FluxForce AML and Fraud Blog

AI Agents vs Automation: Who Weighs, Who Decides

Written by Sahil Kataria | Nov 19, 2026, 6:56:58 PM

Listen To Our Podcast🎧

• 1:26
AI Agents vs Automation: What's Different
Secure. Automate. – The FluxForce Podcast
Your browser does not support audio.

Who weighs the case: a script or an agent?

AI agents vs automation comes down to who weighs the case. A script follows the steps you wrote and does the same thing every time. An AI agent gathers evidence, weighs the case and prepares a recommendation, and that changes what you need to record.

Most firms will run both. Gartner advises using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval. The hard part is knowing which of your processes is which, and what has to change when a process moves from a script to an agent.

This guide gives COOs at regulated firms a side-by-side table across decision traits, the evidence behind it and a checklist for the switch. Last updated 19 November 2026.

In This Article, You'll Learn
  • How agents, assistants and scripts differ on who weighs, who decides and what's recorded.
  • Which jobs should stay with rules or RPA, and which are worth an agent.
  • What "checks before judgment" means.
  • Which laws care about the difference, and from when.
  • What to check before you replace a script with an agent.

Request the side-by-side table across decision traits

Bring one process you're thinking of moving from a script to an agent. We'll go through the decision traits with you and show you what an agent prepares for a case in shadow mode, with a person making the call.
Get the side-by-side table across decision traits

The side-by-side table across decision traits

Compare scripted automation, AI assistants and AI agents on who weighs the case, who decides and what gets recorded, then put agents only where a case needs weighing.

  1. List the decisions in a process and mark the ones a rule already makes well.
  2. Mark the exceptions that need evidence from several systems and a judgment.
  3. Run your rules first as hard checks, then let an agent weigh only what passes (checks before judgment).
  4. Send every recommendation to a named reviewer and record why each case went the way it did.
  5. Widen autonomy only for an item type where shadow mode results match your reviewers and internal audit can replay the record.
"
Direct answer

AI agents vs automation comes down to who weighs the case. A script follows the steps you wrote and does the same thing every time. An AI agent gathers evidence, weighs the case and prepares a recommendation for a person to decide, so it needs a record of why. Most firms will run both, with rules first.

AI agents vs automation across decision traits

Scripted automation means rules engines, robotic process automation (RPA) bots and workflow tools that follow steps someone wrote in advance. An AI assistant answers a person's question. An AI agent works through a task across several systems and prepares an outcome.

  • Scripted automation runs each case through a fixed rule and acts on the result. An AI assistant answers a question, and the person decides what to do with the answer. An AI agent gathers evidence, weighs the case and prepares a recommendation. The recommendation goes to the reviewer, who makes the call.
Scripts follow a rule, assistants answer, agents weigh and prepare, and a person decides.
Decision trait Scripted automation (rules, RPA, workflows) AI assistant AI agent
What it's given A fixed sequence of steps A question from a person A goal, tools and a mandate
Who weighs the case Nobody; the rule fires or it doesn't The person asking The agent gathers evidence and weighs it
Who decides The rule's author, in advance The person A reviewer, unless the firm has let the agent act alone for that item type
Same input, same output? Yes Not always Not always, so the record matters more
What's recorded That a step ran, and when Usually the chat, if anything What it looked at, which checks ran, why it recommended what it did, and who decided
How it fails Breaks loudly when a screen or format changes Gives a wrong or made-up answer Takes a wrong step that later steps build on
Who's accountable The firm The firm The firm
How you stop it Switch off the bot or the rule Stop using it A kill switch for each agent, with every stop recorded
Best fit High-volume work with clear rules Lookup and drafting Exceptions and cases that need a decision

Three rows matter most for a COO. "Who weighs the case" is where the agent adds value, "what's recorded" is what you'll be asked for when a decision is questioned, and "who's accountable" never moves. In Moffatt v. Air Canada, a 2024 decision, the British Columbia tribunal said it makes no difference whether information comes from a static page or a chatbot, and Air Canada was responsible for both.

