Why enterprise AI agents fail and what failed runs cost
Learn why enterprise AI agents fail, how to diagnose the cause, and how to calculate cost per completed workflow including retries and human intervention.
Blog
Ideas, field notes, and product updates from Monarch.
Featured · Benchmarks
Monarch completes nearly 70% more workflows than Codex
Benchmark results: Monarch completed 129 of 150 workflows to Codex’s 76 using the same model and reasoning level. See completion and model costs.
Engineering · Governance
Secure your agents: but how?
An agent security reference architecture: Sandboxed processes, LLM gateways, observability, forward proxies, and content inspection.
Labs · Research
Self-Evolving Harnesses #1: How far can your setup climb?
Get better results from the same AI model by improving how it’s used.
Latest articles
No articles match these filters. Clear a filter to see more.
Guides
AI agent integration platforms: How to choose
Compare Workato, Monarch, Composio, and Merge Agent Handler, then separate integration needs from the choice to build with Microsoft, Google, OpenAI, or Claude.
How AI agents work with legacy systems without a suitable public API
See how AI agents can use authorised private web interfaces and UI paths in legacy systems, with an application map, approval steps, and verified results.
Why AI agents need an application map
An application map connects records, actions, permissions, and validation checks so AI agents can execute workflows with explicit business rules.
Secure your agents: but how?
An agent security reference architecture: Sandboxed processes, LLM gateways, observability, forward proxies, and content inspection.
Monarch completes nearly 70% more workflows than Codex
Benchmark results: Monarch completed 129 of 150 workflows to Codex’s 76 using the same model and reasoning level. See completion and model costs.
Self-Evolving Harnesses #1: How far can your setup climb?
Well then, how much better can an AI agent get without a better model nowadays?
So, you need a mainframe development environment
How we built a mainframe playground and vibecoded in COBOL.
We're thinking about AI safety all wrong
Practical safeguards for AI workflows: Permissions, protected credentials, deterministic actions, monitoring, and testing before production.
Mainframes aren’t dinosaurs
Why mainframes still power demanding business workloads, and how AI agents can work with their integrations and terminal applications.
Your next software builder might work in finance
What OpenAI’s non-engineering teams suggest about enterprise work over the next three years.
Who owns an AI workflow after the pilot?
An AI workflow needs someone accountable for its results after launch, with the time, evidence, and authority to keep improving it.
Monarch Sydney panel with BCG X and Airtree: When AI moves faster than the enterprise
Sam Senior, Michael Schniering of BCG X, and Claire Smith on what separates the companies getting returns from AI. Full video and transcript.
Your business can do more than your AI connectors allow
A connector logo does not mean your agents can finish the work. Expect deeper application discovery, reusable execution, and people leading the business change.
Monarch vs Microsoft Copilot, Copilot Studio, Foundry, and Fabric
Microsoft's AI products serve different roles. See how Monarch discovers application operations and runs cross-system agentic workflows alongside your existing stack.
Your AI strategy won’t work if everyone’s job stays the same
AI adoption requires changes to people’s workloads, incentives, and responsibilities. Give them time to improve the work and authority to lead the agents doing it.
AI agents for insurance claims: Build the evidence before the settlement
An insurer's AI workflow should connect the policy, repair estimate, and claim history, then help the team complete the next action with the customer properly informed.
AI agents for financial hardship: Stop making customers repeat their story
A useful banking agent connects the customer's request, account history, and existing support arrangement, then helps the team carry its decision into the systems that administer the loan.
AI agents for airline disruption: Close the gap between rebooking and a place to stay
Rebooking leaves work across hotel suppliers, ground transport, and case systems. Start with one overnight assistance workflow and verify every arrangement.
Permission-aware enterprise agents: Whose authority applies?
Follow an agent's authority from the requesting user through connected accounts, business approval, execution, and removal of the access it creates.
AI agents in a customised ERP: Get the change request right
Follow a supplier change request through custom fields, related records, and approval rules, then verify exactly what the ERP has saved and submitted.
Computer use vs Monarch: Enterprise agent workflows
Compare how agents reach application actions, what gets reused and repaired, and how each approach proves the work actually happened.
Monarch vs CData Connect AI for enterprise agents
Compare supported queries and writes, curated tools, and application discovery through a workflow that takes a held order to an approved release.
Connecting AI agents to internal apps without an MCP server
Build one useful business operation, establish its permissions and completion checks, then choose how your agent will call it.
Monarch vs Arcade for enterprise AI agents
Compare authorised tool execution, custom application work, and deployment choices, then test the complete workflow your team needs to run.
Monarch vs Composio for enterprise AI agents
Compare application access, authentication, and workflow ownership, including the work needed when a required action sits outside a ready-made tool.
AI agents for Australian energy retailers: Start with billing investigations
A useful energy workflow connects the complaint, account, metering evidence, and reviewer. Prove that investigation before expanding into financial actions.
Monarch vs UiPath for enterprise AI agents
Compare how Monarch and UiPath connect existing agents to enterprise systems, reuse automations, and keep control of work across applications.
