Enterprise AI Agents
Most companies have agents in a chat window and nothing in production. The work that pays for itself sits inside your systems: the ticket queue, the CRM, the invoices nobody wants to reconcile. I build agents that do that work under defined permissions, plus the orchestration layer that keeps them coordinated and lets you see what they did.
Enterprise AI agents are software agents that execute multi-step work inside a company's own systems: reading tickets, updating records, drafting responses, escalating to a human, all under permissions someone defined on purpose. AI agent orchestration is the layer that keeps a fleet of them coordinated, observable, and safe. It routes each task to the agent that should own it, retries what fails, and hands off to a person where a mistake would be expensive. Oleg Sotnikov designs and ships that layer: the agents, the integrations into your systems, and the guardrails around both. He runs AppMaster on a two-person team with agents in production, processing 11B+ tokens a month. AI agents for business earn their keep on exactly that discipline: defined permissions, visible behavior, and a human wherever the cost of error is high.
If all you need is the phone answered, that is a simpler and cheaper product: the AI receptionist ships as a fixed $10,000 implementation. This page is about agents that work across your internal systems, which is scoped per engagement. See the AI receptionist package
What Gets Built
Six parts of a working agent system. Most engagements need all six, and the order follows wherever the work already hurts.
Workflow agents inside your systems
The agent works where the work already lives: CRM, ERP, ticketing, internal APIs. An MCP server per system does the connecting, so each system exposes one described interface instead of a bespoke integration per agent.
Agent orchestration
Routing tasks to the agent that should own them, retries and reroutes when something fails, state that survives a restart, and clean handoffs between agents and people. One agent is a demo; a fleet needs a supervisor.
Permissions and guardrails
Least privilege per agent, scoped credentials that expire, and an audit trail of every action with the input that caused it. My open-source sallyport vault follows the same pattern: the agent gets the operation, never the key.
Knowledge grounded in your data
Answers come from your documents, tickets, and records rather than from what a model absorbed in training. Retrieval quality, freshness, and who is allowed to see what matter more here than the choice of model.
Human-in-the-loop by design
Approval gates in front of anything expensive to get wrong: refunds, contracts, messages a customer will read. The agent prepares the work, a person signs it off, and gates come down later only when the logs have earned it.
Observability and evals
A trace of what each agent did and why, dashboards for cost and failure rate, and evals that run on every change. Agent behavior drifts quietly, so it needs the same instrumentation as any other production system.
How It Works
Map the workflow and permissions
We take one real workflow end to end: who does what today, which systems it touches, and what an agent may and may not do in each. You get a design with the permission model already decided, not a slide deck.
Build and orchestrate against real cases
Agents get built against your actual tickets, records, and edge cases rather than a clean demo set. Orchestration, retries, and human handoffs arrive as soon as there is more than one agent in play.
Run behind evals, then hand over
Nothing reaches customers before traces, cost dashboards, and evals are in place. Your team gets the code, the dashboards, and a runbook; I stay on as fractional CTO only if you want the ownership to stay with me.
Why Me
- AppMaster runs on an AI-first team of two people, down from 25, with agents in production processing 11B+ tokens a month at 99.99% uptime
- I build and maintain open-source agent infrastructure: sallyport, a Mac vault that runs authenticated actions for AI agents over MCP, so a credential never has to sit inside a prompt
- 25+ years integrating enterprise systems, starting in enterprise IT in the 2000s, across 1,000+ projects and 9 startups founded
Related Work
The pieces around an agent program, each with its own page.
Frequently Asked Questions
What are enterprise AI agents?
Enterprise AI agents are software agents that carry out multi-step work inside a company's systems instead of only answering questions in a chat window. A support agent reads the ticket, looks up the account, updates the record, drafts the reply, and escalates anything outside its remit. What makes them enterprise-grade is the permission model, the audit trail, and the evals around them; the choice of LLM matters less than people expect.
What is AI agent orchestration?
AI agent orchestration is the layer that coordinates several agents and the people working alongside them. It routes each task to the agent that should handle it, retries or reroutes failures, keeps state across steps that may take minutes or days, and hands work to a human at defined points. Without it you have a collection of demos that each work alone, and no view of what the system did as a whole.
How are custom AI agents different from an AI receptionist?
The AI receptionist is a productized single-purpose agent: it answers the phone, books meetings, and qualifies leads, and it ships as a fixed $10,000 implementation. Custom enterprise agents are a different shape of work — several agents across several systems, your permission model, your workflows — so they are scoped per engagement. If phone answering is the whole job, take the package: it costs less and ships sooner.
How do you keep AI agents safe?
Least-privilege credentials so each agent can only touch what its job requires, approval gates in front of actions that move money or reach a customer, an audit trail of every action with the input that triggered it, and evals that run on each change so behavior drift surfaces before customers meet it. Credentials stay out of the agent entirely: my open-source sallyport vault runs the authenticated action and returns the result, so the agent gets the operation and never the key.
How much do enterprise AI agents cost?
Scoped per engagement, because the cost sits in your systems and permissions rather than in the model. Most companies start with the $5,000 Team & AI Audit or an AI proof of concept on a single workflow, which puts a real number on the build before anyone commits to it. Ongoing agent work usually runs inside a fractional CTO engagement at $5,000–10,000 a month.
Put Agents Where the Work Is
Bring one workflow you would like to hand over. Half an hour is usually enough to tell whether an agent belongs there and what it would take to run it safely.
If agents are the wrong tool for that workflow, I will say so on the call.
Related reading
Agents in production, MCP integration, and what it takes to run them safely.
