Assistants, workflows, agents: what agentic AI means for a 30-person firm
Agentic AI is the buzzword of the year. Here is the plain-English difference between an assistant, a workflow and an agent, and the order to adopt them in.
"Agentic" is doing a lot of work in vendor marketing at the moment. Underneath it are three quite different things, and a business that confuses them either buys too much autonomy too early or dismisses the whole area as hype. Here is the distinction we use.
Assistants: answer from your information
An assistant answers questions and drafts text using your own documents, knowledge base, customer records or data. It reads; it does not act. A good one cites its sources and only sees what the person asking is entitled to see.
Examples: "What did we agree with this client in March?", "Summarise this 40-page report", "Draft a reply to this email in our house style."
This is the first thing most businesses should deploy, because the failure mode is mild (a wrong answer a person reads) and the governance is mostly about permissions you should already have.
Workflows: act, with rules and approval
A workflow takes a request, understands it, retrieves what it needs, applies your business rules, updates a system, creates a task and waits for a person to approve anything consequential. The model does the understanding; code does everything else.
Examples: new-client onboarding, document requests, supplier renewals, monthly reporting.
This is where the measurable time saving is. It is also where the audit trail matters, because the workflow changes systems of record.
Agents: decide what to do next
An agent is software that can read information, retrieve data, make recommendations, call APIs and execute actions, choosing its own sequence of steps to reach a goal. It is the thing the word "agentic" is usually pointing at.
Agents are useful for open-ended work where the steps cannot be written down in advance. They are also the hardest to govern, because the sequence is not fixed and the failure mode is an action rather than an answer.
The adoption order
- Fix permissions and data hygiene. Every AI system inherits them.
- An assistant over your own documents, in an environment matched to how sensitive they are.
- One workflow, measured.
- Agents, only where the steps genuinely cannot be scripted, and only with the controls above.
None of this needs a data-science team. It needs the estate operated properly, an honest view of which data is sensitive, and someone accountable for the log.
Want help with this in your business?
Talk to Foundry — we’ll talk through your situation, no obligation.