Agentic AI vs Automation vs Chatbots: The Decision Framework

Overview

• Chatbots are best for answering, routing, and basic interaction.
• Automation is best for stable, repetitive, rule-driven workflows.
• Agentic AI is best for dynamic, multi-step work that requires reasoning and action.
• Human-in-the-loop control is best for high-risk decisions and exceptions.
• Most enterprises need a hybrid architecture, not a single tool.
• ROI usually improves when automation handles the deterministic steps and agents handle the variable ones.
• Governance matters as much as capability. NIST and HBR both emphasize oversight, trust, and risk controls for modern AI systems (HBR).

Market Context: Why this debate matters now

Enterprises want execution, not demos

The market has moved beyond “cool chatbot” thinking. Executives now want systems that save time, reduce labor, cut errors, and improve service levels. That is why the discussion has shifted from conversational AI to operational AI.

In practical terms, the center of gravity has moved from interface innovation to workflow redesign. A lot of first-wave AI spending went into chat layers, copilots, and proof-of-concept demos that looked impressive in meetings but did not materially change how work moved across the business. That is no longer enough. Leadership teams now expect AI programs to produce one or more of four measurable outcomes: lower operating cost, faster throughput, better decision quality, or higher revenue conversion.

AI adoption is broad, but operational maturity is not

McKinsey’s State of AI 2024 shows strong adoption, but broad adoption does not mean strong execution. Many organizations are still stuck at the pilot or feature layer instead of redesigning workflows around the technology (McKinsey).

That gap between adoption and value realization is the real market story. Plenty of companies now have access to language models, RPA tools, AI search, or copilots. Far fewer have:

• restructured processes around those capabilities,
• instrumented the workflows for measurement,
• clarified governance and approval rights,
• and defined where deterministic automation ends and agentic reasoning begins.

McKinsey’s survey is useful because it highlights that adoption is not the same thing as operational transformation. A company can have AI in one or more business functions and still fail to capture durable value if the technology sits on top of unchanged processes.

This is where the decision framework matters. If a business misclassifies a workflow, it often creates one of two failure patterns:

1. It applies a chatbot to a task that actually needs multi-step action and orchestration.
2. It applies an expensive agentic stack to a task that could have been solved cleanly with workflow automation.

The economic upside is real, but it is uneven

McKinsey’s estimate of $2.6 trillion to $4.4 trillion in annual value is often cited, but the important point is not the headline number alone. The more useful insight is that value is concentrated in specific functions where language, decisions, and information flows dominate the work: customer operations, software engineering, marketing and sales, and R&D (McKinsey).

That matters because it helps leadership teams avoid generic AI strategies. Not every department has the same automation profile. Not every workflow benefits from autonomy. Some workflows are mostly retrieval and response. Some are process-heavy and deterministic. Others are exception-heavy and context-driven. The architecture should follow that reality.

A useful way to read the market is this:

• Chatbots monetize information access.
• Automation monetizes process stability.
• Agentic AI monetizes decision variability.
• Human-in-the-loop protects value in high-risk environments.

Gartner’s signal is strong, but so is its warning

Gartner identified agentic AI as a top strategic technology trend for 2025 and projects that by 2028, 15% of day-to-day work decisions will be made autonomously by agentic AI (Gartner). It also projects that 33% of enterprise software applications will include agentic AI by 2028 (Technology Magazine summary of Gartner).

Those are material signals. But the more important number for operators is Gartner’s warning that over 40% of agentic AI projects may be canceled by the end of 2027 because of poor ROI discipline, cost issues, and weak controls .

That tells you two things:

• adoption pressure is real,
• but indiscriminate rollout will fail.

So yes, agentic AI is strategically important. But it should be deployed selectively, on workflows where adaptability creates economic advantage over fixed automation.

Hyperautomation is still the foundation layer

It is easy to treat agentic AI as a replacement for automation. That is a mistake. In most enterprises, automation is still the foundation. Gartner has long positioned hyperautomation as a major market and has pointed to meaningful cost reduction when organizations automate and redesign processes together.

The practical implication is simple: do not skip the process layer. If your data is fragmented, approvals are unclear, systems are disconnected, and exceptions are unmanaged, adding agents will not solve the root problem. It will just automate confusion faster.

