In-House ML Team, Big-Four Consultancy, or Embedded AI Squad: Who Ships Your AI?
Disclosure in the first sentence, as always: Conectia sells one of the three options below. We run embedded AI engineering squads, so read our take on the other two models knowing where we sit. We'd rather you catch us being fair than catch us being subtle.
The reason this comparison is worth your time is a number. MIT's Project NANDA reviewed 300+ enterprise AI initiatives and reported that 95% of generative AI pilots delivered no measurable P&L return — against an estimated $30–40 billion of enterprise investment. One report, not peer-reviewed, and the sample skews large-enterprise; take the precision with salt. But the shape matches what founders tell us in calls: the demo took a weekend, and production is nowhere in sight a year later. The same report found that pilots built with external partners reached deployment far more often than tools built internally from scratch — which makes who builds it the most consequential decision in your AI budget.
Before choosing who builds, it's worth naming what actually kills the 95%. A weekend demo and a production system differ by an enumerable list, not by magic: access to real data instead of a curated sample; evals that define what «good output» means before customers define it for you; security review of every tool the model can touch; cost ceilings so a retry loop can't spend your quarter; observability that tells you why an answer changed; and a person on call who owns the system when it misbehaves. Pilots stall because nobody budgeted for that list — and each of the three models below buys you a different subset of it.
You have three honest ways to buy that build. Each is a legitimate model optimized for a different owner of the outcome.
The three models at a glance
| In-house ML team | Global consultancy | Embedded implementation squad | |
|---|---|---|---|
| Unit of engagement | Permanent hires | Program / transformation | Senior engineers inside your workflow |
| Time to first commit | Months (hiring pipeline) | Weeks–months (scoping, staffing) | Days — Conectia matches in 72h |
| Annual cost, 3-person AI capability (US) | ~$1M+ all-in before infra | Program-priced, typically higher | 26–71% below equivalent local hires |
| Where knowledge lives after year one | Your company | Partly the firm's methodology | Your repo, your team |
| Built to optimize | Long-term ownership | Governance, scale, accountability | Shipping to production |
| Fails when | You can't hire or retain the talent | The mission is «ship this feature», not «transform this org» | Nobody internal owns the product |
The cost line deserves its footnote: 2025–2026 US salary surveys put a senior ML engineer between roughly $160K base and $350K+ total compensation at large tech companies. Three of those, plus payroll costs, recruiting, and the months of vacancy while you search, clears a million dollars a year before a single GPU bill — our arithmetic, stated as an estimate, not a survey figure.
In-house wins the long game — if you can afford the short one
Building your own ML team is the right call more often than a services company should admit. The knowledge compounds inside your walls. The team that built the system carries it through every model migration and every incident at 3 a.m. If AI is your product — not a feature of it — this is the destination, and every other model is at best a bridge to it.
The honest costs are time and fragility. The hiring pipeline for senior AI talent runs months, in a market where the strongest candidates field competing offers from labs paying compensation you won't match. A three-person team has zero redundancy: one resignation and your roadmap loses a quarter. And the MIT finding cuts hardest here — internally built tools were the pattern most likely to stall before production, because the first in-house team is usually learning production AI on your roadmap.
The global consultancy wins when the problem is the organization
The big firms — the strategy houses and the Big Four — get dismissed in engineering circles with a line we won't repeat as if it were analysis. The fair version: a global consultancy is built to optimize governance at scale. When AI has to land across forty business units, survive a board risk committee, and produce an audit trail a regulator will read, that machinery is the product, and nobody else has it. Enterprise procurement knows how to buy it, and there is real value in an accountable name the board already trusts.
The structural trade-off follows from the same shape. A model built to optimize governance prices everything through the program layer, and its scarcest resource — the people who actually write production code — sits several layers below the people who own your relationship. When the engagement ends, the methodology stays, but the engineers rotate to the next client. That is the shape working as designed: it optimizes for transformation, not for shipping. If your problem is a production system that has to exist in ninety days, you're buying the wrong optimization.
The embedded squad is built for one thing: production
The third model — ours, remember — embeds senior engineers directly into your team: your repo, your standup, your standards. At Conectia the squad arrives through a vetting process designed by active CTOs that passes 3% of candidates, matched to you in 72 hours, at a flat rate 26–71% below an equivalent local hire with compliance and payroll included, and a 14-day Pilot Sprint so you judge shipped output before committing to anything.
What this model refuses to do, stated plainly: it won't own the mission for you. An embedded squad amplifies an owner; it cannot replace one. If nobody in your company will own the product decisions — what to build, what «good» means, when to ship — the squad has nowhere to plug in, and a consultancy that takes the whole mission off your hands will serve you better. The knowledge is supposed to compound in your team, and that requires a team for it to compound in.
The honest decision line
- Build in-house when AI is the core product, you can win senior-talent offers, and you can absorb a two-quarter ramp. Everything else can be a bridge to this.
- Hire the global consultancy when the deliverable is organizational: multi-unit rollout, regulatory posture, board-level accountability. Buy the governance; don't expect startup shipping speed from a model that was never optimized for it.
- Embed a squad when the deliverable is a production system, you have (or want to keep) product ownership internal, and the constraint is senior engineering capacity — not strategy.
- The common hybrid we actually see: an embedded squad ships the first production system while your first in-house hires learn alongside it, and the squad's exit is planned from the start. The MIT 95% is mostly companies that chose none of the three deliberately.
Three questions to ask all three of us
- «Who exactly writes the code, and where are they in ninety days?» In-house: your employees. Consultancy: ask for names, not org charts. Squad: the same engineers, in your repo, until the planned handover.
- «What ships in the first thirty days?» A hiring plan, a discovery phase, and a Pilot Sprint are three different answers to the same question. All three are defensible; make sure the one you're buying matches your deadline.
- «When this ends, what do I own?» Code, documentation, evals, and the operational knowledge to run it — or a deck and a dependency. Get the answer in writing.
The 95% failure rate isn't an argument against AI; it's an argument against buying the wrong model for your situation and calling it a pilot. Pick the owner first — the model follows. If the owner is your own team and what's missing is senior hands, talk to a CTO; if the mission needs a bigger machine than ours, the other two doors are real, and we'll tell you so on the call.
