The Watson Weekly is sponsored by Avalara — the agentic AI platform automating global tax and compliance for leading eCommerce brands with simple plans, transparent pricing, and complete coverage across Shopify, Stripe, WooCommerce, BigCommerce, and more.

I went into our latest Watson Weekly webinar hoping someone would tell me how much a company should spend on agentic AI. Nobody did. I came away with three other lessons that turned out to be more useful.

My guests were Shawn Mandel, CTO of Dentalcorp, Jason Cottrell, CEO of Orium and president of the MACH Alliance, and Anatolii Iakimets, director of product marketing at KIBO.

Dentalcorp supports more than 650 dental practices. Shawn runs technology for all of them with a small team, so he doesn't get a dozen shots at this.

His first use case needs a real sponsor in the business and enough value to pay for what sits underneath it, meaning data access, permissions and an architecture the next use case can build on. Big banks and digital natives with deep engineering benches can experiment their way forward, which most mid-sized companies can't.

Shawn has stopped calling these projects pilots. Each one is a step toward a platform, and he wants them designed that way from day one. He's also been building his own AI applications since last November, because he won't ask his team to do something he doesn't understand himself.

Shawn asks who an agent works for, how it gets onboarded and how it gets switched off. Companies already answer those questions for their employees. Most haven't asked them about software.

Jason took it further. HR builds the systems, but every leader in the company still manages their own people, and he expects agents to go the same way. He described executive teams of eight or ten people all looking at the CIO to solve this, until each of them realizes the processes that need rebuilding are their own.

Jason would rather see a company running five agents on safe, permissioned access to its systems than one with hundreds and no controls. Anatolii told buyers to press vendors on audit logs and human review with the same energy they bring to use cases.

Shawn told me most agentic pitches hold up until he asks for a demo and an engagement plan, and then they tend to fall apart. He was fairly generous about it. The category is roughly ten months old, and nobody builds mature enterprise software in ten months.

Jason's advice was to expect a vendor to get you about 80% of the way and to budget for finishing the last 20% yourself. That only works when the use case is narrow. The projects he sees going wrong try to do too much, like spending $20 million on one general-purpose agent that does everything.

Anatolii warned against chasing agent counts, which reminded me of companies bragging about lines of code. Shawn's caution was more practical. Test your architecture now for the day a contract ends and you want to switch models, because by then untangling it gets expensive.

I still don't have the number I went looking for. Shawn's honest answer was that it depends on the team and data you already have. I suspect plenty of boards will push back on that, and I'm not convinced the CTOs are wrong.