Everyone has an AI plan until reality hits

AUTHORS
PUBLISHED
Nov 18, 2026
CATEGORY
DURATION
8 MIN
  • AI/ML
  • Durable Execution
  • Retries

Revity This is a guest post from Evgueni Iakhontov, Co-founder & Partner at Revity.

Mike Tyson once said, “Everyone has a plan until they get punched in the mouth.” He was talking about boxing, but anyone who has tried to automate a business process will appreciate the sentiment. The bruises simply appear on budgets, delivery schedules, and occasionally the reputations of those who approved them.

There is something satisfying about watching an AI agent perform its first successful task. It receives a request, makes a decision, calls several systems, and returns the correct result. The demo is flawless, the executives are impressed, and someone inevitably asks whether it can be deployed across the business by Friday, preferably before Monday’s board meeting.

Nobody asks what happens when things go wrong. Not yet, anyway.

A perfectly reasonable plan#

Consider a retailer looking to automate complaint handling. A customer reports a faulty product, an AI agent retrieves their purchase history, checks the warranty, and determines whether a refund is appropriate. Larger refunds require manager approval. Once approved, the agent processes the payment, updates the CRM, and notifies the customer.

The team builds a prototype, and everything works. The agent handles complaints in minutes rather than hours, at a fraction of the cost of having someone do it manually.

The business case practically writes itself. Someone estimates a 40% productivity improvement. Another suggests 60%. By the time the figures reach the executive committee, the savings have acquired a confidence that would make even the AI model envious.

Yet the plan rests on a fragile assumption: that customers, employees, and computer systems will behave as expected. Unfortunately, they rarely do.

The trouble with the real world#

The first complication is usually human.

A customer forgets their order number and takes three days to respond when asked for it. Another complaint needs approval from a manager who has just left for two weeks in Bali. Their out-of-office message is cheerful but offers little guidance on who might approve a $500 refund.

Even when the manager is at work, the request may disappear into an inbox containing 147 unread emails and a meeting invitation titled “Quick catch-up,” which everyone knows will take at least an hour.

People get busy, go on leave, or simply forget. A process that depends on someone responding needs to account for what happens when they do not. The alternative is to hope for the best, which remains a surprisingly popular approach to workflow design.

Then there are the technical complications.

Suppose the agent sends a refund request to the payment system, but the request times out. Did the refund go through? Retrying seems sensible, unless the first attempt succeeded and the customer is about to receive their money twice. The customer will probably be delighted. Finance, somewhat less so.

Perhaps the refund succeeds but the CRM update fails. The payment system now shows a completed refund while the CRM insists the complaint is unresolved. The systems disagree about what happened. Anyone who has worked in a large organisation will recognise this as a fairly normal Tuesday.

AI introduces another complication. An agent might misunderstand the complaint, apply the wrong policy, or confidently recommend a refund that should have been declined. Confidence, unfortunately, is not always a reliable indicator of competence. This is true of AI models and occasionally the rest of us.

You need to decide which actions AI can take independently, which require validation and when a human should intervene.

This is the kind of challenge we help businesses solve at Revity. We work with organisations to identify opportunities for AI automation, redesign business processes where necessary, and build AI agents that integrate with existing systems. Much of the work is not actually about the AI model itself, but about making sure the entire process works reliably.

The cost nobody mentioned in the demo#

None of these problems are new. We have dealt with retries, failed transactions, and human approvals for decades. AI simply makes it practical to automate more complex processes, bringing those familiar engineering challenges along for the ride.

Every process needs to remember what has happened and what remains to be done. Transient failures need retries. Human approvals require reminders and escalation rules. Systems need to recover from interruptions without losing progress or repeating transactions.

Someone must build, test, and maintain all of this. None of it tends to feature in the original demo, where the agent completes its task in 30 seconds and everyone goes home happy.

This is why discussions about AI automation costs can be misleading. We spend considerable time comparing model prices and token consumption, yet those may represent only a small part of the total cost.

For example, an agent might consume $0.10 worth of AI tokens handling a complaint, but that tells us little about the cost of processing it reliably.

The original ROI spreadsheet may need another tab or two. Possibly a new workbook.

The missing layer: Durable Execution#

Fortunately, we do not need to build all this infrastructure from scratch. Temporal provides Durable Execution for processes that span multiple systems, wait for human input, and need to recover from failures. It records a Workflow’s progress in an Event History so the process can resume from its recorded state after an interruption.

In our refund example, a Workflow coordinates the process, while Activities handle individual operations such as issuing a refund or updating the CRM. If the payment system is unavailable, Temporal can retry the Activity according to its Retry Policy. If manager approval is required, the Workflow can wait for days with its state persisted. If a Worker running the application code crashes, another Worker can resume execution.

Retries still need to be safe. An Activity may run more than once if its result is not recorded, even when the payment has already gone through. The payment integration therefore needs idempotency: retrying the same refund request must not issue another refund. The payment system must recognise duplicate requests, typically through an idempotency key.

Temporal saves engineering teams from building and maintaining much of the infrastructure needed to track progress, schedule retries, and resume execution after interruptions.

Of course, Temporal cannot persuade the manager in Bali to check their emails or decide whether an AI recommendation makes sense. Those responsibilities still sit with the people designing the process. But it does allow engineers to spend less time worrying about what happens when things fail and more time building the actual business process.

Does it make financial sense?#

Imagine a business automating 10 processes involving multiple systems and occasional human intervention. Without a workflow platform, engineers must build mechanisms for retries, state management and recovery. They might create reusable components, but before long they are maintaining something resembling their own workflow platform.

Congratulations. The business is now developing workflow infrastructure, although that was never the plan. Someone will eventually give it a name and request a budget for version 2.

Temporal offers an alternative, allowing engineers to concentrate on business processes rather than rebuilding common infrastructure.

There are costs, naturally. Temporal Cloud has usage charges, while self-hosting requires infrastructure and operational support. Developers also need to learn the platform. For a short-lived automation involving one or two API calls and basic recovery requirements, it may not be worth the additional complexity.

But for processes involving multiple systems, human approvals and transactions that must recover from failure, the economics can look quite different.

This leads to a metric I believe deserves more attention: the cost per successfully completed business process.

That includes AI usage, infrastructure, retries, failed transactions, human intervention, and ongoing support. It is this total cost per completed process that should be compared with the existing manual process.

If Temporal helps reduce that cost, it improves the potential return on AI automation and may make previously uneconomical processes worth automating.

Beyond the happy path#

Getting an AI agent to complete a task once in a sandbox is very different from handling thousands of customer requests across fragmented enterprise tools.

You cannot anticipate every exception. What you can do is design systems that tolerate failure, prevent duplicate transactions, and apply clear rules for escalating to a human.

At Revity, we approach AI automation as an engineering and business problem. Building an agent that can perform a task once is relatively straightforward. Making sure it can handle real customers, unexpected failures, and all the complications of running a business is where much of the work lies.

The objective, after all, is not to produce an impressive demo. It is to automate a real business process reliably, at a cost that makes commercial sense.

Mike Tyson would probably approve.

Temporal Cloud

Ready to see for yourself?

Sign up for Temporal Cloud today and get $150 in free credits.

Build invincible applications

It sounds like magic, we promise it's not.