Sunday evening. Hana opens her task list at Sandline. One price needs her say-so. The PDF is still waiting beside it.
“Do we need an agent for this?”
AI agents vs workflows comes down to who picks the next step. Let's build the same order three ways and see where Hana belongs.
AI agents vs workflows: who chooses the next step?
A workflow is a route you've drawn in advance. It can branch when a price differs. An agent chooses some next actions for itself.
Think of following a delivery route versus finding a failed delivery. One has agreed stops. The other means choosing what to check next.
Anthropic draws this split: workflow code sets the path; an agent's model chooses steps and tools. AI calls can sit inside a fixed workflow.
One PO, three ways to read it
Palmgate Clinics sends Sandline Supplies PO 7731: a one-page PDF. Sara would normally retype its six lines into a sales order.
Four lines have clean references: helmets, vests, cartons and floor cleaner. Two need a closer look.
| PO detail | What Sandline must check |
|---|---|
| Lines 01 to 04 | References match the catalogue |
| Line 05: “Gloves nitrile M 100pcs”; 40 boxes at AED 28.00; no reference | Intended match: NG-200. Its name, “Nitrile gloves, box of 100”, doesn't confirm size M |
Line 06: TPE-48; 30 rolls at AED 6.00 |
Catalogue: AED 6.50. Arithmetic passes: 30 × 6.00 = 180.00 |
| Order total | AED 2,760.00 plus AED 138.00 VAT = AED 2,898.00 |
Download the fictional six-line PO 7731 (PDF)
Each build needs the same result: a known customer, supported products, agreed units and checked prices. A person must settle disputes before sales order lines exist.
Rules: follow the map
Karim maps fields in a known layout. Exact references become proposed lines. A kept-up alias list can map agreed wording to a SKU.
Lines 01 to 04 pass. Line 05 stops unless its wording has an alias. The price check catches line 06.
Rules handle exact codes and sums without guessing. The catch is upkeep. Changed layouts need new rules; unknown layouts should stop.
AI in a workflow: read, then check
In Beezr, Load data puts Products into a list. The AI reads the PDF with that list in its instructions. It returns proposed lines in a set shape.
The path stays fixed: read, propose, check, review, create accepted lines. The gloves may match NG-200, but Sara must still confirm size. Line 06 goes to approval on a branch the builder set.
The catch: a filled field can be wrong. Check the customer, quantity, unit and product identifiers.
An agent: choose the next lookup
Picture a hypothetical agent checking products and their attributes. It reads each result, then chooses whether another lookup would help.
This suits a search whose next move varies. The UK government's paper describes agents choosing from given capabilities to reach a goal.
Here, the agent can propose a fix. It cannot change catalogue prices or approve its own exception. PO 7731 requires procurement approval for substitutes too.
Keep those limits in fixed controls. A prompt alone isn't the gate. The catch is testing each possible loop; tomorrow's route may differ.
What does each route cost you?
This compares designs, not measured speed. Anthropic says agents often trade time and cost for task performance. Measure your own work.
| Dimension | Rules-only | Workflow with an AI step | Agentic investigation |
|---|---|---|---|
| Control | Builder specifies parsing, mappings and branches | Builder specifies the sequence; the model proposes interpreted content | Model selects permitted investigative actions; fixed controls guard commitment |
| Variability handled | Layouts and aliases you explicitly support | Wording and documents within the configured AI task | Different sequences of investigation |
| Cost drivers | Rule maintenance and exception handling | Model use, checks and exception handling | Model use across loops and tools, plus testing the paths |
| Latency drivers | Parsing and lookups; human delay for unknowns | Model response plus lookups and review | Number of chosen iterations, plus tools and review |
| Failure handling | Stop on unmapped or invalid values | Reject invalid output or route a failed check | Stop on limits, missing evidence or disallowed actions |
| Auditability | Record which rule and branch ran | Keep the proposal and check results with the run | Keep the action sequence, tool results and final outcome |
Count the time to clear exceptions too. A fast answer that leaves Sara guessing can mean doing the job again.
Human in the loop: what can the person change?