Hype is creating architecture mistakes

This is where leaders need discipline. Some vendors call every assistant an “agent.” That confuses the buying process. If the system cannot reason across multiple steps, use tools, manage state, and pursue a goal, it is not really agentic in any meaningful enterprise sense.

A good executive test is to ask four questions:

1. Can the system choose between actions based on changing context?
2. Can it interact with multiple tools or systems to move the task forward?
3. Can it recover from intermediate failures or missing information?
4. Can we observe, constrain, and audit its behavior?

If the answer is no, you may not need an agent at all. You may need a chatbot, a workflow engine, or a better integration layer.

Governance is now part of architecture, not a post-launch step

This debate also matters now because AI governance has moved from policy discussion into system design. NIST’s generative AI risk guidance makes that clear by focusing on inaccuracy, misuse, privacy, security, and trustworthiness across the AI lifecycle .HBR’s writing on AI supervision reaches the same conclusion from a management angle: the right level of human control depends on the business risk and should be calibrated rather than assumed .

That is why human-in-the-loop is included in this comparison. It is not just a safety feature. It is often the difference between a deployable system and a non-deployable one.

The near-term winners will be disciplined operators

The companies that will win this cycle are not the ones deploying the most AI features. They are the ones that:

• choose workflows with clear economics,
• decompose them correctly,
• assign the right architecture to each segment,
• add observability and governance from day one,
• and scale only after proving value.

That is the frame to use for the rest of this article.

Definitions: what each system actually is

Chatbots

A chatbot is a conversational interface. It accepts input and returns responses, usually through text or voice. It can be simple and scripted or powered by a large language model. Its core purpose is interaction.

Best use cases:

• FAQs
• lead capture
• support triage
• appointment booking
• internal helpdesk Q&A

A chatbot may call APIs, but that alone does not make it an agentic system.

Automation

Automation is the execution of predefined steps using rules, workflows, triggers, APIs, RPA, or orchestration tools. It shines when the process is stable and predictable.

Best use cases:

• invoice routing
• CRM updates
• employee onboarding tasks
• form processing
• report generation
• scheduled data syncs

If your main goal is process efficiency, this is the category to review first: AI Automation.

Agentic AI

Agentic AI is a goal-driven system that can plan, reason, retrieve context, use tools, take actions, and adapt its path based on what it learns. That is the important difference. It does not just answer. It moves the task forward.

Best use cases:

• claims triage
• prior authorization support
• logistics exception handling
• enterprise research workflows
• cross-system case resolution
• knowledge operations with incomplete context

For the deeper architecture behind this model, see our internal guide: What Is Agentic AI? Architecture, Use Cases & The Complete Guide.

Human-in-the-loop

Human-in-the-loop means a person reviews, approves, edits, or overrides AI output at defined checkpoints. This is not a weakness. It is a control mechanism.

Best use cases:

• lending decisions
• insurance approvals
• clinical summaries
• legal review
• financial operations with thresholds
• high-risk customer communications

The 4-column comparison table

Dimension	Chatbots	Automation	Agentic AI	Human-in-the-loop
Core purpose	Answer and guide	Execute fixed steps	Plan, reason, and act	Review and control outcomes
Best for	FAQs, routing, support	Repetitive workflows	Dynamic multi-step tasks	High-risk or ambiguous tasks
Input type	Natural language	Structured data and rules	Mixed structured and unstructured context	Any, where human judgment is needed
Decision style	Limited response logic	Deterministic logic	Adaptive reasoning	Human judgment supported by AI
Tool use	Light or optional	Predefined integrations	Core feature across systems	Secondary to review process
Variability tolerance	Low to medium	Low	High	High
Cost profile	Low to medium	Low to medium	Medium to high	Medium to high
Governance need	Moderate	Moderate	High	High
Common KPI	Containment rate	Labor savings	Autonomous completion rate	Risk reduction
Main failure mode	Wrong answer	Breaks on exceptions	Wrong action or hallucination	Slower throughput

What this means in the real world

Most enterprises should not pick one box and stop there. A mature operating model often looks like this:

• chatbot handles intake,
• automation handles routing and deterministic steps,
• agentic AI handles messy or variable work,
• human reviewers step in at risk checkpoints.