The first figure shows the three lanes side by side. Scripted automation runs each case through a fixed rule and acts on the result. An AI assistant answers a question, and the person decides what to do with the answer. An AI agent gathers evidence, weighs the case and prepares a recommendation. The recommendation goes to the reviewer, who makes the call.

What's the AI agent and automation difference?

The AI agent and automation difference starts with inference. The EU AI Act defines an AI system as a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions. A rule doesn't infer anything. US bank supervisors draw the same line. Their April 2026 model risk guidance excludes deterministic rule-based processes from the definition of a model.

RPA sits on the scripted side of that line. IBM puts it simply: RPA is process-driven and AI is data-driven.

Agents sit on the other side. Anthropic describes agents as systems where the model directs its own process and tool use, and contrasts them with workflows that follow predefined code paths. OpenAI's guide to building agents gives a payments example. A traditional rules engine works like a checklist, flagging transactions on preset criteria, while an agent evaluates context and subtle patterns even when no clear rule is broken.

If you're comparing "agentic AI" with single agents, our post on agentic AI vs AI agents covers it, and our agentic AI glossary entry gives the short definition. For accountability, the line that matters sits between rules that don't weigh anything and software that does.

Watch the labels. Gartner warns of "agent washing", the rebranding of AI assistants, RPA and chatbots without real agentic capability. It estimates only about 130 of the thousands of agentic AI vendors are real.

AI agents vs traditional automation: who weighs and who decides

Traditional automation is strong where the rule is the decision. If a payment over a threshold always needs a second approver, a rule should route it every time. The Wolfsberg Group's 2025 statement on monitoring says rules-based methods are effective for well-defined risks.

  • Each case goes through your hard checks first. A case that fails one of the hard checks goes straight to the reviewer with the check that fired. Cases that pass the hard checks go to the agent, which weighs them against your own history. The agent sends its recommendation and explanation to the reviewer. The reviewer makes the call, and the reason goes into the evidence record.
Rules run first as hard checks, the agent weighs what passes, and a person makes the call.

Rules struggle where the case needs weighing. In 2021, FATF said reliance on a rules-based approach is one of the contributing factors for defensive and over-reporting. A 2018 Bank Policy Institute survey found that participating banks reviewed about 16 million alerts in 2017 and filed over 640,000 suspicious activity reports. In a 2024 letter to FinCEN, banking trade groups cited BPI survey data showing banks spend 21.41 hours for every SAR filed.

Agents bring judgment, and judgment brings variation. A Gartner analyst wrote in November 2025 that AI agents can deliver different outcomes to the same question, unlike logic tables or decision trees. Swapping rules for agents would trade one problem for another, so run them in order instead.

Checks before judgment is our name for that order. It's the FluxForce version of what the industry calls decision intelligence. Your thresholds, policies and rules run first as hard checks. Then the agent weighs the case with models trained on your own history and precedent from similar past cases. A language model only drafts and summarizes, and a person makes the call.

The second figure shows checks before judgment. Each case goes through your hard checks first. A case that fails one of the hard checks goes straight to the reviewer with the check that fired. Cases that pass the hard checks go to the agent, which weighs them against your own history. The agent sends its recommendation and explanation to the reviewer. The reviewer makes the call, and the reason goes into the evidence record.

The Wolfsberg Group describes a similar mix, with large language models producing investigator-ready summaries that clarify the underlying risk.

Key insight: let the rules make the decisions they already make well, and give the agent only the cases a rule can't settle, with a person deciding each one.

What gets recorded, and why it matters

A script's record answers "did it run?" An agent's record has to answer "why?"

  • Scripted automation writes a bot log of what it did and when. An AI agent writes an evidence record of what it looked at, which checks ran and why it recommended what it did. The bot log lets internal audit confirm the script ran as configured. The evidence record lets internal audit replay why each case was decided the way it was.
A bot log shows a script ran. An evidence record shows why a case went the way it did.

COSO's 2024 guidance on RPA asks whether audit logs are maintained for all bot operations, and whether bots have a formal change management process with approval, testing and documentation. ISACA wrote in 2022 that a bot needs its own system credentials so its activity can be tracked.

