Separate AI company or in-house team: when it pays off

In short: a separate AI company makes sense when a business wants to build a new product or business model, with rules, costs and talent unlike the core business. To improve existing processes such as orders, customer service or marketing, a dedicated team running a measurable pilot, tied to real data and decision makers, is almost always enough.

Key takeaways

Spinning AI out buys speed and freedom from internal rules, but creates an integration problem to solve from day one.
A separate company pays off when AI becomes a product to sell, or an activity whose costs and risks need to be ring-fenced.
For an SME or a mid-sized B2B company, the most effective alternative is a dedicated team with a pilot on a real process.
Before the structure, choose the route: buy, work with a partner or build in-house. Each one calls for a different kind of team.
Separating the structure does not separate the accountability: governance, privacy and the AI Act still apply.

Why this matters now

According to Istat, Italy's national statistics office, 16.4% of Italian companies with at least 10 employees used at least one AI technology in 2025, up from 8.2% in 2024. Large companies went from 32.5% to 53.1%, SMEs from 7.7% to 15.7%: adoption doubled, but the gap widened. Among companies that considered AI without adopting it, the top barrier was lack of skills (58.6%), followed by legal uncertainty (47.3%) and missing or poor-quality data (45.2%).

That leaves an organizational question: who owns AI? Large groups answer with dedicated structures, sometimes new companies; online stores and B2B firms face it on a smaller scale. It is a recurring theme in the AI for business category.

What is a separate AI company, and how is it different from an internal unit?

A separate AI company is an independent legal entity, such as a spin-off or a subsidiary, with its own accounts, governance and sometimes its own shareholders, that builds AI solutions for the group or for the market. An internal unit has a dedicated budget and leader but stays inside the same company.

Between those two extremes there are lighter options:

An AI center of excellence (AI CoE) is an internal team of experts that sets strategy, standards and priorities so that AI adoption does not fragment across departments.
A dedicated team is a small group of people, with part of their time protected, tasked with getting one or two use cases into production.
A pilot project is a test limited in time and scope, with a metric and a success threshold agreed before it starts.

Why do companies create a separate company for AI?

To work with different rules and timelines from the core business, to ring-fence costs and risks, and to attract talent or capital to a project with a life of its own. Processes built to protect the existing business slow down experiments with models that change every few months.

The expected benefits are two. The first is speed: a unit with its own rules runs pilots, tests and quick iterations, and a failure stays small and contained instead of weighing on the main business. The second is focus: dedicated people and budget are not swallowed by day-to-day operational emergencies. Both are real, but they only pay off if the results flow back into the parent company's processes.

A recent example: in June 2026 Snap turned its internal generative AI video team into a separate company, Dotmo, citing the high cost of doing that work in-house and the fact that the activity is not among its current core priorities. Snap keeps a large equity stake, and Dotmo may raise outside funding later, which an internal department cannot do.

Separation is not forever, though. In April 2023 Google merged DeepMind and the Brain team from Google Research into a single unit, Google DeepMind, to accelerate progress. Once AI becomes central to the main product, being close to the business matters more than independence.

When does a separate AI company make sense, and when does it not?

It makes sense when AI is the product, not a tool to improve the product. If the goal is to cut quote turnaround time or answer customers faster, spinning AI out creates more problems than it solves.

Microsoft's Cloud Adoption Framework agrees: an AI center of excellence usually builds on existing teams, and a standalone AI team is warranted only if current teams cannot support adoption or critical risks exist.

Model	When it makes sense	Effort	Main risk
Separate company	AI sold as a product, outside partners or investors, risks to ring-fence	High	Isolation from the business and its data
Dedicated business unit	AI serving several brands, product lines or countries in one group	Medium-high	Clashes over priorities and budget
AI center of excellence	Many projects needing shared standards	Medium	Becoming a bottleneck
Dedicated team	The first one to three use cases in a mid-sized company	Low-medium	Protected time that exists only on paper
Pilot project	Testing a use case before investing	Low	Staying an experiment with no follow-up

Before setting up a separate company, check four conditions: an offer someone outside the group would pay for, a budget that can sustain years without quick returns, a team with proven skills, and a clear decision on who owns data and technology. If one is missing, the lighter option is almost always right.

Should you build AI in-house, buy it or work with a partner?

Most companies should start with existing tools or a specialized partner, and build in-house only what creates a distinctive advantage. The choice shapes the structure: if you buy or partner, you need a team that integrates and governs, not a company that builds models.

