 where clinicians given access to the tool achieved 74 percent diagnostic accuracy, while the AI working alone reached 92 percent.

Another pathology study of more than 100 slides showed a different risk: clinicians without confident diagnoses were unduly swayed by AI recommendations of negative results.

"That is the flip side, the risk of building in too much confidence," Raley said. 

"There is quite a lot of consideration around how you want to bring these tools in and how you need to upskill the people that are going to be using them."

She said organisations working with low-resource languages such as te reo Māori found that traditional, purpose-trained models outperformed generative AI tools, and mixed-language conversations required hybrid approaches combining both model types.

"It is not just about the model, it is how you put things together," said Raley.

The Co:Lab conference is for the pathology sector and was held in Wellington on September 2-3, 2026.

To comment on or discuss this news story, go to the eHealthNews category on the HiNZ eHealth Forum 

If you would like to provide feedback on this news story, please contact the editor Rebecca McBeth.

You’ve read this article for free, but good journalism takes time and resource to produce. Please consider supporting eHealthNews by becoming a member of HiNZ, for just $17 a month.

Read more AI & Analytics news

Return to eHealthNews.nz home page