Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOmar Baruzzo’s argument is that an agent project usually runs into trouble when it stops reading business data and starts changing it. A read-only demo can look finished. Once the agent has to post to an ERP or another system of record, the hard problems are integration, data integrity, permissions, retries and reconciliation. Those are systems-engineering problems, and a better model does not solve them.
This is one practitioner’s analysis, not a measured ranking of why projects fail. No source we reviewed gives a representative failure-cause rate. Gartner and NIST guidance does support the broader point that the surrounding controls matter as much as the model.
As an Amazon Associate I earn from qualifying purchases.
Why read-only demos mislead
Baruzzo’s central observation is that reading is forgiving and writing is not. An agent that summarizes invoices, answers questions about stock or drafts a recommendation can be wrong without lasting damage, because a person sees the output before anything changes. The moment the agent writes, its mistakes become records in a system that other processes, reports and people rely on.
That is why a project can look successful in a demo and then stall. The model’s behavior may be fine. The missing pieces are the ones a demo never exercises.
#1 Best Overall
What goes wrong at the write boundary
These examples come from Baruzzo’s essay. They are his operational observations, and the governance and monitoring sources below do not independently verify each ERP-specific case.
Data integrity and server-side validation
If an agent’s output goes straight into a business system, the system itself has to decide whether the write is acceptable. Validation belongs on the server side, where the data rules live, not in the agent’s prompt. A prompt asks the model to behave. Validation enforces the rule whatever the model does.
Irreversible postings
Some writes cannot be quietly undone. A posted accounting entry, for example, may need a formal reversal rather than a deletion. If the agent can trigger such an action, “we’ll fix it afterwards” is not a plan. Decide up front which actions are reversible and which need a gate in front of them.
Queued batch writes
When writes are queued and applied in batches, an error may not show up when the agent acts. It surfaces later, when the batch runs, and the failure is separated from its cause. That makes it harder to tell which agent action produced which result.
Retries against non-idempotent endpoints
Baruzzo specifically calls out non-idempotent writes as a duplicate-order risk. If a call times out and the agent retries, an endpoint that does not recognize repeats can create the order twice. Ordinary retry logic, which works for reads, becomes dangerous for writes unless each request carries a key that lets the system detect duplicates.
Scoped identities
The agent should act under an identity with only the permissions its task needs. A broad service account turns any error, or any manipulated instruction, into broad access. A narrow identity limits the damage and makes the audit trail readable.
Imperfect test environments
Test systems often differ from production in data volume, master data quality, configuration and integrations. An agent that passes in a clean sandbox can meet unfamiliar cases in production. Check the test environment against real production data conditions before trusting results from it.
Recommended Free Tools
Reconciliation
Changes made by an agent should be traceable and comparable against what was intended. Reconciliation is how a team finds out that what the agent did and what the business expected have drifted apart.
Match controls to autonomy and consequence
Gartner’s guidance, published May 26, 2026, divides agents into four autonomy levels and says controls should scale with autonomy and access scope. It is a useful way to compare projects, because “agent” covers very different risk profiles.
| Autonomy level (Gartner) | What the agent does | Questions to settle |
|---|---|---|
| Observe | Reads and reports | What data can it see, and is that access scoped? |
| Advise | Recommends; people execute | How are recommendations checked, and are they recorded? |
| Act with approval | Writes only after explicit sign-off | Is the approval meaningful and auditable? |
| Act autonomously | Writes within guardrails | Who monitors it, who can stop it, and how is it rolled back? |
Gartner’s Shiva Varma, Senior Director Analyst, put the failure in these terms: “Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure.” The practical reading is to choose a control set for each workflow, not apply one uniform policy to every agent.
Human approval is not a complete control
Putting a person in front of every write feels safe, but Gartner says approvals need meaningful workflows and audit trails, and warns about approval fatigue. A reviewer who clicks through hundreds of requests is not a control. The approval screen should show what will change, in terms the approver can judge.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
For autonomous operation, Gartner calls for continuous monitoring, enforced guardrails, rollback mechanisms, circuit breakers and clear ownership. A circuit breaker is the answer to the question of what stops the agent when something looks wrong, and a named owner is the answer to who decides.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Evaluation and monitoring continue after launch
NIST describes post-deployment monitoring as a way to check behavior in real-world conditions and track unexpected outputs and consequences. Its published report also says methods and best practices remain nascent and scattered, so teams should expect to build some of this themselves rather than adopt a settled standard.
One NIST effort still in development is an evaluation-probe approach. It checks factual grounding against a human-curated corpus and accumulates results in a machine-readable audit trail. The useful idea is that evaluation can leave evidence behind, which fits the reconciliation point above.
NIST’s voluntary AI Risk Management Framework covers risk across design, development, use and evaluation, with associated guidance on monitoring, incident response, recovery and change management. It is a framework teams can draw on. It is not a mandatory certification and does not guarantee trustworthiness. OWASP’s GenAI Security Project also covers agentic AI systems and is a reasonable starting point for security research, though it does not validate any particular enterprise architecture.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesWhat the numbers do and do not say
Gartner predicts that 40% of enterprises will demote or decommission autonomous AI agents by 2027 because of governance gaps found only after production incidents. That is a forecast about enterprise governance outcomes. It is not a statistic that 40% of agent projects fail, and it does not say what share of stalls come from integration. Treat claims about how often the model is or is not the cause as unsupported unless they cite a representative study.
Quick Recap
A production-readiness check before an agent gets write access
- Classify the workflow: observe, advise, act with approval or act autonomously.
- Scope the identity: give the agent only the permissions the task requires.
- Validate on the server: enforce data rules in the target system, not only in prompts.
- Make writes safe to repeat: use idempotency keys or equivalent duplicate detection before enabling retries.
- List irreversible actions: put approval or a hard block in front of anything that cannot be cleanly undone.
- Test against production reality: compare test data and configuration with production.
- Design the approval screen: show the actual change, and watch for reviewer fatigue.
- Plan reconciliation: record what the agent did and compare it with what was intended.
- Name an owner and a stop mechanism: include a circuit breaker and a tested rollback.
- Monitor in production: track unexpected outputs and consequences after launch, not only before.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