One signal comes from The GenAI Divide: State of AI in Business 2025, a report by MIT's NANDA project covered by Fortune in August 2025: only about 5% of the generative AI pilots studied achieved rapid revenue acceleration, while the vast majority stalled with little to no measurable impact on P&L. According to the authors, the problem is not model quality but integration with business processes. Tools bought from specialized vendors and partnerships succeeded about 67% of the time, while internal builds succeeded only about one-third as often. The study also found that more than half of generative AI budgets go to sales and marketing tools, while the biggest returns come from back-office automation. It is based on interviews, a survey and an analysis of public deployments, so read it as a directional signal, not a rule.

Option	When it makes sense	What you need in-house	Example
Buy software with built-in AI	Standard processes: customer service, copy, document analysis	An owner who configures it, trains people and checks results	Help desk with suggested replies for an online store
Specialized partner	Custom processes linking ERP, online store and CRM	A business owner and someone who knows the data and systems	Automatic reading of requests for quotation for a B2B distributor
In-house build	AI that becomes a distinctive advantage on your own data	A stable technical team and a multi-year budget	Demand forecasting on years of sales history
Separate company	AI sold on the market, outside investors	Governance, contracts on data and intellectual property	A spin-off such as Dotmo

For practical examples of back-office automation, see the article on business process automation.

What lighter alternatives does an SME have?

A dedicated team with a written mandate and a measurable pilot: it costs little, stays close to data and customers and, if it works, extends to other processes.

The minimum team: a business owner accountable for the process, someone who knows the data and systems (ERP, online store, CRM), a privacy contact and, if needed, an external partner. The stack can stay simple: an orchestrator such as n8n or Make, a model such as OpenAI's via API, and the data the company already has. As I explain in the article on enterprise AI and operational context, the value comes from connecting AI to real processes.

Two hypothetical scenarios:

E-commerce. A homeware store with a few thousand products uses AI to draft product pages from supplier data and replies about orders and returns, with human approval. It tracks writing hours per page and time to first response.
B2B. An industrial components distributor extracts part numbers and quantities from emailed requests for quotation, checks availability in the ERP and prepares a draft quote for the sales rep. It tracks time from request to quote and coding errors.

For small businesses the path is shorter, as I describe in the article on the future of AI for small businesses. If budget is the constraint, look at public funding for AI in SMEs.

Who should lead AI in a company?

Top management, with a business owner for each project. When AI becomes strategic, even companies that give it its own structure connect it directly to the people who make decisions.

In March 2024 Microsoft created a new organization, Microsoft AI, focused on Copilot and its other consumer AI products and research: its CEO joined the senior leadership team, reporting directly to Satya Nadella. The structure is autonomous, but not far from where decisions are made.

A mid-sized company or an SME does not need a new executive. The sponsor is the owner or CEO, who sets priorities and settles conflicts between departments; each project has a business owner who owns the process and is accountable for it. Involving the core business people who will use the solution from the start is the most practical cultural lever, because the tool is shaped around their needs instead of arriving ready-made.

Should AI skills be centralized or distributed?

Centralize first, then distribute. Microsoft's Cloud Adoption Framework says companies at an early stage benefit from a centralized center of excellence that consolidates expertise and foundational practices and speeds up adoption; as adoption matures, the center should move to an advisory role that supports AI use across departments. A centralized model ensures control and consistency, while an advisory approach provides flexibility.

A common middle ground is hub-and-spoke: a small central core (the hub) manages standards, security, vendors and shared data, while contacts in each department (the spokes) propose, use and improve solutions. In a mid-sized B2B company the hub can be the dedicated team described above, and the spokes the contacts in sales, customer service and finance.

What are the risks of separating AI from the business?

The main risk is building solutions the business never adopts: value appears when a real process gets faster, cheaper or sells more, not when a prototype ships. Other risks:

Data access: if another group company does the development, you need contracts, permissions and clear privacy roles, and everything takes longer.
Duplicated costs: administration, licenses, infrastructure and security are managed twice.
Two cultures: different rules and incentives strain collaboration with day-to-day operations.
Accountability: under the AI Act, whoever uses an AI system has obligations of their own, even if another group company built it.