That is usually the practical architecture.Visual placement note: Insert one 16:9 comparison diagram here named with ALT text: “Agentic AI vs automation vs chatbots comparison of manual legacy systems and Agix approach”. Use dark background dark red for manual legacy systems, Agix orange for Agix components, and plain bold AGIX watermark at bottom-right.

Visual placement note: Insert one 16:9 flowchart here named with ALT text: “Agentic AI vs automation vs chatbots decision tree for architecture selection”.

Step 1: Is the task mainly about answering a question?

If yes, start with a chatbot. Do not overbuild.

Step 2: Is the workflow fixed and rules-based?

If yes, use automation. It will usually be cheaper, simpler, and easier to govern.

Step 3: Does the task require reasoning across multiple steps?

If yes, you are in agentic AI territory.

Step 4: Does the system need to take action across business tools?

If yes, simple conversational AI is not enough. You need orchestration and guardrails.

Step 5: Is the cost of being wrong high?

If yes, keep a human in the loop.

Simple decision tree in text form

1. Need answers only? → Chatbot
2. Need fixed execution? → Automation
3. Need planning + tool use + adaptation? → Agentic AI
4. Need compliance, approval, or expert review? → Human-in-the-loop

Examples: where each approach fits best

Chatbot example: customer support front door

A retail company gets thousands of repetitive customer questions:

• order status
• refund policy
• shipping times
• account help

A chatbot is enough if the goal is fast self-service and ticket deflection. IBM notes that AI support systems can improve service speed and consistency when designed correctly (IBM).

Automation example: invoice processing

A finance team receives invoices in structured formats, validates them against purchase orders, and routes them for approval. That is a classic automation case. You do not need an agent to “reason” through a mostly deterministic path.

Agentic AI example: claims triage

An insurance or healthcare workflow often includes:

• unstructured documents,
• policy interpretation,
• missing data,
• multi-step follow-up,
• tool calls across systems,
• exception handling.

That is where agentic AI adds value. It can retrieve context, evaluate rules in context, request missing information, and move the case forward.

Human-in-the-loop example: medical or financial approval

If the output affects diagnosis, claims approval, pricing, lending, or legal liability, do not fully automate the final step. Use AI to summarize, recommend, and prepare the case, but keep expert approval in place.

Detailed industry examples: healthcare, real estate, and logistics

Healthcare example 1: patient access and appointment triage

A multi-location healthcare group receives high volumes of appointment requests, insurance questions, specialty routing requests, and follow-up calls. The old workflow relies on phone trees, front-desk staff, and manual scheduling handoffs.

What works best

• Use a chatbot for symptom-neutral intake, FAQ handling, clinic hours, provider discovery, and appointment request capture.
• Use automation to update scheduling systems, confirm appointments, trigger reminders, and send forms.
• Use agentic AI only when routing depends on multiple variables like provider availability, insurance match, visit type, and prior patient context.
• Keep human-in-the-loop for anything that touches diagnosis, escalation based on risk language, or regulated clinical communication.

Why this matters
The front door of care is often a communication problem first, not an agent problem. A simple but well-integrated chatbot plus workflow automation can remove a lot of manual call load. Agentic logic becomes useful only when the routing becomes too variable for rules alone.

Healthcare example 2: prior authorization support

Prior authorization is one of the clearest examples of where fixed automation runs out of road. The workflow can involve:

• payer-specific rules,
• clinical documents,
• missing lab results,
• inconsistent forms,
• appeals,
• and back-and-forth coordination between staff and payers.

What works best

• Use automation for document collection, status updates, and submission packaging.
• Use agentic AI to review the case bundle, identify missing evidence, match documentation against payer requirements, and recommend next actions.
• Use human-in-the-loop for final submission decisions or disputed cases.

Case-style outcome logic
This is a strong candidate for manual work reduction because much of the burden is administrative and document-heavy rather than clinical decision-making. The agent is not practicing medicine. It is coordinating information work under human oversight.

Healthcare example 3: clinical summarization and physician review

Hospitals and provider groups often struggle with clinician time spent reading notes, discharge summaries, messages, and cross-system records.

What works best

• Use automation to gather records from EHR-connected sources.
• Use agentic AI to synthesize longitudinal context, summarize relevant findings, and flag missing information.
• Use human-in-the-loop because a clinician must remain responsible for interpretation and action.

Why this should not be fully autonomous
The issue is not whether the model can summarize. It often can. The issue is that medical action requires professional judgment, accountability, and review. This is a classic assistive use case, not a fully autonomous agentic ai.

