In 2026 the question "should we use AI?" turned into a harder one: who should build it? The answer used to be a choice between buying a tool and hiring a consultancy. Then OpenAI, Anthropic, AWS and Microsoft all launched units that send engineers into their customers' offices, and forward deployed engineering became a category of its own. Here's an honest comparison of the five routes, what each is good at, and the questions that separate a real partner from a good pitch.
The five ways to get AI built
- Buy an off-the-shelf AI product. A SaaS tool with AI built in: a support chatbot, a transcription service, an AI feature in software you already use.
- Hire a strategy consultancy. An assessment, a roadmap and a business case, usually followed by a separate build.
- Commission an agency or freelancer. A team that builds to a brief you provide.
- Build an in-house team. Hire AI and software engineers and own the capability yourself.
- Work with a forward deployed engineering partner. Engineers who come to your office, decide with you what to build, then build it, deploy it and hand it over.
How they compare
| Route | Good at | Watch out for |
|---|---|---|
| Off-the-shelf product | Speed and cost for common, standard problems | It won't bend to your process, and your data sits on the vendor's terms |
| Strategy consultancy | Organisation-wide priorities, governance and change management | A deck is not a system; the build is someone else's problem |
| Agency or freelancer | Building a well-defined thing to a clear brief | It's only as good as the brief, and you have to write it |
| In-house team | Long-term capability and deep product knowledge | Slow to hire, and one resignation can stall everything |
| Forward deployed partner | Deciding what to build and building it, in one pair of hands | Needs real time from your people during the visit and the build |
When each route is the right one
Buy off the shelf when the problem is common and your process is standard: meeting transcription, spam filtering, proofreading. If a product already does the job, we'll say so rather than quote for a build.
Hire a consultancy when the question is organisational rather than technical: which business units go first, what governance you need, how to bring several thousand people along. Just budget separately for whoever will build the result.
Use an agency when you know exactly what you need and can specify it: screens, integrations, acceptance criteria. The risk sits in the specification, not the code.
Hire in-house when AI is becoming part of your product or a permanent operating capability, with years of roadmap ahead. Expect a long search: the skills are scarce, and the same engineers are being recruited by OpenAI, Anthropic, AWS, Google and Microsoft, all expanding their forward deployed teams at once (TechCrunch). Building alone is also the riskier route: MIT's 2025 research found that buying from specialised vendors and working with partners succeeded about 67% of the time, while internal builds succeeded only a third as often (Fortune). Many companies start with a partner and hire once the value is proven, so the first hire inherits a working system instead of a blank page.
Choose a forward deployed partner when you know the pain but not the solution, the work crosses several people and systems, and you want working software in months rather than a plan in weeks. The person who studies your process also writes the code, so nothing gets lost between the diagnosis and the build.
You can mix the routes
These options aren't exclusive, and the best setups usually combine them. Buy the commodity tools, including the AI features already in your office software. Use a partner for the processes that are specific to your business and worth owning. Hire once there's a working system and a roadmap to keep someone busy. The mistake is using one route for everything: buying a tool for a process it doesn't fit, or hiring a team before you know what it should build.
Why the big AI companies may not be your first call
If forward deployed engineering is the answer, why not go straight to OpenAI's Deployment Company, Ode, AWS or Microsoft Frontier Company? For some organisations you should. But they're built around global enterprises, such as Microsoft's launch clients Unilever and Novo Nordisk (Microsoft), and, in OpenAI's case, around the more than 2,000 businesses its partners back (OpenAI). Ode is the exception on size, aimed at mid-size companies, though its private equity backers will steer their own portfolio companies to it (TechCrunch). We look at who each programme serves in our guide to forward deployed engineering.
The other question is technology. Three of the four build on their own stack: OpenAI's engineers connect OpenAI models, AWS deploys into your AWS environment and Ode is built around Anthropic's Claude, while Microsoft says its approach is multi-model. A specialist partner can be model-agnostic, choosing the model and the hosting to fit your data rather than a vendor's roadmap, including European or self-hosted models when the data is sensitive. It's also close enough to spend days in your office rather than an hour on a video call.
Ten questions to ask any AI implementation partner
- Will you see the work before you quote? For anything that spans several people or systems, a partner who scopes from a sales call alone is guessing.
- Who writes the code? The person in the meeting, or someone you'll never meet?
- Who owns the code, the data and the prompts? It should be you, in writing.
- Where is our data processed and stored? Which country, which provider, and under which processor agreement?
- Which models will you use, and why? A good answer depends on your data and your risk, not on a partnership badge.
- How will we measure success? Against which baseline, taken when?
- What happens when the model is wrong? Confidence thresholds, human review and fallbacks should be designed in, not promised later.
- How do you test changes? Look for evaluation sets built from real examples, as in our guide to evaluating LLM quality.
- Is the price fixed? Open-ended hourly billing puts all of the risk on you.
- What does the handover look like? Documentation, training and access, so you can run it yourselves or move to someone else.
If a partner hesitates on the first three, keep looking. They're the questions that reveal whether you're buying engineering or a sales process.
Red flags in an AI proposal
- The use case will be found later. An open-ended discovery phase has a way of quietly becoming the project.
- The demo runs on public data. Anything looks good on clean sample data. Ask to see it on a handful of your real, anonymised cases.
- Success is measured in usage. Logins and prompts sent aren't outcomes. Hours saved, errors avoided and faster responses are.
- Pricing per agent or per seat before anyone has seen your process. It tells you the product came first and your problem second.
- Nobody asks about your data. If personal or patient data doesn't come up in the first conversation, it will come up later, expensively.
- The exit is vague. If leaving means losing your prompts, your data or your integrations, you're renting, not owning.
Where Neurova fits
We work as a forward deployed engineering partner, from Eindhoven, for clients across Europe. If the work spans several people or teams, the engineer who will build your system spends one to three days in your office first, depending on how many teams are involved. Either way, you get a written scope with a fixed price and a date. Weekly demos on real data follow, then deployment into your environment with staff training. You own the code, and your data stays in Europe, handled to the standards we apply to medical software.
What we look for during those days is set out in which business processes to automate with AI first, and the mistakes we're there to avoid are in why most AI pilots never make it to production.
The right route depends less on the size of your ambition than on how well you know the problem and who will own the result. Neurova AI builds custom AI applications, agents and integrations end to end. If you'd like to meet the person who would build yours before deciding, ask for an on-site visit.
Frequently asked questions
What is the difference between a forward deployed engineer and an AI consultant? An AI consultant typically assesses your organisation and recommends what to do, leaving the build to someone else. A forward deployed engineer looks at the work on site and then builds, deploys and supports the system. The consultant's deliverable is advice; the engineer's is working software.
Should we hire an in-house AI engineer or use a partner? Hire when AI is a permanent part of your product or operations and there is a roadmap to keep a team busy for years. Use a partner when you need one to three systems built and running, or want proof of value first. Many companies start with a partner and hire once the work is proven.
What should we ask an AI implementation partner before signing? Ask whether they will see how the work is done before quoting, who will write the code, who owns the code and data, where data is processed and under which agreement, how success will be measured, what happens when the model is wrong, and what the handover looks like if you part ways.