Three engagement models: a managed service, a Center of Excellence, an embedded pod. One rule applies to all three, which is that what done means is agreed and baselined before any work starts.
Most AI engagements are sold as access. You buy licenses, or you buy hours, and what happens after that is your problem. We sell the other side of it: an agreed result, measured against a baseline both sides signed before anyone started building. What varies is not whether we commit to an outcome. It is who ends up owning the work once we have.
That question has three honest answers, so we run three engagement models. A managed service where we own the outcome outright. A Center of Excellence engagement where the capability transfers to your teams and we leave. An embedded pod where your roadmap stays yours and we supply the practitioners who have built these systems before.
None of them is the discount version of another. They answer different questions about your organization: do you want this capability, or do you want the result it produces?
Figure 1: who is on the hook, and when that changes
This is the model for organizations that would rather hold a partner accountable for a result than a vendor accountable for uptime. We take end-to-end ownership of an agreed scope of work, whether that is a domain, a workflow, or a set of engineering functions, and run it.
The scope is a process, not a product footprint.
We take ownership of the work itself: the claims that have to clear, the files that have to close, the tickets that have to be resolved. Not a seat count, not an environment, not a support tier.
Targets are contracted before the engagement starts.
Cost, quality, speed and reliability are baselined against what the process does today, then written into the agreement as commitments. Contracted, not projected. The distinction is the whole model.
Our team, our platform and our agents run it together.
AIOS carries the execution, our practitioners own the exceptions, and human oversight is a step in the process rather than a promise in a deck. You see the same execution record we do.
Quarterly reviews are held against the agreed numbers.
If the outcomes are met, the review is short. If they are not, that is our problem to resolve, and it shows up in our economics that quarter, not at a renewal meeting a year later.
This is the model our pricing was built for. When a task completes against locked criteria, we earn the outcome fee; when it does not, we do not. How that is measured and billed is in Pricing what you can prove.
Some organizations do not want a partner running the work in five years. They want their own teams doing it, to their own standards, across domains we will never touch. That is a different engagement, and pretending otherwise would just build a dependency we would later have to defend.
We stand up and operate a Center of Excellence.
Governance, agent design patterns, toolchain standards and delivery practices: the parts that are hard to invent once and impossible to invent five times in five business units.
Your teams build capability by delivering, not by attending.
Practitioners work side by side on live engagements. Training programs produce certificates. Shipping a governed agent into production with someone who has done it before produces engineers.
The COE produces reusable patterns and blueprints.
Everything the COE builds is written down as something another team can adapt to their own domain and workflow. The output is a library, not a set of one-off implementations.
There is a defined handover point.
The engagement is designed to end. Success is internal capability that sustains itself after we leave, on a date that was named at the start.
The third case is the simplest to describe and the easiest to get wrong. You already have a roadmap or a platform. You do not need a strategy engagement or a Center of Excellence. You need people who have built enterprise Agentic AI systems before, working inside your delivery cycle at your cadence.
A pod of engineers who have done this in production.
Direct experience building enterprise Agentic AI systems: the failure modes, the governance conversations, the integration work that is never in the demo.
The pod operates inside your structure.
Your sprint cadence, your toolchain, your governance. The pod is accountable to your delivery milestones, which means your engineering leadership stays in charge of sequencing.
Scope is a working system, not a staffing level.
The engagement is defined around a specific outcome: a system running in production, with the architectural and operational decisions documented well enough for your team to own it.
Commercials are sized to the initiative.
Fixed capacity, time and materials, or a structure built around the milestone, whichever fits the initiative. The pricing model flexes; the definition of done does not.
| Managed Service | COE Enablement | Engineering Pod | |
|---|---|---|---|
| Who is accountable | We are, for the result | Shared, then you are | You are, for the roadmap |
| Who runs the work | Our team, platform and agents | Your teams, with our practitioners | Our pod, in your delivery cycle |
| Where capability lands | With us, under contract | With your COE, by design | With your team, documented |
| How scope is set | A domain or workflow, end to end | Governance, patterns, standards | A specific system in production |
| How it ends | Renews against performance | A defined handover point | The shipped system |
Whichever model you pick, the first thing we do is measure what the process costs to run today: volumes, handoffs, exception rates, cycle time, error rate, and what all of that consumes in people and time. Nothing is committed before that assessment, because a commitment made against a number nobody measured is a guess with a signature on it.
Figure 2: the front door
The baseline is the artifact that matters. It is signed before work starts, and it is what both sides are judged against afterward. It is also the reason we occasionally end an assessment by telling an organization that the process they brought us is not a fit. Once a baseline exists, that conclusion is a fact rather than an opinion.
We would rather lose the deal at the assessment than win it and spend a year arguing about what the number was before we arrived.
Mitchell Gunnels, founder, cvlSoft
The failure we see most often is an organization buying capability transfer when what it actually wanted was the result. The COE gets stood up, the patterns get written, and then the teams who were supposed to absorb the capability are still carrying their day jobs. The engagement was not wrong; it was aimed at the wrong question.
The reverse happens too. A team with a strong platform group and a live roadmap does not need us to own the domain. It needs three engineers who have already made these mistakes, inside the sprint, hitting the milestone. Selling that team a managed service would be selling them accountability they were never going to hand over.
So we ask two questions before proposing anything. Where should this capability live in three years, and who is going to be accountable when it misses? The first answer picks the model. The second one is what we put in writing.
Across all three: outcomes are defined before work starts, measured against a baseline both sides agreed to, and reviewed on a fixed cadence against the same evidence. Human oversight is built into the operating model rather than bolted on for the audit. And the work runs on the same platform and the same execution record in every case, so what we tell you happened is what the system recorded.
The models exist because organizations are different. The commitment does not vary, and it is the only part of this we will not negotiate: we agree what done means before we start, and we are measured against it.
Sources
We embed until it works, then you pay for what worked. Bring the process you would most like to stop staffing.