Approval gives permission before an action. Review checks a field or match. Intervention stops or redirects work that's going off course.
A notice after the order is saved is a receipt. AWS's approval example pauses, sends approve and reject links, then follows the answer.
Sandline · example
Hana shouldn't have to solve the PDF twice
Line 06 fails the price check. The workflow waits and gives Hana a task.
Her record shows the PDF, source wording, proposed SKU and failed check. It keeps both prices: AED 6.00 and AED 6.50. She sees the choice: 30 rolls, AED 15.00 under list.
In Beezr, she can approve, reject with a reason or ask for clarification. She can also send it back to an earlier approval step. Approved and rejected paths stay separate; the Smart Script creates lines after approval.
Build this review record with modules, fields and a failed-check message. If you request the model's uncertainty, treat it as a hint. Only a score you've tested should drive a rule.
Which approach does your process need?
Ask what varies and what a wrong action would cost. Varied wording can fit a fixed process. A search that changes as you learn may suit an agent.
Give the model only the freedom the job needs
- If data and mappings are clear → use rules.
- If documents vary but steps are known → use AI within a workflow.
- If results decide the next lookup → consider an agent with limited tools and stopping rules.
- If wrong terms would matter → require review before accepting them.
Myth / Reality
“Adding AI makes a workflow autonomous.” AI can extract, classify or draft on a fixed path. Ask who controls the steps.
Our take
Our take. Start with the known process. Add freedom only when a bounded AI step leaves a proven problem unsolved. A bigger demo isn't a reason.
How do you test the result?
Give each build the same documents and catalogue. Use PO 7731, an incomplete order and one missing a customer identifier. Write the expected review routes first.
Anthropic's test guidance separates the transcript, or run record, from the outcome: the final state. Check both. Did the order exist, and did it wait for approval?
Test approved and rejected branches. Beezr's run history shows steps, timings and errors. Its step-by-step debugger helps find the failed stage.
How Beezr does it
An answer has a name; a decision has a step. Beezr saves the AI result under your chosen name. Later steps use it: a Decision with Smart Script conditions, an approval and a Smart Script to create records.
Key takeaways
- If you're technical: check who controls each action. Test final records and run evidence on the same cases.
- If you're not: ask what the approver sees and what rejection does.
- Count the full work of clearing exceptions.
Build the PDF-PO flow in Beezr
Keep the order process fixed. Give AI the reading step.
- Prepare the record. Add the PDF file field, proposed order and review fields.
- Load and read. Load Products into a list in the AI instructions. Request structured proposed lines.
- Check. Set a Decision with Smart Script acceptance checks.
- Pause. Send failed totals or unresolved matches to approval. Keep approved and rejected paths separate.
- Create lines. Put the Smart Script after checks and approval. Inspect run history in tests.
Questions readers ask
Agentic workflow vs AI agent: what's the difference?
“Agentic workflow” may mean fixed AI steps, a bounded agent inside a workflow, or a model choosing the whole path. Ask which.
Beezr labels its AI step “AI agent”. Here that means one model call on your fixed path. It returns an answer; later steps act.
AI agents vs AI automation: how do they relate?
AI automation covers both fixed workflows with AI and model-directed actions.
What are agentic workflow examples?
Sandline's fixed PO flow reads, matches, checks, reviews and creates lines. See document processing. The hypothetical agent above chooses its lookups.
How do Beezr approval tasks reach the right person?
Choose named people, a role, a record's user field or a Smart Script. Tasks can be claimed or reassigned, so someone can cover for Hana.
Set priority and an overdue limit in days. Late tasks show in red.
Are agents always slower or more expensive?
Don't assume so. Compare calls, tools, upkeep and human review. Loops add work; unresolved exceptions cost time too.
Sandline Supplies is a fictional example company.
Sources
- Building effective agents · Anthropic · 2024-12-19
- AI Insights: Agentic AI · UK Government · Undated; accessed 2026-10-17
- Deploying a workflow that waits for human approval · AWS · Undated; accessed 2026-10-17
- Demystifying evals for AI agents · Anthropic · 2026-01-09