AI won't transform a company you refuse to rebuild
Most enterprises use AI to run a slightly faster version of the company they already have. The bigger move is to rebuild how it works while it keeps running at full speed.
AI agents for telecommunications: Recover orders that fall between CRM and provisioning
An accepted broadband order can still leave a customer without service. Recovering it means following the work through provisioning, inventory, and billing, then proving the records agree.
AI agents for employee offboarding: Close the gaps across HR and IT
Carry an approved departure through access removal, work handover, and records preservation, with evidence of what each system actually completed.
The map your AI needs is already inside your systems
An enterprise AI ontology is the map that lets agents operate your business. Most of it already lives in your software. Discover it, don't rebuild it from memory.
AI agents for supplier delays: Give planners a delivery date they can use
Follow a late purchase-order line from a supplier's revised promise through split deliveries, production decisions, and actual receipt.
AI agents for accounts payable: Resolve the invoice exception before the payment run
Follow an invoice discrepancy through purchasing, receipts, supplier correspondence, and AP, then verify that the approved correction reached the records that control payment.
MCP is the wrong default for enterprise AI
MCP makes connecting agents to your systems easy — not dependable. Why it is the wrong default for enterprise workflows, and what to run instead.
Monarch vs Google Gemini: What your Google stack covers
Google already supplies models, workplace AI, connectors, and agent infrastructure. Compare the application access your workflow still needs.
How to evaluate an enterprise AI pilot: A scorecard for completed work
Define which cases enter the pilot, what counts as a useful result, and how you will compare it with today's operation before the first live run.
Monarch vs Workato for enterprise AI agent integration
Compare how Workato and Monarch make enterprise application operations available to AI agents, and test the maintenance work behind a complete workflow.
Same model. 23% more workflows completed.
We ran 591 AutomationBench workflows with and without Monarch. Opus 5's completion rate rose 23%; GPT-5.6 Sol and Kimi K3 improved too.
Build with Claude or OpenAI directly, or add Monarch?
Compare the same frontier models with existing or custom tools and with Monarch, including the application work your team must build and maintain.
Workato vs MuleSoft vs Automation Anywhere vs Monarch
Compare Workato, MuleSoft, Automation Anywhere, and Monarch on workflow completion, agent capabilities, total cost, and Australian deployment requirements.
Why enterprise AI agents fail and what failed runs cost
Learn why enterprise AI agents fail, how to diagnose the cause, and how to calculate cost per completed workflow including retries and human intervention.
AI agent integration platforms: How to choose
Compare Workato, Monarch, Composio, and Merge Agent Handler, then separate integration needs from the choice to build with Microsoft, Google, OpenAI, or Claude.
How AI agents work with legacy systems without a suitable public API
See how AI agents can use authorised private web interfaces and UI paths in legacy systems, with an application map, approval steps, and verified results.
Why AI agents need an application map
An application map connects records, actions, permissions, and validation checks so AI agents can execute workflows with explicit business rules.
Browse all 41 articles and reports
No articles match your search. Try a broader term or clear the search.
AI strategy
- AI won't transform a company you refuse to rebuild
- How to evaluate an enterprise AI pilot: A scorecard for completed work
- MCP is the wrong default for enterprise AI
- Monarch Sydney panel with BCG X and Airtree: When AI moves faster than the enterprise
- The map your AI needs is already inside your systems
- Who owns an AI workflow after the pilot?
- Your AI strategy won’t work if everyone’s job stays the same
- Your next software builder might work in finance
Agent reliability
Application access
- AI agent integration platforms: How to choose
- Build with Claude or OpenAI directly, or add Monarch?
- Computer use vs Monarch: Enterprise agent workflows
- Connecting AI agents to internal apps without an MCP server
- How AI agents work with legacy systems without a suitable public API
- Mainframes aren’t dinosaurs
- Monarch vs Arcade for enterprise AI agents
- Monarch vs CData Connect AI for enterprise agents
- Monarch vs Composio for enterprise AI agents
- Monarch vs Google Gemini: What your Google stack covers
- Monarch vs Microsoft Copilot, Copilot Studio, Foundry, and Fabric
- Monarch vs Workato for enterprise AI agent integration
- So, you need a mainframe development environment
- Why AI agents need an application map
- Your business can do more than your AI connectors allow
Governance
Reliability
Transformation
- AI agents for accounts payable: Resolve the invoice exception before the payment run
- AI agents for airline disruption: Close the gap between rebooking and a place to stay
- AI agents for Australian energy retailers: Start with billing investigations
- AI agents for employee offboarding: Close the gaps across HR and IT
- AI agents for financial hardship: Stop making customers repeat their story
- AI agents for insurance claims: Build the evidence before the settlement
- AI agents for supplier delays: Give planners a delivery date they can use
- AI agents for telecommunications: Recover orders that fall between CRM and provisioning
- AI agents in a customised ERP: Get the change request right
- Monarch vs UiPath for enterprise AI agents
- Workato vs MuleSoft vs Automation Anywhere vs Monarch