Real estate example 1: inbound lead qualification

A real estate business receives buyer and seller inquiries across website forms, ads, chat, calls, and listing platforms. Leads vary widely in seriousness, budget, location, financing readiness, and timeline.

What works best

• Use a chatbot to capture inquiry details 24/7 and answer common questions.
• Use automation to push lead data into CRM, assign agents by geography, and trigger nurture campaigns.
• Use agentic AI when qualification requires contextual follow-up, mortgage-readiness signals, multi-property comparison, or next-best-action suggestions.
• Use human-in-the-loop when advice approaches contractual, financial, or negotiation territory.

Case-style benefit
The chatbot handles speed. Automation handles consistency. Agentic logic improves qualification quality. Human agents stay focused on relationship and deal judgment.

Real estate example 2: document-heavy transaction coordination

Real estate closings involve contracts, addenda, disclosures, lender updates, inspections, title work, and multiple stakeholders. There is a lot of repetitive coordination, but also a lot of exception handling.

What works best

• Use automation for checklist management, reminders, form routing, and milestone notifications.
• Use agentic AI to review transaction status, identify missing documents, draft stakeholder follow-ups, and sequence next actions.
• Use human-in-the-loop for legal interpretation, contract changes, and negotiation-critical messages.

Why agentic AI helps here
Traditional automation works well until one party misses a step, a document comes back incomplete, or the timeline shifts. That is where agentic AI adds value by adapting the follow-up logic instead of just firing prebuilt reminders.

Real estate example 3: portfolio operations and tenant communication

Property managers handle maintenance requests, lease questions, escalations, vendor coordination, and compliance communication across many units.

What works best

• Use a chatbot for tenant FAQs and service request intake.
• Use automation for work-order creation, dispatching, rent reminder workflows, and status updates.
• Use agentic AI to prioritize cases, summarize multi-thread tenant history, recommend vendor actions, and coordinate resolution paths.
• Use human-in-the-loop when disputes, legal notice, safety, or high-value tenant retention issues arise.

Case-style outcome logic
This is a strong operational intelligence use case because the value comes from reducing fragmented coordination rather than generating content for its own sake.

Logistics example 1: shipment exception management

Logistics teams live in exception mode. Delays, routing issues, customs documents, inventory mismatches, and customer escalations can all break a fixed plan.

What works best

• Use automation for standard status updates, milestone notifications, POD capture, and routine ticket creation.
• Use agentic AI to detect exceptions, retrieve context from TMS/WMS/ERP systems, propose alternate routing or next actions, and coordinate communications.
• Use human-in-the-loop for cost-sensitive rebooking, customer commitments, or cross-border compliance issues.

Why this is not a chatbot problem
Customers may interact through chat, but the real business challenge is operational decisioning across systems. This is exactly where agentic AI can outperform a simple conversational layer.

Logistics example 2: warehouse issue resolution

A warehouse team deals with damaged goods, pick errors, stock discrepancies, labor bottlenecks, and urgent reallocation decisions.

What works best

• Use automation for scan-based workflows, task assignment, and inventory state updates.
• Use agentic AI to analyze discrepancy patterns, pull related shipment or inventory history, propose resolution paths, and create action queues.
• Use human-in-the-loop when shrinkage, safety, or high-value inventory exposure is involved.

Case-style benefit
This is an operations use case where the agent’s value comes from compressing decision time, not replacing the physical workflow.

Logistics example 3: customer service for B2B freight accounts

Large B2B customers do not just ask “Where is my shipment?” They ask:

• why did a milestone slip,
• what can be done next,
• what is the impact on downstream delivery,
• who is responsible,
• and what recovery option is available.

What works best

• Use a chatbot for first-line status questions.
• Use automation for standard alerts and case creation.
• Use agentic AI to synthesize shipment data, customer commitments, exception history, and available recovery options into a usable resolution path.
• Use human-in-the-loop for account-critical decisions, penalties, or SLA-sensitive commitments.

Why this matters
The difference between a good freight experience and a bad one is often not visibility alone. It is exception response quality. That is a decisioning problem, not just a messaging problem.

