Enterprise AI Agents With Engineers Attached: Why Self Managed Wins

On 22 July 2026, OpenAI announced Presence, a way of buying enterprise AI agents that involves no checkout page at all. Reporting by AI News describes it as a managed offering delivered through a limited availability programme, with deployments led by OpenAI's own forward deployed engineers and a short list of approved systems integrators. The company is explicit that it is not a self serve product. Each engagement begins with a single job, such as settling a billing dispute or clearing an IT service request, and the agent is granted only the knowledge and system access that one job requires.

The reasoning behind it is sound. Gartner has warned that more than 40 percent of agentic AI projects will be scrapped before the end of 2027, and it puts the blame on governance, unclear business value and thin operational discipline rather than on what the models can do. Two years of enterprise experiments have made the same point: the hard part of a production agent is integration, permissions and change management, not intelligence. Sending engineers to do that work answers the real failure mode. It also, quietly, moves the bottleneck from software to staffing.

Aphelion AI is a private enterprise AI platform built to give businesses their own AI agents inside infrastructure they own and control, with no data leaving the building and no dependence on an external provider's roadmap. We share the diagnosis of why agent projects fail. We reach a different conclusion about the cure. Aphelion is self managed by design: we do the setup, we apply the default guardrails, and your team runs everything after that, with the Aphelion AI Agent alongside them to help.

The Delivery Model Is the Product

What is striking about Presence is that the technology is not the differentiator. OpenAI now offers broadly the same agent capability through three routes, separated less by what the software can do than by who does the work. Access to the managed route depends on workflow fit, implementation readiness and, tellingly, available delivery capacity.

That last criterion is the whole story. Software scales freely. Engineers cleared into a bank's core systems do not. The forward deployed engineer model borrowed from the data analytics world involves people embedded in customer operations for months, and those economics look nothing like the economics of metered inference. When the constraint on your AI roadmap is another company's hiring pipeline, you are no longer buying a platform. You are queueing for a consultancy.

There is a second issue worth writing into any contract. When the model vendor is also the implementation partner, the lines of accountability for a policy misapplied in production need to be stated explicitly rather than assumed. A single throat to choke is comfortable right up until the moment you need to choke it.

How Aphelion Sets You Up

Self managed does not mean self assembled. Nobody should be handed a bare platform and wished luck. Our setup process front loads the expertise so your team inherits something that already works, and the sequence is deliberate:

• Scope the first jobs. We start where OpenAI starts, with specific work rather than a general assistant. A named process, a defined outcome, a measurable result.
• Deploy privately. The platform goes into infrastructure you own or exclusively control, so your prompts, documents and outputs never leave your governed environment.
• Build the initial agents. We configure them around your processes, load the knowledge they need, and connect them to the systems they must reach.
• Apply the default guardrails. Scope limits, permission boundaries, escalation thresholds and full session logging are switched on before anything speaks to a colleague or a customer.
• Hand over properly. Your team learns the interface, the prompt builder and the enrichment tooling, then takes ownership of the deployment outright.

After that final step, the relationship changes shape. We are not embedded in your operations for months, and you are not paying for us to be. You have a working system, a pattern to copy, and the ability to build the next agent yourself on a Tuesday afternoon rather than in the next available delivery slot.

Guardrails That Ship Switched On

Most of what makes a managed deployment feel safe is not exotic. It is testing before launch, boundaries that intervene when an interaction drifts outside defined limits, records that a reviewer can audit, escalation paths that give a human structured context rather than a cold transcript, and controlled rollout with the ability to roll back. These are sensible controls. They are also perfectly capable of being defaults rather than deliverables.

Aphelion applies that control set during setup so your agents start life inside sensible limits, and your team then adjusts them as confidence grows. Because the deployment is private, several of the hardest governance problems simply do not arise. There is no third party retaining your prompts, no shared platform to vet, and no external training pipeline to worry about. Compliance work becomes a review of your own logs and your own access model, which is materially cheaper and considerably faster than assessing infrastructure you are not permitted to inspect.

The same logic applies to what the agent knows. Our approach to data enrichment means the knowledge layer is yours: your policies, your documents, your records, curated by people who understand them. An agent does not become production ready by swallowing a document library, and it certainly does not become safe that way. It becomes ready when the people who own the process have shaped what it knows and what it may do with that knowledge.

"Guardrails are not a professional services deliverable. They are a product feature. If safe defaults only arrive when an engineer is in the room, the platform has not finished being built."

The Agent That Helps You Run the Agents

The obvious objection to self management is skill. Enterprise teams are busy, and prompt design, scoping and escalation policy are unfamiliar territory for most of them. This is exactly the gap the managed model is priced to fill.

