DigitalTransformationTalk: Empower AI agents to drive business processes at scale

On 15 April 2025, DigitalTransformationTalk host Kevin Crane was joined by Ayman Husain, Customer Engineering Leader, Google;Christian Hull, Chief Technology Officer, Tax Canary; and Sathya Sethuraman, Field CTO, Camund.

Views on news 

While traditional task automation tools (such as RPA) excel at isolated functions, it takes a lot more integration and co-ordination to keep a business process running efficiently from start to finish. That’s where end-to-end process orchestration comes in. Process orchestration makes automation not only efficient, but also resilient and adaptable to changing needs. In the old days, if you wanted a change in software, you had to raise a ticket and a change took ages.

 

Today, technology is more democratised but changes in technology and governance, thankfully, are still closely aligned. Without orchestration businesses can create the most diabolical technical risks you will ever have. While earlier you knew what outcome to expect, with AI, you don’t have to have that outcome understood ahead of time. Instead, the AI system will ask itself questions if it sees something that hasn’t happened before. There are one or two more things not covered by the article. First, integration is not out-of-the-box ready yet and it shouldn’t be oversimplified. It will invariably be complex and painful. Secondly, some businesses will find integration too complex and costly and will refrain from AI deployment. Thirdly, there is also the fear of losing jobs to AI that can withhold adoption. 

Why small AI prototypes fail on an enterprise level?

70-90% of pilots fail to reach production due to technical hurdles and organisational barriers or a loss of momentum. Typical issues include the lack of good quality data, the high cost of integration and governance shortfall. Although failure rates are high, it’s also an upside that unviable projects fail fast, as this will save the company money. It’s often naivety that prevents projects from succeeding. It takes several years of a seasoned professional’s career to learn things about coding or why governance really matters. Businesses must also stop thinking about technology and the business as silos, which is 1990s thinking. Instead, you must rebuild everything around the product. AI and genAI allow businesses to build new prototypes really quickly. And as these technologies are good at scaling, it’s usually the business or governance issues that stand in the way of upscaling. Organisations have huge amounts of data – some of it unstructured or overstructured or sitting in a vault that no one has access to. The problem often is though that data is too expensive or the RoI is missing. 

Although transparency is important in the context of AI models, a Stanford survey has shown that the transparency scores among AI companies is only 37 out of one hundred. Iterative communication between business and technical teams during MLOps is also key. You also need a data owner, who knows everything about a data set and is aware of confidentiality issues. Data scientists must get support from the business to get a full understanding of enterprise data. They will be aware of the high costs of computing.  RPA had the same issue in its heyday – it was too expensive, which prevented wider adoption.

Even with automated processes, there are points where it’s humans making a probabilistic decision. AI is today the only technology that allows us to bring automation to workflows that couldn’t be automated previously. 

The panel’s advice

• With the most sophisticated genAI models, to answer one question can cost as much as $6,700, while these models are still on par with the performance of a graduate, who will do the job at a much more reasonable price.
• Big tech is good at innovating fast, while big enterprises are best at killing innovation by governance.
• You always need industry specific use cases. There is no AI for all.
• Divide your processes into deterministic and probabilistic. Automate deterministic workflows with rule-based algorithms.
• Create sandboxes to learn AI by doing.