Why That Warning Is a Thing to Worry About
The demo went well. The agent located the data, finished the assignment, and made everyone’s afternoon a bit brighter.
Someone then enquired, “What happens if the supplier changes its price, the approval is missing and the finance system is down?”
That’s when the actual talk starts.
With companies thinking about their next AI bets in September, it’s worth taking a closer look at a pre-existing Gartner prediction that more than 40% of agentic AI projects will be abandoned by the end of 2027. Although not a measurable failure rate of AI in general, this is a projected consequence of escalating expenses, questionable economic value or insufficient risk controls, in June 2025.
At Voyantt, we hear a practical issue underlying that warning: What has to be in place for an AI agent to be a reliable part of the working day?
What Makes an AI Agent Different from Conversational Chatbots
A simple conversational chatbot may answer a query, or write a response. An AI agent may also loop through a series of tasks using associated tools and systems. OpenAI describes AI agents as autonomous systems designed to execute complex tasks for a user while operating within specific safety boundaries. Rather than just answering questions or generating text, these agents can use tools, and draw conclusions to achieve a set goal.
For instance, while a chatbot may generate a sales follow-up, an agent may review the customer’s history, identify an overdue follow-up, write the email, have it approved, send it, and log the contact.
This will reduce too many mouse clicks to just a few. This also implies that the organization must determine what the agent can access, what actions it can do and when it has to ask a human.
Some AI Projects Can Get Scrapped. Here’s Why.
Gartner’s concerns lead to three typical business questions: Can we afford to keep doing this? Is it doing enough good to warrant the price? Have we capped what it can do, and is that reasonable?
Cancellation isn’t always a sign that the technology failed. A functioning project may duplicate another endeavour, lose financing, or address an issue that has fallen down the priority list. These are instances of business reasons for cancellation, not further discoveries in the Gartner prediction.
So, what should a team consider before launching a larger project?
Start With a Genuine Problem: Something Worth Solving
“We need an AI agent” is still a vague sentence; it triggers a lot of questions.
“Our purchasing team spends hours checking routine orders across several systems” provides everyone with something valuable to work with.
Before you decide on a platform, learn who performs the job, where it gets stuck, who pays for errors, and what choices demand judgement. Then determine whether an agent is a good match. If the work follows predictable guidelines then good old fashioned automation can probably handle the job just well.
When a challenge is obvious , it is easy to quantify success and difficult to confuse a flashy demo for a good investment .
Clean up Your Workflow to Ensure Greater Productivity
If two departments have different approval rules an agent will not resolve this by processing requests faster. Same for duplicate supplier records and missing budget info.
First, map the process. Choose the system with the authoritative data, cut out the unnecessary steps, and find the person in charge of exceptions.
It’s the digital version of clearing your kitchen counter before you cook. A little preparation goes a long way.
Plan for the Days When Things Go Sideways (Hey, It Rhymes!)
Live work involves incomplete requests, unavailable systems, and stuff nobody thought to put in the demo.
The agent needs a firm answer. Should it retry, request missing information, wait or hand the task over to a human?
What happens if a connection drops immediately after a purchase order is placed? The workflow should verify that the first attempt was successful before trying it again. Otherwise, one short outage could become two orders of the same supplies.
These are the details that make the difference between employees being able to trust, or not, the system on a normal, slightly chaotic, Tuesday.
An AI Agent Should Enjoy Indpendence, But Within Human Control
In May 2026, Gartner predicted a different (kind of scary) future: By 2027, 40% of enterprises would need to “decommission” or “demote” existing autonomous AI agents because live incidents would reveal gaps in governance.
Governance is the set of rules, responsibilities and checks that govern what an agent can do. Gartner’s May 26, 2026 announcement “demote” means to decrease the independence of an agent. An agent who issued refunds on his own might instead make recommendations to a manager. To decommission is to remove it from service.
It’s a big difference. The above forecast counts enterprises, the previous one counts projects. Neither says that 40% of all agents have failed already.