Aphelion fills it differently. The Aphelion AI Agent works as an in platform guide for the people running the deployment. It explains what a setting does, drafts and refines prompts, suggests scope and escalation boundaries for a new use case, and helps interpret what the session logs are telling you. When a support lead wants an agent for refund queries, they describe the job in plain language and get help turning it into a configured, guarded agent, without opening a ticket with anyone.

The compounding effect matters more than the convenience. Every change your team makes builds internal capability that stays with you. Under a managed model, the operating knowledge accumulates inside the vendor, which is precisely why the next change also has to go through them. Our system integration work follows the same principle: we establish the first connections to your CRM, ERP or document store, and your team extends the pattern to the systems that surface later, which they always do.

Governance Without a Consulting Retainer

It is worth being fair to the managed approach. The evidence that enterprises struggle with agent governance is real, the early named customers of these programmes are serious organisations, and hands on delivery genuinely helps a first agent reach production. Our disagreement is about what happens next. Four differences decide it:

• Speed of change. A self managed platform changes when your business changes. A managed one changes when the vendor has delivery capacity, and capacity is the resource being rationed.
• Where the knowledge lives. Internal capability compounds across every agent you build. Vendor capability compounds inside the vendor, and you rent access to it each time.
• Clarity of accountability. When you own the deployment, the configuration and the logs, there is no ambiguity about who is answerable for a decision the agent made.
• Predictable cost. Per deployment implementation pricing with no public reference point is difficult to budget against. A platform you run yourself has a cost you can forecast.

None of this requires you to be a large organisation with a machine learning team. It requires a platform that assumes competent business users rather than embedded consultants, and that ships the safety rails as standard. You can read more about the team building on that principle on our About page.

Owning the Loop, Not Renting It

The most interesting detail in the Presence launch is the improvement loop: production sessions and escalations are read back, changes are proposed, and the customer's team tests and approves them before rollout. That loop is the real asset in any agent deployment. It is how a system that works on day one still works in month eighteen.

The question every buyer should ask is who owns it. If the loop runs on your infrastructure, over your logs, driven by your people, it is a capability you keep. If it runs as a service, it is a subscription you service, and it stops the moment the arrangement does.

Aphelion is built for the first version. We set your agents up, apply the guardrails, and hand over a working deployment that your team owns, runs and improves. The expertise arrives once, at the start, and then stays with you rather than leaving with the engineers.

Frequently Asked Questions

What is a self managed AI agent?

A self managed AI agent is one your own team configures, supervises and changes, rather than one that only a vendor's engineers can touch. The supplier provides the platform, the initial setup and a working set of guardrails, and after that your staff decide what the agent knows, which systems it reaches, what it is allowed to do unsupervised and when it must hand over to a person. The distinction matters because it determines how fast you can adapt. A self managed agent changes when your business changes, while a vendor managed one changes when the vendor has capacity.

How does Aphelion AI set up and hand over an AI agent?

Aphelion deploys the platform inside infrastructure you own or exclusively control, builds your initial agents around the jobs you actually need done, connects them to your knowledge and systems, and applies our default guardrails before anything goes live. From that point the deployment is yours. Your team creates new agents, adjusts scope, edits prompts and tunes behaviour through the interface, with the Aphelion AI Agent on hand to explain settings and draft changes. There is no ticket queue and no waiting for a consultant's diary to clear.

What guardrails come as standard with a private AI agent?

Aphelion agents ship with a default guardrail set covering scope limits, permission boundaries drawn from your existing access model, escalation to a human at defined thresholds, and full session logging. Because the deployment is private, every prompt, document and response stays inside your governed environment rather than passing to a shared external platform, which simplifies GDPR, ISO and internal audit work considerably. Auditors review your own logs and your own controls instead of assessing a third party's infrastructure you cannot inspect.

Can a self managed AI agent integrate with existing business systems?

Yes, and integration is usually where agent projects succeed or stall. Aphelion treats system integration and data enrichment as core platform capabilities rather than bespoke consulting work, connecting agents to the CRMs, ERPs, document stores and databases you already run. We build the first connections during setup so the pattern is established, then your team extends from there. Because each new connection is configuration on a modular platform rather than a fresh development project, adding a system that surfaced late is incremental work instead of an expensive rebuild.

Self managed private AI versus a vendor managed deployment: which is better for enterprises?

A vendor managed deployment puts experienced engineers inside your project, which genuinely helps a first agent reach production, but it makes your roadmap dependent on the vendor's delivery capacity and leaves the operating knowledge outside your walls. A self managed private deployment carries a short learning curve at the start and then compounds, because every change your team makes builds internal capability rather than another invoice. For organisations that expect to run many agents over several years, self managed wins on speed, cost and control, and it removes the awkward question of who is accountable when the model supplier is also the implementation partner.