Agents take more steps, so they need more. The Financial Stability Board's October 2026 sound practices note that an AI agent can take hundreds of intermediate steps and make errors in any of them. They say traceability should let a firm follow what triggered an agent's action, what information and tools it accessed, what steps it took and whether it stayed within its permissions. The FCA's October 2026 good practice examples include firms that kept records of prompts, outputs, model versions, filters and escalation decisions.

Under the EU AI Act, high-risk AI systems must allow automatic recording of events over their lifetime. Deployers must keep those logs for at least six months, and financial institutions keep them inside the documentation their financial services law already requires.

The third figure compares the two records. Scripted automation writes a bot log of what it did and when. An AI agent writes an evidence record of what it looked at, which checks ran and why it recommended what it did. The bot log lets internal audit confirm the script ran as configured. The evidence record lets internal audit replay why each case was decided the way it was.

The gap between what was decided and what you can show later is the replay gap. With scripts it's small, because the rule is the reason. With agents it can be large unless the record is built in from day one.

AI agent vs automation workflow: which one fits which job?

An automation workflow follows a path someone drew, and an AI agent picks its own path. For most departments the answer is a workflow with an agent where a case needs weighing. IMDA's agentic AI framework says most agentic implementations in enterprise deployments are already hybrids of deterministic rules and some autonomy.

The builders of these models say the same. Anthropic recommends finding the simplest solution possible, which might mean not building an agentic system at all. OpenAI's guide says that if a use case doesn't meet its criteria for an agent, a deterministic solution may suffice. IMDA says not all use cases are suitable for agents, and some may be better served by deterministic workflows.

Cost points the same way. AI agents could cost 30 times more than RPA bots for high-frequency process executions, according to the same Gartner analyst. He also says most organizations aren't replacing their RPA or business process automation tools. Forrester expects ROI and governance challenges to keep most organizations running deterministic automation through 2026.

If the job looks like this Use Why
Same input, same answer, thousands a day Rules or RPA Cheap, predictable, easy to audit
Fixed steps across several systems Workflow automation The path is known; a person handles exceptions
An exception that needs evidence from several places and a judgment AI agent inside a workflow The agent weighs and prepares; a reviewer decides
A decision that's irreversible or legally significant Agent prepares, person decides every case High risk always goes to a person

To sort one process, we recommend three passes:

  1. Mark every decision a rule already makes well, and leave it with the rule.
  2. Mark the exceptions that need evidence from several systems and a judgment. Those are the agent's work.
  3. Mark anything irreversible or legally significant, and keep a person on every one of those cases.

Finance is a good test. Ardent Partners puts the average invoice exception rate in accounts payable at 18.4%, and says exceptions are typically the biggest single reason AP benchmarks aren't lower. Exceptions are the minority of invoices, and they're the ones that need weighing, so they're where an agent earns its cost. Fraud raises the stakes. In the AFP's 2025 survey, 79% of organizations reported attempted or actual payments fraud in 2024.

AI vs RPA: what changes for accountability and the law

In the EU, GDPR gives people the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects. In the UK, Article 22A of the UK GDPR, in force since February 2026, says a decision counts as solely automated where there is no meaningful human involvement in taking it. The EU's new anti-money laundering regulation will require meaningful human intervention in decisions made with AI or automated processes to enter, refuse or maintain a business relationship with a customer. That regulation applies from 10 July 2027.

For high-risk AI under the EU AI Act, the person overseeing the system must be able to disregard, override or reverse its output. They must also be able to interrupt it through a "stop" button or a similar procedure that brings it to a halt in a safe state. Regulation (EU) 2026/1744 moved those duties, and the logging duties above, to 2 December 2027 for Annex III systems.

In the US, the same April 2026 bank model risk guidance puts generative and agentic AI outside its scope, and its model definition leaves out rule-based processes. Neither your bots nor your agents sit inside today's US model guidance, so your own controls have to fill the gap. The FTC said in 2024 that there is no AI exemption from the laws on the books.

Singapore's financial supervisor is more specific about stopping AI. In its October 2026 guidelines on AI risk management, MAS describes kill switches as mechanisms to deactivate AI quickly if it exceeds risk tolerances.