Gartner recommends matching controls to the level of autonomy and access an agent has. The oversight required to read a document is different from that required to commit company funds.
For buying, an agent could fill in an order. Purchases above a certain value must be authorized by an authorized person. There may be a second agent to help check the details but this is not necessarily an independent safeguard of approval.
Start With One Agent and Give It a Clear Job
OpenAI recommends making the most of a single agent before introducing multiple agents. Its guide also emphasizes clear instructions, well-defined tools, safeguards, and human intervention when needed.
That’s a good place to start for a pilot. If a single agent can dependably finish the process, then there should be a clear reason for any other agent to join the team.
Specialist agents may aid when a job requires varied expertise or the instructions become too complicated for one agent to continuously follow. They also offer handoffs for coordination and extra areas to examine if things go wrong.
Only testing will tell whether the additional support is working.
How Voyantt Would Approach a Purchasing Workflow
Illustrative example: this is a proposed approach, not a verified client engagement or a report of delivered results.
Imagine a business that gets email requests for purchases. Workers manually check supplier details, investigate costs, validate spending plans, and monitor approvals. “A straightforward order might take days to complete the rounds, but none of the processes are very difficult on their own.
Voyantt will start with the workers. We would examine the procedure, pinpoint the delays, and establish a small trial program for regular purchases from authorised vendors.
Connect the Right Information
The proposed workflow would connect the request inbox or form with supplier records, the finance or enterprise resource planning system, and an approval channel such as Microsoft Teams or email.
The agent would extract the request details, flag missing information, check the approved supplier, and prepare a draft purchase order. Conflicting records would go to a person for review.
Access would be limited to what the task requires. Permission to read supplier information, for example, would not grant permission to change bank details or release payments.
Put Approval Where It Belongs
During the pilot, all purchase orders would have to be approved by a person before being submitted. Purchases beyond a company specified threshold would be routed to an authorised budget holder.
The approver would view the supplier, products, total cost, budget details and exceptions. Once approved, the process may place the order and record the result. Any major deviation or change in the source or amount would need further assessment.
If it’s not approved, it’s a pending order. Silence wouldn’t equal a yes.
Decide What Success Looks Like
A pilot target, for instance, would be to minimise the average time from a full request to an accepted purchase order by 30%, while remaining within buying regulations. That would be an objective to test not a promised or attained outcome.
We would also track errors, duplicate orders, staff review time and operational expenses. If someone spends fifteen minutes tweaking the outcome then saving 10 minutes is less interesting.
Testing might include missing pricing, denied approvals, duplicate requests and system breakdowns. A identified company owner would analyse the facts to decide whether to extend the process.
Voyantt would combine process design, application integration, approval controls and assessment so the company could see the whole workflow at work.
Five Questions to Answer Before Scaling
A useful pilot should leave the team with clear answers:
- Value: What improved after accounting for operating costs and employee review time?
- Reliability: How often does the workflow finish correctly, including unusual cases?
- Authority: What can the agent access or change, and who approves consequential actions?
- Recovery: Can the team stop the workflow, investigate errors, and correct affected records?
- Ownership: Who owns the connections, reviews performance and approves changes?
Answers to the above questions help the business make decisions on if to grow, modify or stop. A pilot that reveals a poor investment early has still done useful work.
Those answers help the business decide whether to expand, adjust, or stop. A pilot that reveals a poor investment early has still done useful work.
A Practical Way Forward
Gartner also forecasts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. That remains a prediction, but it allows for wider adoption even as individual projects are canceled.
For Voyantt, the next step is practical: choose a workflow, define the boundaries, test the awkward cases, and let the results guide what comes next.
Have a process that involves too much copying, checking, or chasing? Bring it to us. We can help assess where agentic AI fits and shape a pilot with clear approval points and measurable goals.
Let’s talk about making that workflow work better. Just send us a message, and we take it on from there!
