In September 2026, Gartner predicted that 70% of enterprises will abandon agentic AI built by embedded vendor engineering teams by 2028, "trapped by soaring costs and unable to evolve it on their own."
‍

Gartner separately predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or weak risk controls.


Few of these projects fail on launch day. Most launch well and then slowly stop being useful. The challenge is clear. It’s the handover. It’s the ongoing maintenance and upkeep. 


When a vendor builds an AI agent and leaves, the business inherits a solution it might lack the expertise, resources or processes to maintain. There is no shortage of vendors and consultants offering to build AI agents, or guides explaining how to build them. But getting an agent into production is only part of the work.


For an AI deployment to deliver lasting value, someone needs to be accountable for its ongoing performance. That means monitoring results, updating behavioural logic, testing changes, managing risks and continually improving the experience.


Gartner also points to the potential for traditional services and partner models to deliver more cost-effective, predictable results once the scope of a project is clearly defined.


This is where the delivery model matters. A managed AI service extends beyond implementation, with a partner responsible for helping the solution perform, adapt and improve over time. The AI agents that deliver lasting value have a plan for what happens after launch, and a partner accountable for helping them get there.
‍

Why handed-over AI sales agents tend to stall

An AI agent that talks to customers is never finished. Prices change, policies change and products are discontinued. Customers start asking questions nobody anticipated. The AI models underneath improve every few months, and an agent built on last year's technology starts to look slow and limited next to what customers use elsewhere.
‍

When a build team hands an agent over and moves on, all of that lands on someone inside the business. Usually that person already has a full-time job. The knowledge falls out of date, and accuracy slips. Eventually someone asks why the business is paying for it.
‍

MIT's 2025 study of enterprise AI deployments, The GenAI Divide, points in the same direction as Gartner. Organisations that bought AI from specialist partners succeeded roughly twice the rate of those that built internally. 

‍

The cost of building in-house

It’s common to think that building an agent internally means paying for it once and owning it.


In reality, ownership comes with ongoing work. The team needs to debug and retune the agent as customer behaviour and volumes change, rewrite prompts as models evolve, monitor the platform, configure alerts and plan upgrades as the underlying technology changes.

The initial build is only the beginning. Someone still has to be responsible to keep it working.


Another cost that's often overlooked in AI deployments is the cost of maintaining quality. An agent that works well on launch day drifts if nobody is watching it. Answers go stale as the underlying data changes. Edge cases surface that nobody planned for. Model updates shift behaviour in ways that only show up in real conversations.


That last point is well documented. Researchers from Stanford and UC Berkeley compared two versions of GPT-4 released three months apart. They found its behaviour on the same tasks changed substantially, including how well it followed instructions. Their conclusion: anyone relying on these models needs to monitor them continuously.


Holding quality at the level customers expect is continuous work. For an in-house team, it competes with everything else on the roadmap. Every hour your team spends maintaining an AI platform is an hour not spent on the work only your organisation can do.

A managed service takes the build, the operation and the ongoing improvement off your internal roadmap. It also makes the cost predictable, because the monitoring, retuning and model upgrades sit with the partner rather than your team.
‍

What it looks like when you Partner 

Australian Red Cross Lifeblood started working with Clevertar in 2021 on a single job. Jamie, its AI agent, answered donor FAQs using what was then standard intent-matching technology.


Jamie has changed a lot in the five years since. In 2022 it got better at understanding natural language. In 2023 it was rebuilt on a large language model. In 2025 it moved into Lifeblood's donor portal and handled the donor eligibility quiz, guiding people through it as a conversation. Jamie explains why each question is asked and even handles sensitive topics. 

The results show what that continuity buys:

• Quiz completion reached 85% in Jamie.
  Between April 2025 and April 2026, 6,781 people started the quiz through Jamie and 5,442 finished it.
  ‍
• In December 2025, a national call for donors drove demand to fifty times normal levels.
  The website struggled and phone queues grew. Jamie handled about 15,000 interactions in the following days, covering eligibility screening and donor centre information. If even a third of those chats had become phone calls, the contact centre would have struggled.
  ‍
• Lifeblood's own contact centre team started using Jamie as an internal knowledge base, because it was the fastest way to find a reliable answer.
  
  

None of that came from a single build. It came from a partnership with clear outcomes that were well defined from the start and revisited as things changed.
‍

When both sides agree on what success looks like, whether that’s completion rates, deflection, customer satisfaction or cost to serve, every decision after launch has a reference point. A new model is adopted because it moves those numbers, not because it’s new. A knowledge update gets priority because the data shows where customers are getting stuck. A new use case is added only once the current one is working.


Each improvement should build on the last. The agent gets more accurate, more useful and more trusted over time, by customers and by the staff working alongside it. That’s what turns a deployment into a lasting part of how a business engages its customers.
‍

What a managed AI service actually covers

At Clevertar, we take ownership of defined parts of a client's customer engagement. We build the agent specific to your needs, run it and keep improving it, so you don't have to become an AI platform operator. In practice, that means: 

Before launch

• Scope agreed upfront.
  What the first phase includes is written down before the build starts, along with the measures it will be judged on.
  ‍
• An agent built on your content.
  It’s trained on your products, policies and FAQs, speaks in your brand voice, and has guardrails against off-brand or unsafe answers. The knowledge base belongs to you.
  ‍
• Tested before it goes live.
  Your team tests a prototype against an agreed set of real customer questions, and we refine it until sign off.
  ‍

After launch

• The platform is our job.
  Hosting, uptime monitoring and AI model costs are covered, and the agent’s knowledge refreshes from your content when needed.
  ‍
• A named contact and monthly reporting.
  One person at Clevertar knows your account, reports on performance each month and makes content updates as your products and policies change.
  ‍
• Ongoing tuning.
  We adjust how the agent handles conversations as real usage shows where customers get stuck. As the underlying AI improves, we move the agent forward where it lifts results.

‍

AI sales and support agents that deliver lasting value have a plan for after launch, clearly defined outcomes, and a partner accountable for reaching them.
‍

Before you sign off on an AI agent, ask who will be responsible for it in month six, not just in week one.
‍

If you’d rather have a partner who stays accountable for the results, talk to us about a proof of value. You’ll see the agent working with your own customers before you commit to anything longer. Either way, make sure someone owns the answer to that question before launch day.




Sources

• Gartner, "Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028" (20260929): gartner.com
• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (20250625): gartner.com
• MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Fortune: fortune.com
• Clevertar, "How Lifeblood built an AI that donors actually trust": clevertar.com
• Chen, Zaharia and Zou, “How Is ChatGPT’s Behavior Changing over Time?”, Harvard Data Science Review (2024): arxiv.org

‍