Most organisations looking for an enterprise AI company are not looking for a model. Models are widely available. They are looking for an engineering partner that can make AI work inside an operation that already exists, with its own systems, its own data, its own regulators and its own ways of failing. This guide sets out how to judge an AI engineering company in the USA or elsewhere, and what to ask a custom AI development company before you sign.
What an enterprise AI engineering partner actually does
The model is usually the smallest part of the work. The larger part is everything around it: reaching the data where it lives, deciding what the system may see for each role, connecting to the systems of record, writing down what it may do on its own and what goes to a person, measuring whether it is good enough, and keeping it working after launch.
That is why an enterprise AI engineering partner should also be a software company. Integration, data layers, access control and custom applications are where most projects succeed or fail. Ask to see that work, not only the AI.
Sector experience that transfers
No partner has built your exact system before, but the patterns carry across sectors when the engineering is sound. Look for evidence across the kinds of operation that resemble yours, and ask what transferred.
Opsian's published work spans a deliberately broad set of sectors:
- Defence: defence work in the United States, alongside cybersecurity and surveillance enhancement for the Royal Thai Navy delivered through Section 31 International.
- Finance: a machine learning voice agent on live regulated calls for Capital One, with policy-bounded actions and escalation to a person with the full context of the call.
- Retail: Gap Inc.'s first internal GPT, with retrieval over internal documents, access control per business unit and an evaluation harness.
- Logistics: a logistics network inside Expeditors' own data centre and an AI-assisted transportation management system for On The Rock Logistics.
- Manufacturing: an AI-enabled ERP with K.C. Metalsheet across sales, procurement, warehouse and production.
- Healthcare: a proposed hospital intelligence layer and the NavCare staffing prototype at Navavej International Hospital.
The common thread is not the industry. It is operations where decisions have to be explained afterwards. The industries section lists the sectors Opsian builds for and where clients and partners already exist.
Where delivery and infrastructure sit
Ask where the engineering team is, where support comes from, and where the infrastructure runs. These are three separate questions, and the answers affect response times, contracts and data protection.
Opsian is a US company. Its delivery and infrastructure are based in the United States, and it delivers for clients wherever they operate. The location of the team and the location of the data are separate decisions: where data must stay in a particular country or inside a client's own facility, systems can be deployed privately, on premise or in-region. The Expeditors network, built inside the company's own data centre, is an example of the second.
How the engagement is structured
Good partners make it easy to start small and stop cleanly. Opsian works in three models:
- Audit: a structured examination of the operation, delivered as a document the client owns and can take to any builder. Most engagements start here.
- Build: architecture, engineering and delivery with a fixed scope taken from the audit findings, an evaluation harness and handover documentation.
- Embedded team: Opsian engineers operating as the client's AI function on an ongoing basis.
Ask who you will actually work with. At Opsian, clients work directly with the company's leadership and the engineers building the system, from the first call to live operation, rather than through a chain of handoffs.
Governance and evaluation built in
In a regulated operation, the boundary around an AI system matters more than its fluency. Ask how the partner decides what a system may do, how permissions are enforced, how every action is recorded and how exceptions reach a person. Our note on AI agents and governed AI covers this in detail.
Then ask how quality is measured. The answer should be an evaluation set built from real questions and expected answers, run before launch and after every change to the model, the prompts or the data. Without it, nobody knows whether last week's change made the system better or worse.
Support after launch
An AI system changes behaviour as the data around it changes, as model versions change and as the operation asks more of it. Ask what happens after the launch date: monitoring, on-call response, scheduled re-evaluation, cost tracking and incident reviews.
Opsian's Managed AI Operations runs what it ships, under a service level agreement, with an on-call engineer, runbooks for known failures and a written review after every incident. Engineering, support and the infrastructure it runs on are based in the United States.
Questions to ask in the first meeting
- Which of your published case studies is closest to our operation, and what was its status: proposal, prototype, pilot or live?
- Where will our data be processed and stored, and who can reach it?
- What will the system be allowed to do on its own, and what goes to a person?
- How will you measure quality before launch and after every change?
- Who supports it at three in the morning, and under what terms?
- If the audit says we should not build, will you say so in writing?
Where to start
If you are comparing partners, commission an audit first. It gives every vendor the same written findings to answer, and it gives you a document you keep whichever way you decide. Opsian's anonymised sample audits show the finished document for hospitality, healthcare, manufacturing, professional and financial services, retail, and construction. When you are ready, book an audit.
Questions
- What does an enterprise AI company do?
- It designs, builds and supports AI systems inside an organisation's existing operation, connected to its own data and systems, with the permissions, records and evaluation that a regulated or high-volume environment requires. The work usually covers custom software and integration as well as models.
- What is the difference between custom AI development and buying an AI product?
- A product solves a defined problem the same way for every customer. Custom AI development fits the system to your processes, data, obligations and existing software. Many engagements combine the two, with released products where they fit and custom work around them.
- What does US-based delivery mean for clients outside the United States?
- At Opsian, the engineering, support and infrastructure are based in the United States, and clients can be anywhere. Where a client's data has to stay in a particular place, systems can be deployed privately, on premise or in-region, so the location of the team and the location of the data are separate decisions.
- How long does a custom AI project take?
- It depends on scope. As one published example, the AI-enabled ERP for K.C. Metalsheet is planned in five phases from the requirements and data audit to go-live, with a target of full launch within six months of activation. An audit first gives you a timeline built on your own operation.
