AI agents vs. automation.

Automation is software that executes a defined sequence of steps. Same inputs, same outputs, every time. The intelligence sits with the engineer who wrote the rules. AI agents are software that is given a goal and a set of tools and decides — at runtime — which tool to use, in which order, with which inputs, and when to stop. The intelligence sits in the model.

The two are not competitors. They are different layers of the same stack.

Choose deterministic automation when the workflow is well understood, the inputs are structured (forms, APIs, databases), failure modes are predictable, latency must be low and cost per run must be near zero. Invoice routing, KYC document hand-offs, cron-driven data pipelines, scheduled reports, system-to-system integrations — all automation territory. Putting an LLM in the middle of these is usually negligence: slower, more expensive, less reliable.

Choose an AI agent when the workflow requires reasoning over unstructured inputs (emails, documents, web pages, conversations), the path branches in too many ways to enumerate, the system has to recover from surprises, or you would otherwise need a human to decide what to do next. Investor research, sales triage, support handling, deal screening, contract review, executive briefings, outbound personalisation — all agent territory.

The strongest systems use deterministic flows as the rails and agents as the conductors. A pipeline routes inbound documents deterministically; an agent reads the messy ones and decides what to do. A CRM fires triggers on rules; an agent drafts the personalised follow-up. Wrap consequential agent actions in a human approval step until evaluations show the agent is right often enough to act alone.

Sales operations. Lead routing, CRM hygiene, opportunity-stage updates and reminder cadences are deterministic automation — predictable, high-volume, near-zero cost. Drafting the personalised follow-up for a complex enterprise deal, after reading the prospect's recent product launches and the last six months of email thread, is agent work. Combine them: the rules engine fires the trigger, the agent writes the message, a human approves before send.

Finance and operations. Reconciling invoices to POs, scheduled reporting, anomaly thresholds on transaction data — classic automation. Investigating an exception, summarising the root cause from a chain of emails and ERP entries, and drafting the resolution memo — agent. The agent never touches the ledger directly; it produces a recommendation that an operator approves.

Investor relations and capital markets. Calendar and CRM hygiene, mandate-tag updates, distribution of standard decks — automation. Researching a new family office, mapping their thesis against the current fund, drafting the first-touch email — agent. This is exactly the split Raise Platform runs in production, and it is the dominant pattern for any judgement-heavy workflow.

Customer support. Ticket routing, SLA timers, macro responses for known issues — automation. Triaging an unfamiliar issue, summarising the user's history and proposing the right resolution path — agent. The user never sees the seam: the same interface fronts both layers.

Agents introduce a class of risk that automation does not: probabilistic outputs and emergent behaviour. The mitigation is structural, not aspirational. Every consequential action runs through a human approval step until evaluations show the agent is right often enough to act alone. Tools the agent can call are scoped narrowly and audited. Evaluations are continuous and blocking: a regression in agent behaviour holds the deploy. None of this is exotic — it is the same operational rigour any critical system requires, applied to a probabilistic core.

Built AI ships both — see AI Agents for production agents by business function and AI Automation for end-to-end automation and internal tools. In Raise Platform, both layers run side by side: deterministic data pipelines feed agents that reason about investors and outreach.