The FCA says firms are responsible for the AI systems they develop, procure and deploy, including their outputs and outcomes. No rule lets a firm hand accountability to software.

What does this look like in practice?

The scenario below follows one approval process from a script to a script plus an agent. It uses supplier bank-detail changes because the rule there should never move, and the judgment around it is where the time goes.

Illustrative scenario: supplier bank-detail changes at an insurer

Illustrative scenario, not a customer result.

This scenario is invented to show the process, not drawn from a real customer.

A 900-person insurer's accounts payable team gets a few dozen supplier bank-detail change requests a month. An RPA bot holds every change and opens a callback task, but the queue is long and nobody can see which requests look odd.

Nothing the script did was removed, and a person still makes every call.

  1. The COO and finance director agree the rules stay exactly as they are: every change is still held, and no payment moves to new details without a callback.
  2. An agent is added after the hold. It gathers the supplier's history, recent invoices, the request's sender and any lookalike email domains.
  3. In eight weeks of shadow mode, the agent prepares a risk note and priority for each request while the team works the queue as before, and the team records where it disagrees.
  4. The finance director lets the agent sort the queue by its priority. The callback and the approval stay with a person for every request.
  5. At year end, internal audit replays a sample of changes from the evidence record and sees what the agent looked at and who approved each one.

Who decides what when an agent replaces a script?

Moving work from a script to an agent moves the judgment. Accountability stays with the firm and the people it names.

Many firms already draw this line. KPMG's third-quarter 2026 survey found 49% of leaders have defined high-risk use cases where they don't allow autonomous decision-making by agents. In Deloitte's 2026 State of AI in the Enterprise, close to three-quarters of companies planned to deploy agentic AI within two years, yet only 21% of those companies reported a mature model for agent governance. The FSB says firms should agree a risk tolerance for each agentic use case at the start, which informs how much autonomy and access they grant.

What are the roles and boundaries when an agent replaces a script?

Rules keep making the decisions they already make well; the agent weighs exceptions and prepares the case; the reviewer decides; the process owner sets the mandate; risk or compliance sets the risk bands; IT operates the kill switch.

ComponentResponsibilityBoundary
COOPicks which processes stay scripted and which get an agent.Doesn't approve individual cases.
Process ownerWrites the agent mandate by risk band and keeps the hard checks current.Doesn't let a vendor write the mandate.
Reviewer or approverDecides each case the mandate sends them, using the prepared file.Doesn't approve without the evidence record.
Risk or complianceSets the risk bands and checks shadow mode results before autonomy widens.Doesn't run day-to-day cases.
IT and securityGives each agent and bot its own identity, the minimum access and a tested kill switch.Doesn't widen access without the process owner.
AI agentsGather evidence, run the checks, weigh the case and prepare a recommendation.Don't approve payments, reject claims or file reports.

The reviewer decides every case the mandate sends them, high risk always goes to a person, risk or compliance decides when autonomy widens, autonomy beyond preparing the case is set by the firm band by band, and any agent can be stopped with its kill switch.

How does FluxForce approach AI agents and automation?

FluxForce is the platform your departments run AI agents on. The agents investigate and prepare the work, a person makes the call, and every step leaves evidence an examiner can replay.

  • Take in passes the department's records to understand. Understand passes one assembled view of the case to weigh. Weigh passes the checked and scored recommendation to prove. Prove passes the evidence record to report and learn.
Take in, understand, weigh, prove, report and learn, with governance and data protection on every step.
Public stage What happens
Take in Agents connect beside the ERP, core system or case tool you already run, and nothing is migrated
Understand The agent assembles the supplier, claim or customer into one view before anything is scored
Weigh Your rules run first as hard checks, then the agent scores the case and recommends, with an explanation
Prove Every step goes into a tamper-evident evidence record with its inputs, time, actor and the rules of that day
Report and learn The agent drafts the report, a person approves it, and models change only after a person approves

Governance controls apply on every step. Each agent works inside a mandate you set, high risk always goes to a person, and every agent has a kill switch. Data protection applies on every step.

