Deciding where to build your artificial intelligence (AI) agents has become one of the most consequential and confusing calls organizations need to make. Vendors on all sides, from hyperscalers to software-as-a-service (saas) providers to point AI solutions providers, are all clamoring to make the case for why you should build agents on their platforms. In our work building agents at Valoir and our conversations with customers who have successfully built agents and put them into production we found there are really five factors to consider:
Where does the workflow and data already live?
An agent that needs deep access to case records, entitlements, or customer history performs best and requires the least integration work when it's built on the platform that owns that data and those workflows. For example, an agent handling contact center escalations should live in something like Genesys; an agent touching the sales pipeline gains from proximity to Salesforce; agents dealing with IT tickets likely belong closer to ServiceNow. Point solutions can do the job with APIs and Model Context Protocol (MCP) servers, but every hop adds latency, cost, and friction. The challenge becomes a bit more nuanced when vendors, departments, and technology overlaps.
What are your skills and resources?
The comparative time to value and skill levels required to build, deploy, and manage AI over time are critical, so an important consideration has to be your existing skills and expertise both today and over time. A hyperscaler agent that took three strong engineers six weeks to build will still need someone watching it and managing it in production a year later. Platform-native agents shift a lot of that maintenance burden onto the vendor, which is part of why they're attractive to teams without a standing AI engineering function. You'll want to consider in-house capacity as a moving target, particularly as the technology continues to rapidly evolve, and recognize that if you haven't already rethought hiring or upskilling, you're behind.
How much do you want to depend on one vendor?
Agents built on platforms like Salesforce, ServiceNow, or Genesys are faster to stand up but tie your AI and automation roadmap to that vendor's release cadence, pricing, and model choices. Hyperscaler-built agents give you more control (usually) over model selection and portability at the cost of more engineering lift. We've found that for most organizations, the cost and burden of supporting and managing agents over time far outweighs the benefit of the build-on-hyperscalers approach unless you're actually an AI business. Point solutions might often be best of breed for a specific task but add another vendor to manage and another party with access to your data.
What is the vendor's cost structure?
Platform vendors often bundle agent capability into existing licensing, which looks cheap until you hit usage-based AI credits or per-seat premiums. Hyperscaler costs scale with compute and can be hard to forecast without real usage data. Point solutions tend to have the most transparent pricing but the least room to negotiate once you're dependent on them. Cost structure matters more than list price, but be warned: cost and price structures are continuing to evolve, and no vendor has nailed it yet. You'll want to balance your appetite for predictable pricing with your desire for pricing tied clearly to usage. Moving forward, we'll see more of a focus on value-based pricing with clear metrics and telemetry on not just token usage but relative value of that usage based on business outcomes. No one's really there yet.
What about governance, compliance, and trust?
Where the agent runs determines where data flows, what's logged, and who and what can audit decisions. Ask the hard questions about certifications, data residency, observability, and auditability as well as things like kill switches that will identify anomalies and pull the plug before AI goes off the rails. Any vendor that can't answer your questions with answers that are satisfactory and understandable to you should be out by default. AI is a trust game, and if you can't understand it, you can't explain it to your users, and they won't trust it.
In short, where to build agents depends on a number of factors that influence your time to value, overall total cost of ownership, potential benefit, and risk. Picking one platform for everything is likely the wrong bet. Luckily, moving forward there will be more agent portability and orchestration across systems and applications and the skills transferability between platforms will be less of an issue. In the meantime, focusing on agentic projects that deliver rapid time to value (think three to six months) will enable you to reap the benefits from AI without locking you into a particular agentic strategy in a rapidly evolving landscape.
Rebecca Wettemann is founder and CEO of Valoir.