Monday morning. Sara has a customer PO open and the product list beside it. A delivery question waits. “Can't we give this to an agent?”
What is agentic AI going to choose here? And when should it ask Sara? Let's follow one order, then build the bounded version in Beezr.
What is agentic AI in simple terms?
Agentic AI is AI with room to choose actions toward a goal. Think of a stockroom assistant asked to prepare an order. They can check shelves and ask about missing items.
That doesn't give them permission to change the customer's deal.
The UK government's paper contrasts this with software that follows instructions set in advance. The agent picks among the abilities it's been given. Ask which choices it can make, rather than counting AI calls.
Agentic AI vs generative AI: what changes?
Generative AI creates content, such as text or images. That's the academic definition. A convincing answer alone doesn't show that software can act.
An AI agent is software that chooses and takes actions toward a goal. Agentic AI is the system built from one or more agents. That split follows the UK paper.
| Term | What you're describing | Purchase-order example | The catch |
|---|---|---|---|
| Generative AI | A content-producing capability | Produce a proposed list of order lines | A plausible list still needs checking |
| An AI agent | Software choosing actions toward a goal | Search products, inspect a result, choose another lookup | Every available action needs a clear boundary |
| Agentic AI | A system built from agents, plus its goal, limits and checks | Coordinate the lookups and checks toward an order ready for review | The whole system needs an owner, including failed runs |
These aren't stages you have to climb. You can use content generation in a fixed workflow and stop there. Or give an agent a small research job, with approvals elsewhere.
Names vary. Anthropic's distinction is useful: workflows follow paths set in advance. Agents let models direct their process and tool use.
How does an agentic AI loop work?
An agent acts, gets a result, then chooses its next move. Anthropic's account stresses tool feedback and clear stopping rules.
Here's a draft agent for Palmgate's PO. Sandline's Beezr example below uses a fixed workflow with one AI step.
An agent's loop around PO 7731
- Gather — read the six-line PO and fetch possible product list matches.
- Decide — choose another allowed lookup or offer a match.
- Act — run the lookup or return the draft mapping.
- Check — compare the result with the goal and business checks.
- Ask a person — stop for a price difference or weak product evidence.
Palmgate wrote “Gloves nitrile M 100pcs”. An agent might search, check packaging, then ask for more detail. Its job is to find a match it can support.
Sandline lists NG-200 as “Nitrile gloves, box of 100”. That's the intended item here, but the product list says nothing about size M. Flag the “M” for Sara; don't invent a size or substitute.
What does this look like in Beezr?
Sandline · example
Sara gets a proposal, then the workflow checks it
Palmgate Clinics sends a six-line PDF PO. Sara attaches it to Sandline's record in Beezr. A Load data step puts the product catalogue in a workflow list.
The AI step reads the PDF with that list in its instructions. It returns draft order lines matched to catalogue products. A later Decision runs the builder's checks.
A price difference goes to approval. An unclear match goes back for detail. Only after the checks and approval does a Smart Script create sales order lines.
Beezr calls this step “AI agent”. It asks a language model and returns text or structured data. The workflow decides what happens next.
| AI step setting | Choices |
|---|---|
| Provider and model | Builder-selected, from OpenAI, Anthropic, Google or OpenRouter |
| Files | PDF, Word, Excel, PowerPoint, text and HTML |
Sara gets draft matches. Hana gets a route for price exceptions. Karim can check failed workflow runs.
How Beezr does it
Put the check in its own step. Save the AI answer under a chosen name for later steps. Follow it with a Decision or approval task, since the model can be wrong.
An approval task makes the workflow wait. A yes and a no take their own branches.
Agentic AI vs automation: which fits?
| Approach | Good fit | Honest catch |
|---|---|---|
| Rules-only automation | Known fields, explicit mappings and stable checks | New wording needs a mapping or a person |
| Workflow with an AI step | Variable documents inside a known approval process | Extraction and matching remain fallible |
| Agentic execution | Investigation where each result changes the next useful action | You must bound actions and evaluate different possible paths |
Give choice only to the part that needs it
- If fields and IDs are agreed → use rules and parsers. There may be nothing to interpret.
- If documents vary but the process is known → use a workflow with AI. Let AI suggest; let checks decide.
- If each discovery changes the next useful action → consider a bounded agent. Set allowed actions and stops first.
- If a wrong action could commit too much → require approval first. Later review can't undo every consequence.
Start with the product record. Does it state the unit, pack and traits the customer asked for? A prompt won't fill gaps in the stock list.
Then define “done”. Sandline needs checked customers, matches and prices, not just six filled lines. A correct refusal counts too.
Keep reading apart from committing to an order. A right to look up products isn't a right to create orders. Enforce this where actions run, outside the prompt.
Keep the proposal, failed check and source document together. Beezr's run history shows steps, timings and errors. Approval history shows who decided.
The UK paper also stresses tests, sound data and checks on how the system runs.
What should you test before trusting it?
Use tests where you know the right answer. Try an exact SKU and odd wording. Add a missing unit, a size with no evidence and a price mismatch.
Should each test offer a match, ask or stop? Write that down first.
Check the final records and actions, not just the words. Anthropic's guide to tests treats the outcome and the run transcript as distinct things. Run tests more than once: results can vary.
For Sandline, the right lines must reach the right state. Price review must stay in place.
Don't trust instructions in files sent to you. They may try to turn an agent from its task: that's prompt injection. A PO supplies data; it can't rewrite rules or grant tools.
Myth / Reality
“An agentic system should need no people.” You choose which actions get freedom. An agent can research alone and stop before a price exception.
Build that stop in.
Our take
Our take. Write the job description first. What may AI read or offer? Which action needs a person?
More tools won't settle vague answers.
Key takeaways
- If you're technical: find what chooses the next action. Enforce limits outside the model's answer.
- If you're not: ask to see a failed match and rejected approval.
- A workflow with an AI step may be enough.
- Measure outcomes, including the right stop for review.
Put an AI step inside a governed workflow
Build Sandline's document flow in Beezr, starting with one process.
- Create records — add PO and sales order line modules. Include a PDF file field.
- Load Products — put the product list in a list for the AI instructions.
- Ask for draft lines — set fields, objects and a list as the answer shape. Save a named result.
- Check and approve — use Smart Script conditions in a Decision. Route exceptions to approval.
- Create accepted lines — run Smart Script after the required branch passes. Check run history during tests.
Questions readers ask
What is agentic AI in simple terms?
AI with some choice over actions toward a goal. State what it may do and when it must stop.
Agentic AI vs AI agents: what's the difference?
Here, an agent is the software participant. Agentic describes the system's freedom to act. Check the actual control flow, since names vary.
Agentic AI vs generative AI: can they work together?
Yes. Content generation can help an agent choose actions and use results.
What are agentic AI examples?
Our bounded investigator picks the next product list lookup from what it finds. A fixed extract-check-approve chain is a workflow under this guide's distinction.
Is agentic AI the same as automation?
It's one approach. Rules follow a designed path; agents choose some actions at runtime. See AI agents vs workflows.
Sandline Supplies is a fictional example company.
Sources
- AI Insights: Agentic AI · UK Government · Undated; accessed 2026-10-17
- Generative AI · Feuerriegel, Hartmann, Janiesch and Zschech; arXiv · 2023-09-13
- Building effective agents · Anthropic · 2024-12-19
- Mitigating prompt injections in browser use · Anthropic · 2025-11-24
- Demystifying evals for AI agents · Anthropic · 2026-01-09