The fourth figure shows the five stages. Take in passes the department's records to understand. Understand passes one assembled view of the case to weigh. Weigh passes the checked and scored recommendation to prove. Prove passes the evidence record to report and learn.

Financial crime and compliance is the first app line on the platform: AI agents that investigate AML, sanctions, fraud and KYC alerts and prepare the case. Your analyst makes the call, with evidence an examiner can replay. You decide how much each agent does on its own, and every one has a kill switch.

Once agents sit beside your scripts, someone has to know which model version made each recommendation and who approved its last change. That's the job of FluxForce AI Model Governance, which keeps one inventory covering our agents' models and your own.

To see which agents run on the platform today, start with our AI agents page. Our guide to AI agents in finance and where they fail covers where autonomy stops inside a regulated firm.

Which sources support this comparison?

Gartner: over 40% of agentic AI projects will be canceled by end of 2027

Agents for decisions, automation for routine workflows, assistants for retrieval; agent washing of RPA and assistants; about 130 real agentic AI vendors; cancellation forecast and reasons.

Analyst forecast and advice (June 25, 2025), not measured data. Checked in the 16 November 2026 Wayback copy (gartner.com returns 403 to scripts).

2026-11-19

Regulation (EU) 2024/1689 (AI Act), consolidated 27 July 2026

AI system defined by inference; automatic logging for high-risk systems; six-month log retention; override and stop-button duties; Annex III credit and recruitment uses.

EU law; high-risk duties for Annex III systems apply from 2 December 2027.

2026-11-19

OCC Bulletin 2026-13 (with Federal Reserve SR 26-2)

Model definition excludes deterministic rule-based processes; generative and agentic AI out of scope.

US bank supervisory guidance, April 17, 2026; nonbinding.

2026-11-19

Saikat Ray (Gartner) via CXOToday

Agents could cost 30x more than RPA bots for high-frequency executions; most organizations are not replacing RPA or BPA; agents can give different outcomes to the same question.

Analyst article republished by CXOToday, November 17, 2025.

2026-11-19

Forrester, Predictions 2026: Automation At The Crossroads

ROI and governance challenges keep most organizations on deterministic automation through 2026.

Analyst prediction, November 3, 2025.

2026-11-19

Financial Stability Board, agentic AI sound practices

Agents take hundreds of intermediate steps; traceability of triggers, tools, steps and permissions; risk tolerance per use case at inception.

International standard-setter publication, October 15, 2026.

2026-11-19

Wolfsberg Group, Monitoring for Suspicious Activity Part 2

Rules-based methods effective for well-defined risks; hybrid approach; investigator-ready summaries.

Industry body statement, 2025; financial-crime monitoring.

2026-11-19

COSO, internal control over robotic process automation

Audit logs for all bot operations; formal change management for bots.

Internal control guidance, 2024.

2026-11-19

Regulation (EU) 2024/1624 (AMLR)

Meaningful human intervention in AI or automated customer-relationship decisions.

EU law applying from 10 July 2027.

2026-11-19

Moffatt v. Air Canada, 2024 BCCRT 149

A company is responsible for all information on its website, whether from a static page or a chatbot.

British Columbia Civil Resolution Tribunal, February 14, 2024; not binding precedent elsewhere.

2026-11-19

Where does FluxForce AI Model Governance fit?

FluxForce AI Model Governance keeps a live record of every model your agents and your own teams run. Every model version is tracked, compared and watched for drift once it's in production, and if performance slips you roll back to the previous version in one click.

For a COO moving work from scripts to agents, that answers the question an auditor asks first, which is which model made a decision and why. Release needs the model owner's approval, every rollback is logged with who did it and why, and there's a kill switch on every agent.

See how AI Model Governance works

FluxForce sits beside your RPA and workflow tools rather than replacing them, and its agents don't approve payments, reject claims or file reports. AI Model Governance doesn't decide which processes get an agent or approve a release: the COO, the model owner and your risk team make those calls. Any autonomy beyond preparing the case is yours to switch on, band by band. If the process you want to change sits outside financial crime, ask which agents exist for it today.

What should a COO check before replacing a script with an agent?

