An AI agent is software that pursues a goal over several steps and decides which tools to use, inside limits you write down. A workflow follows steps you defined in advance. Most SMBs need workflows first. An agent pays off where the work changes from case to case, with tests around it and a person approving what goes out.
I build these systems at LeaderNova, and many first conversations start the same way: someone has been sold an "agent" that turns out to be a chatbot with a new label. This guide is the answer I give them: a plain definition, how to tell whether you need an agent or a workflow, where agents earn their place in a small company, and how to keep them safe.
What is an AI agent?
An AI agent is a language model that works towards a goal, chooses its next step and calls tools to get there: it reads an inbox, searches the web, drafts a document, updates a record. Anthropic draws the line in its engineering guide: workflows run models and tools through predefined code paths, while agents direct their own process and tool use (Anthropic, December 2024).
The definition matters because the market is noisy. Gartner predicted that 40% of enterprise applications would include task-specific AI agents by 2026, up from less than 5% in 2025 (Gartner, August 2025). The same firm warns about "agent washing", vendors relabelling assistants, chatbots and RPA as agents, and estimates that only about 130 of the thousands of vendors claiming agentic AI are real (Gartner, June 2025).
So when someone pitches you an agent, ask two questions. Does it decide its own steps, or does it run a script? And what is it allowed to touch?
Workflow or agent: which do you need?
Start with the simplest thing that works. Anthropic recommends exactly that, adding complexity only when it improves results, and notes that the teams that succeed use simple, composable patterns instead of heavy frameworks (Anthropic, December 2024). McKinsey's review of a year of agentic deployments reaches the same place from the business side: low-variance, standardised work is often better served by rules-based automation, which is more reliable (McKinsey, September 2025).
| Automated workflow | AI agent | |
|---|---|---|
| Who decides the steps | You, in advance | The agent, within written limits |
| Best for | Repeatable work with known inputs: invoice intake, data sync, approvals | Work that changes from case to case: mail triage, research, first drafts |
| How it fails | Stops on an input it doesn't expect | Takes a plausible but wrong step |
| How you test it | Fixed test cases | A golden set of real cases, scored on every change |
| Running cost | Low and predictable | Higher, and it varies with the case |
| Where to start | AI business automation | Agentic AI |
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls (Gartner, June 2025). Choosing an agent where a workflow would do is a quick way to end up in that 40%.
Where do AI agents pay off in an SMB?
An agent earns its place when the input varies, the judgement behind it can be written down, and a good draft saves real time even if a person still checks it. In a small company that usually means jobs like these:
- Finance and investor relations. Preparing the investor update, researching a fund before a call, checking the story against the numbers.
- Marketing in your tone of voice. First drafts of pages, emails and proposals that already sound like you.
- Social content. Turning a talk, a case or a release into posts for each channel.
- Legal and admin. A first read of a contract against your standard positions, deadline tracking, routine paperwork.
- Operations follow-up. Chasing open tasks, preparing the weekly status, nudging the owner of a stalled item.
- The inbox. Triage, a morning digest and replies prepared as drafts.
McKinsey's first lesson is that business value comes from changing the workflow around the agent (McKinsey, September 2025). That is why I measure before building anything: hours spent, cycle time and errors on the job as it runs today, then the same measures after go-live.
How do you keep an AI agent safe?
The OWASP Top 10 for Agentic Applications 2026 lists the risks that are specific to agents. The first three are goal hijack, where someone steers the agent through content it reads; tool misuse; and identity and privilege abuse. The list goes on to memory poisoning, cascading failures and rogue agents (OWASP, December 2025). Its core principle is least agency: give an agent only the autonomy its job needs.
In practice I apply four rules:
- Agent identity. Each agent signs in as itself, with its own credentials and permissions, short-lived and scoped to its job, never with a founder's login. Revoke the identity and the agent stops.
- Least agency, through context engineering. It reads only what its job needs: specific drive folders, the mail it is CC'd on, one calendar. For each task it sees only the facts, memory, records and instructions that task needs. Sensitive data, such as HR files or health data, stays out of its reach.