Cross-industry pattern to remember

Across healthcare, real estate, and logistics, the pattern holds:

• Use chatbots where speed of interaction matters.
• Use automation where the path is stable.
• Use agentic AI where the path changes by case.
• Use human-in-the-loop where judgment, compliance, or risk cannot be delegated.

Implementation Guide

1. Map the workflow first

Before picking the technology, document:

• trigger
• inputs
• outputs
• systems touched
• exception paths
• approval points
• current handling time
• failure cost

This step matters more than most teams think.

Do not start with the model. Start with the workflow map. In most organizations, what looks like “an AI opportunity” is actually a mix of:

• communication delay,
• manual data movement,
• fragmented knowledge,
• missing decision rules,
• and weak exception handling.

Unless you identify those elements clearly, you will buy the wrong architecture.

2. Break the workflow into segments

Classify each step:

• conversational
• deterministic
• judgment-heavy
• approval-required

This usually reveals that the answer is hybrid, not singular.

A good decomposition exercise often looks like this:

• front-end intake → chatbot or form-driven workflow
• standardized routing → automation
• exception analysis → agentic AI
• high-risk decision → human approval

This segmentation also helps with cost control. You do not want an expensive reasoning layer doing work that a simple rules engine can perform more reliably.

3. Define the target operating metric before the technology

Track the baseline:

• average handling time
• manual touches per case
• cost per transaction
• escalation rate
• SLA misses
• rework rate

Without baseline measurement, ROI claims are just noise.

Add one more layer: pick the metric that matters most for the business unit. That might be:

• throughput,
• cost-to-serve,
• approval time,
• first-contact resolution,
• inventory recovery speed,
• or claim cycle time.

If there is no primary metric, the project will drift into generic “AI transformation” language and become hard to govern.

4. Assess data and systems readiness

This is where many programs slow down. Before implementation, validate:

• source systems and APIs,
• document quality,
• data freshness,
• identity and access constraints,
• event triggers,
• auditability,
• and ownership of business rules.

Agentic systems in particular depend on reliable context. If the knowledge base is stale, APIs are incomplete, or document quality is poor, the system will produce unstable results no matter how good the model is.

5. Decide the right system pattern

At this stage, ask directly:

• Does the user just need an answer?
• Does the workflow follow a fixed path?
• Does the system need to reason across changing context?
• What is the financial or compliance cost of being wrong?

Then assign the pattern:

• chatbot for answer-first interaction,
• automation for deterministic execution,
• agentic AI for adaptive multi-step work,
• human-in-the-loop for protected decisions.

This is the core decision framework of the article. Keep the assignment strict.

6. Add architecture controls early

Agentic systems need more than prompts. They need:

• retrieval
• tool permissions
• policy constraints
• fallback logic
• observability
• audit trails
• approval routing

Use design-time controls, not just policy documents. For example:

• restrict which tools the agent can call,
• constrain financial thresholds,
• require approval above risk levels,
• log each step taken,
• capture model confidence or evidence trace,
• route unclear cases to humans.

NIST’s AI RMF guidance is especially relevant here because it treats risk management as an operational lifecycle activity rather than a one-time review (NIST).

7. Build the evaluation framework before launch

Do not wait until after deployment to decide how the system will be judged. Create test sets for:

• standard cases,
• edge cases,
• missing-data scenarios,
• exception-heavy cases,
• compliance-sensitive outputs,
• and escalation triggers.

For agentic workflows, evaluate:

• action accuracy,
• tool-call correctness,
• policy adherence,
• completion quality,
• time-to-resolution,
• and human override frequency.

This is the only way to know whether the system is actually helping operations or just producing plausible-looking output.

8. Pilot narrowly, scale carefully

Pick one workflow with:

• clear value,
• enough volume,
• accessible systems,
• and an accountable business owner.

Then expand based on evidence, not enthusiasm.

The best pilot workflows usually have:

• meaningful manual burden,
• enough repeatability to measure,
• enough variability to justify the chosen architecture,
• and clear economic value if performance improves.

If you are exploring broader deployment patterns, review our Agentic AI Systems service page.

9. Design the human operating model

This is usually ignored in early projects. You need to specify:

• who reviews exceptions,
• who owns policy changes,
• who monitors performance,
• who handles incident response,
• who can override the system,
• and who signs off on scaling.

This matters because successful AI deployment is not just model performance. It is operating discipline. If nobody owns the exceptions, the queue becomes the new bottleneck.

