Enterprise AI

Build enterprise AI systems

My team and I study the work, data, and existing systems before choosing a process change, conventional software, AI, or a combination that can be operated by the business.

The problem

Enterprise AI projects often begin before the underlying work is understood. The process may be undocumented, required data may be incomplete, or an existing system may already solve part of the problem. Some requests assume AI is required before the team has compared it with a process change or conventional software.

How I work

Our forward-deployed engineers work directly with the customer’s team, at the client’s location when needed. We document the process, inspect the data and systems, and define the result the business needs. I help choose the appropriate combination of process improvement, reused or repaired systems, new software, and AI. When an AI system is appropriate, we define its role, access, approvals, evaluation, operation, and ownership.

How this connects to my work

I lead engineering at KnackLabs. Our engineers use Symphony Forge to build and ship the software. When the solution calls for an AI employee, we use Gantry to operate it with defined access, human approvals and activity records. See the products I have built and led.

What an engagement covers

Problem and process discovery

Work with the people responsible for the current process to document tasks, exceptions, decision points, and the expected result.

Data and existing systems

Assess whether required data exists, whether it is usable, and whether current systems should be reused, repaired, integrated, or replaced. A new system of record is built only when the work requires one.

Solution choice

Compare process changes, conventional software, and AI against the required outcome, cost, risk, and operating constraints.

AI system design

Define runtime, memory, tools, identity, access, and trust boundaries for work that benefits from an AI component.

Tools and approvals

Limit the systems and actions available to the AI system. Require human approval for selected actions based on business risk and policy.

Evaluation

Test representative cases, expected failures, and important boundaries before release and when the system changes.

Operations and ownership

Define deployment, monitoring, cost reporting, escalation, recovery, and a named owner who can operate and maintain the system.

Related writing

• The cheapest model can cost you more

  A cheaper model can need more attempts to finish the same task. Compare the full cost, including thinking effort, caching, and review time.