- Human in the loop. Emails, documents, payments and posts stay drafts until someone signs them off, and so does any change to a system of record. When the agent is unsure, it hands the case to a person, and each correction becomes a new test case.
- Everything is logged. Every tool call is recorded, so you can see what the agent did, why, and under which identity.
Where the agent runs matters as much. We deploy agents in your environment, on your cloud and your accounts, or as private AI on your own servers, hosted where you need it (the EU, the US or a specific country) in line with GDPR, CCPA and the other laws that apply. LeaderNova builds the agent, proves it against the tests and hands over the runbook, the tests and the logs. You own it and you run it. For the regulatory side, see the AI Act for SMBs.
Why one agent per job?
A single assistant that does everything is hard to test and harder to secure, because it needs access to everything too. We organise agents horizontally, by job: a finance and investor-relations agent, a marketing agent that writes in your tone, a social content agent, a legal and admin agent, an operations follow-up agent. Each has its own identity, its own access and its own tests.
The team reaches each agent where it already works: in chat, with an @mention in a thread, or by making the agent the owner of a task in your workplace tool. Nobody has to learn a new app.
What keeps this maintainable is the architecture underneath, which the company owns:
- An agentic framework. The architecture every agent runs on: connections to your tools, orchestration, approval gates, evals and logging. It is built from simple, composable parts rather than a heavy third-party framework, and it replaces the personal assistants that nobody maintains.
- A knowledge layer. Agents read the company's approved ways of working and facts from one versioned store, with the source of each fact. Personal notes become shared only after someone reviews them.
- A memory layer. What agents remember across tasks and conversations, such as case history, decisions and preferences, with written rules on what is kept, for how long and who can see it. It is kept separate from the knowledge layer.
- Your systems of record. Agents read from and write to the CRM, the accounts, the ERP, the ticketing tool and the documents, which stay the single source of truth. There is no shadow database, and every agent action can be traced there.
- Tests on real cases. Every agent capability is tested against a golden set of 20 to 30 real cases, before go-live and on every change.
- Orchestration last. Agents hand work to each other only once each one has proven itself alone, and we pilot the hand-off on a single case first.
What does this look like in practice? An operations-heavy engineering company
An operations-heavy engineering company asked us for AI assistants for its team. We started with an Ops & AI audit, and the brief set the conditions: production-grade from day one, data under the company's control, a person approving everything outbound, and no new per-user AI subscriptions.
What is delivered so far:
- email triage and a morning digest with draft replies, live and tested against a golden set of real cases;
- separate AI assistants consolidated onto one agentic framework, with a shared knowledge layer and memory, each capability tested against real cases;
- specialised tools for documents, CRM and business intelligence, inside the tools the team already uses;
- a knowledge layer of approved ways of working, kept apart from personal notes;
- access rules that keep sensitive data out of the AI's reach;
- an agent identity and data-access model for each agent.
The details are in the case study.
How do you start with AI agents?
- Audit and baseline. An Ops & AI audit maps where the time goes and records hours, cycle time and errors for the jobs in scope.
- Pick one job. Choose work that changes from case to case. If it doesn't, build a workflow: it is cheaper and more reliable.
- Write the limits. What the agent reads, what it can do, and what needs a person's approval.
- Build the golden set. 20 to 30 real cases, scored before go-live and again on every change.
- Go live with drafts. The agent prepares, a person approves, and you measure against the baseline.
- Add the next agent. Orchestration comes after the single agents have proven themselves.
You don't need per-user AI subscriptions for any of this if you'd rather not pay for them.
If you are weighing an agent for a specific job, start from the work itself. We'll tell you whether it needs an agent or a workflow, and what it takes to run it safely in your environment. See how we build AI agents
Sources
- Anthropic, "Building effective agents", 19 December 2024: anthropic.com
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025: gartner.com
- Gartner, "Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025", 26 August 2025: gartner.com
- McKinsey QuantumBlack, "One year of agentic AI: Six lessons from the people doing the work", 12 September 2025: mckinsey.com
- OWASP GenAI Security Project, "OWASP Top 10 for Agentic Applications for 2026", 9 December 2025: genai.owasp.org
Last checked: 2 November 2026.