10. Plan for phased autonomy, not instant autonomy

A better rollout model is:

• Phase 1: AI drafts or recommends
• Phase 2: AI executes low-risk steps
• Phase 3: AI handles defined exceptions
• Phase 4: AI operates semi-autonomously within policy boundaries

This reduces risk and helps teams build trust based on evidence rather than promise.

11. Use industry context, not generic templates

The right pattern depends on the domain. Healthcare, fintech, insurance, logistics, retail, and real estate each have different tolerance for error and different workflow shapes. For one industry example, see our Healthcare AI solutions.

A good implementation plan in healthcare will not look the same as one in logistics. The difference is not just regulation. It is workflow topology:

• healthcare is document-heavy and approval-sensitive,
• real estate is communication-heavy and milestone-driven,
• logistics is exception-heavy and system-coordination driven.

So yes, reuse patterns. But do not copy architectures without mapping industry constraints first.

12. Scale only when value and controls both hold

The real scale gate is not “the pilot worked once.” The scale gate is:

• ROI is measurable,
• exception handling is stable,
• governance works,
• business owners trust the system,
• and human review load is sustainable.

If those conditions hold, expand horizontally into adjacent workflows. If not, fix the operating design before scaling.

That is also why many organizations start with AI Automation and then add agentic capability later. It creates a cleaner foundation for controlled autonomy rather than forcing autonomy onto a broken process.

Real-world metrics that actually matter

Executives do not buy AI categories. They buy results.

The useful metrics are operational:

• manual work reduction
• cycle time compression
• exception rate
• human override rate
• containment rate
• cost per case
• quality and compliance performance

At Agix Technologies, the target is usually straightforward: remove friction from the workflow and make the economics work. In the right use cases, that can mean:

• up to 80% less manual work
• faster turnarounds
• fewer escalations
• lower operating cost
• quicker deployment in 4–8 week

Gartner’s market view supports the efficiency focus. Hyperautomation remains a priority for about 90% of large enterprises, but fewer than 20% are mature at measuring it effectively. That gap between adoption and measurement is exactly where many projects fail.

Limitations and risks

Chatbot limitations

• Often shallow
• Weak at long-running tasks
• Can frustrate users if retrieval is poor
• Tends to look smarter than it really is

Automation limitations

• Breaks when the process changes
• Poor fit for ambiguity
• Needs clean inputs and clear logic
• Can create maintenance overhead across systems

Agentic AI limitations

• Higher compute and orchestration costs
• Harder to test and debug
• Greater governance burden
• Risk of hallucination, tool misuse, or wrong sequencing

NIST’s generative AI guidance is useful here. It emphasizes governance, measurement, and control mechanisms for risks such as inaccuracy, privacy issues, security gaps, and misuse.

Human-in-the-loop limitations

• Can slow throughput
• Adds staffing cost
• Creates review queues
• May become a hidden bottleneck if overused

The answer is not “remove humans.” The answer is “place humans where they matter most.”

Practical recommendation for leaders

Use this rule set:

• Choose chatbots when the interaction mostly ends with an answer.
• Choose automation when the path is fixed and repetitive.
• Choose agentic AI when the work is variable, context-heavy, and action-driven.
• Choose human-in-the-loop where risk, regulation, or ambiguity makes oversight non-negotiable.

And most importantly: do not think in categories alone. Think in workflow layers.

The strongest enterprise design is usually a stack:

• conversational front end,
• automation for the predictable path,
• agentic orchestration for the complex path,
• human approval for critical decisions.

Conclusion

This decision is less complicated than the market makes it sound. Chatbots are designed for interaction, AI Voice Agents enable natural conversational engagement across voice channels, automation handles deterministic execution, and agentic AI manages dynamic, goal-driven work. Human-in-the-loop systems remain essential for maintaining quality, judgment, compliance, and trust.

If you match the right technology to the right workflow, you gain efficiency, scalability, and measurable business value without unnecessary complexity. If you mismatch them, you risk costly implementations and disappointing outcomes. Organizations that combine automation, agentic intelligence, and AI Voice Agents strategically are better positioned to deliver superior customer experiences and operational performance. At Enova, we help businesses evaluate where each approach creates the greatest impact, ensuring AI investments generate sustainable results rather than becoming expensive experiments.