  1. List the decisions in the process. Leave the ones a rule already makes well, and mark the exceptions that need evidence and a judgment as the agent's work.
  2. Write the agent mandate by risk band, and name the reviewer for each band and the kill switch owner.
  3. Give the agent its own identity and the minimum access the mandate needs.
  4. Run shadow mode on live cases and compare the agent's recommendations with your reviewers' decisions every week.
  5. Ask internal audit to replay a sample of cases from the evidence record before autonomy widens for any item type.
  6. Test the kill switch and record the test.
MetricDefinitionDecision guardrail
Recommendation matchShare of shadow mode cases where the agent's recommendation matched the reviewer's decision.Widen autonomy only for an item type with explained differences.
Exception handling timeTime from exception raised to the reviewer's decision, before and after the agent.Never trade a missing record for speed.
Replay rateShare of sampled cases internal audit could replay without asking anyone.Any case that can't be replayed is a finding.
Kill switch testA recorded test that each agent stops at once when switched off.No agent goes live without a passed test.
Key takeaways
  • Scripts follow fixed steps; AI agents weigh the case and prepare a recommendation; a person decides.
  • Keep rules and RPA where the rule is the decision; add an agent where a case needs evidence and a judgment.
  • Run rules first as hard checks, then let the agent weigh what passes (checks before judgment).
  • An agent needs an evidence record of why it recommended what it did; a log that it ran is enough only for a script.
  • Your firm stays accountable for scripts, assistants and agents alike.

Conclusion

Scripts follow steps and agents weigh cases, so the choice is where judgment belongs in each process. Keep rules where the rule is the decision, add an agent where a case needs weighing, and keep a person on every call your firm will be asked to defend.

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Before you replace a script, check what the agent costs next to the script it sits beside, which decisions it actually improves and which controls it will run under.

Request the side-by-side table across decision traits

Bring one process you're thinking of moving from a script to an agent. We'll go through the decision traits with you and show you what an agent prepares for a case in shadow mode, with a person making the call.
Get the side-by-side table across decision traits

Frequently Asked Questions

Automation follows steps someone wrote in advance and gives the same output for the same input. AI infers an output from the data in front of it, so it can handle cases no rule anticipated. The EU AI Act's definition of an AI system turns on that ability to infer. An AI agent goes further and works through a task, choosing its steps and tools.

Neither is better in general. RPA is better for high-volume work with clear rules, because it's cheaper and more predictable. A Gartner analyst estimates agents could cost 30 times more than RPA bots for high-frequency work. AI agents are better for high-value, lower-frequency cases where traditional automation adds little. Most firms will use both, with rules running first.

AI can be part of an automated process, but they aren't the same thing. Much automation uses no AI at all, and US bank model guidance treats deterministic rule-based processes as outside its definition of a model. When AI is added, the process starts weighing cases, and you need records of why.

Not at most firms. A Gartner analyst says most organizations aren't replacing their RPA or business process automation tools. Agents are being added where scripts stop, usually at the exceptions.

AI agents are AI systems under the Act's definition, so they're in scope. Whether the high-risk duties apply depends on the use, such as credit scoring or HR decisions. Those duties apply to Annex III systems from 2 December 2027. Scripts that don't infer anything fall outside the definition.

About the author

Sahil Kataria

Founder and CEO of FluxForce

Sahil works on secure AI and financial technology for regulated industries. His engineering background spans identity verification, payment security and compliance automation. He has led teams across Africa, the United States, Europe and India.

At FluxForce.ai, his focus is on explainable AI and auditable financial workflows. He writes for banking and compliance teams about the practical decisions behind these systems: how to assess risk, keep controls visible and introduce automation without losing accountability.

View Sahil Kataria author profile →
AI Model Governance

Primary solution: every model version tracked, compared and replayable.

AI agents

The AI agents on the platform today.

AI agents in finance: what they do and where they fail

Where autonomy stops inside a regulated firm.

Agentic AI vs AI agents

The live post that owns the "agentic ai vs ai agents" keyword.

Kill switch: definition

How every agent can be stopped.

Human-in-the-loop review

The review control behind every case an agent prepares.