According to the European Commission, bans on prohibited AI practices have applied since 2 February 2025, while after the AI Omnibus the rules for high-risk uses, including recruitment and worker management, apply from 2 December 2027. To give governance a structure, NIST's AI Risk Management Framework is a useful reference: a voluntary framework released in 2023, with a generative AI profile added in 2024. I go deeper in the article on AI governance and explainability; for AI in people management, see also the article on AI, HR processes and compliance.

How do you keep the bridge between the AI unit and the core business?

With rules agreed before you start, not after. The risk of isolated innovation applies to a separate company, a business unit or an internal team alike: only the length of the bridge changes.

An internal customer for every project: a business owner who asks for the solution, uses it and signs off on the result.
Shared business metrics: the AI unit is measured on process time, cost or sales, not on the number of prototypes.
Core business people on the project: those who run the process take part in tests and reviews, with protected time.
Rules for going live: who runs the solution after the pilot, with what budget and what support.
A regular check-in with top management: to decide priorities, roll-outs and shutdowns.

The MIT report cited above points the same way: generic tools work for individuals but stall in companies because they do not adapt to workflows. That is why the bridge has to be designed from day one.

Who owns the data, models and code if AI moves to another company?

Settle it in a contract before you start. When one group company builds or runs AI for another, you need clear agreements on three points: ownership and licensing of code and models, use of the parent company's data, and liability for errors.

On personal data, the GDPR offers two reference points. Recital 48 recognizes that companies in a group may have a legitimate interest in transmitting personal data within the group for internal administrative purposes, including client and employee data. But if the AI company processes personal data on behalf of the parent, Article 28 makes it a processor: the relationship must be governed by a contract or other legal act setting out the subject matter, duration, nature and purpose of the processing, the type of personal data and the categories of data subjects. The same applies to an external partner or vendor. This is general guidance: for your actual setup, ask a lawyer and, if you have one, your DPO.

How to apply it in your company: first steps

Pick the process, not the structure. Among repetitive tasks with data already in your systems, pick one with a number to improve, such as quote response time.
Write the mandate down. Appoint a business owner and set a fixed share of weekly time for three or four people.
Agree on the metric and threshold before you start. Measure the baseline and decide which result justifies scaling up; the article on measuring the value of AI offers a framework.
Run the pilot on real data for eight to twelve weeks, with human approval of everything that reaches customers.
Decide at the end: scale, fix or stop. Consider a dedicated unit or a separate company only when several projects have worked and AI is turning into an offer of its own.

It is the same approach behind the automation and AI solutions I build for online stores and B2B companies: one process, one number, one owner.

Mistakes to avoid

Building the structure before the use case: without a concrete problem, a department hunts for justifications, not results.
Separating without a bridge: every project needs someone in the business who uses it and is accountable for it.
Ignoring shadow AI: shadow AI is the use of AI tools the company has not approved, often with business data pasted into a chat; a dedicated team also exists to offer safe alternatives.

Frequently asked questions (FAQ)

What is a separate AI company?

It is an independent legal entity, such as a spin-off or subsidiary, created to build AI solutions with its own budget, governance and sometimes shareholders. It can serve its group or sell its technology on the market.

When does a separate AI company make more sense than an internal team?

When AI becomes a product to sell, when outside capital is needed, or when costs and risks must be kept away from the core business. To improve orders, customer service or marketing, an internal team stays closer to data and customers and costs less.

Does an SME need a dedicated AI department?

Usually not. A team of three or four people with a written mandate and protected time is enough, possibly supported by an external partner.

How long should an AI pilot project last?

Eight to twelve weeks on real data is usually enough for a process with steady volumes, such as quote requests or support tickets. What matters more is agreeing on the metric, the baseline and the success threshold beforehand.

What is the difference between an AI center of excellence and a dedicated team?

A center of excellence coordinates adoption across the whole company through strategy, standards and priorities. A dedicated team gets one or two use cases into production. Most companies start with the team.

Does a separate company remove AI Act obligations?

No. The AI Act distinguishes between whoever develops an AI system and whoever uses it, and the user has its own obligations even if the system comes from another group company. For high-risk uses such as recruitment, the rules apply from 2 December 2027.

Is it better to build AI in-house or buy it?

For standard processes, start with existing tools or a specialized partner; building in-house makes sense when AI becomes a distinctive advantage or a product. According to the MIT report covered by Fortune, purchased tools and partnerships succeed far more often than internal builds.

Sources

To find the right setup and first pilot for your company, book a free assessment.