Gartner has made a blunt prediction. By 2028, it says, 70 per cent of enterprises will abandon agentic AI built by vendor forward-deployed engineering, the model where a vendor embeds its own engineers in your organisation to build the system. The reasons Gartner gives are rising costs and organisations being unable to evolve the systems on their own.
If you are planning an agentic AI implementation, that line is worth reading twice. The risk is not that the technology fails. It is that you end up unable to run what you paid for.
What Gartner actually said
The prediction comes from a Gartner press release dated 29 September, with analyst Mukul Saha quoted. Gartner also expects that through 2028 fewer than 20 per cent of these engagements will turn recurring customer needs into capabilities in the vendor's core product. It calls the pattern a growing risk of "FDE washing".
Saha's advice includes naming an executive sponsor accountable for business outcomes, not simply the implementation budget. He also observed that some providers charge premium fees without the delivery depth, programme management or change management maturity to justify them. The three phase recommendation is simple: set contract terms before signing, embed vendor engineers with your own people during delivery, and follow a pre-agreed exit plan so the organisation can run the solution independently.
A note of caution. This is a prediction, not measured data, and Gartner's press release includes no survey results. Predictions of this kind can be directionally useful while being wrong on the number. I would treat the 70 per cent as a prompt for questions, not a statistic to quote.
Why this matters now
Vendors are moving quickly into this space. Cognizant launched Activate on 5 October, a unit aimed at companies with $1 billion to $5 billion in annual revenue, offering pre-configured solution kits and a single accountable engagement lead. That is a reasonable offer, and may suit some buyers well. It also shows how crowded the market for implementation partners is becoming.
What this suggests is a buying decision that is easy to get wrong. Delivery speed is visible in a proposal. Your ability to run the system after the vendor leaves is not.
What agentic AI implementation should look like on day one
Most of the exit plan is decided at the start, not the end.
Insist that your own staff are embedded in the build, with named counterparts for each vendor engineer. Require documentation, prompts, evaluation sets and integration code to be delivered into repositories you control, as the work proceeds rather than in a bundle at the end. Agree who owns intellectual property in anything built on your data. And write the exit criteria into the contract: what you must be able to do independently, how it will be tested and when.
In practice, this tends to be overlooked because it feels pessimistic at the start of a hopeful project. It is the same instinct as asking for a handover plan when you hire a contractor.
A composite scenario
A mid-market insurer engages a vendor to build an agent for claims triage. The vendor's engineers deliver on time and the results are good. A year later the insurer wants to change the escalation rules and add a new product line. Nobody internal understands how the agent's logic is structured, the evaluation tests sit in the vendor's environment and a change request is quoted at a premium. The agent works. The organisation cannot move it. This is the pattern Gartner is describing, and it is avoidable.
Where Australian organisations should look
For AI implementation services in Australia, the same logic applies whichever partner you choose, including us. Ask any partner what you will be able to do alone at the end, and ask them to put it in writing. A partner who is comfortable with that question is usually a good sign.
What This Means for Your Organisation
What we see across organisations is that implementation partners are judged on delivery and then, quietly, on dependency. The organisations that avoid lock-in treat capability transfer as a deliverable with its own acceptance tests. That means someone internal who can explain how the agent works, change it and test it without calling the vendor. If you cannot name that person at the start, you are buying a service rather than building a capability, which may be fine if you choose it deliberately.
Key Takeaways
- Gartner predicts 70 per cent of enterprises will abandon agentic AI built by vendor forward-deployed engineering by 2028, but this is a prediction, not measured data.
- The core risk is dependency: costs rise and the organisation cannot evolve the system on its own.
- Build capability transfer into the contract from day one, with embedded staff, code and evaluation sets in your repositories and clear exit criteria.
- Ask every implementation partner what you will be able to do independently at the end, and get the answer in writing.
How Trusenta Can Help
- Custom AI Development builds agentic systems with your team alongside, so the knowledge and code stay with you.
- AI Integration Services connects agents to your existing systems in ways your own engineers can maintain.
- Fractional CIO gives you independent senior oversight of vendor contracts and exit terms.
Conclusion
A good implementation leaves you more capable than it found you. Judge any agentic AI plan by whether you could carry on without the people who built it, and write that test into the agreement before the first line of code is written.
