guide
How to hire an ML engineering agency (2026)
The questions that separate an engineering team from a prompt wrapper, and what a serious proposal contains.
Read// build vs buy
Off-the-shelf AI tool vs Custom ML build
This decision gets made badly in both directions. Teams buy a SaaS tool for a problem that is specific to their data and spend a year fighting the integration. Other teams commission a custom model for a problem that three off-the-shelf products already solve well.
The deciding factor is almost never sophistication. It is whether your data looks like the data the vendor trained on, and whether the decision the model makes is close enough to your economics to be worth owning.
Side by side
Nine dimensions that decide it.
| Dimension | Off-the-shelf AI tool | Custom ML build |
|---|---|---|
| Time to first value | Days to weeks. Configuration, not construction. | Weeks to months. Data audit, feature work, evaluation, shadow deployment. |
| Up-front cost | Low. Subscription, sometimes with an onboarding fee. | Higher. A scoped engineering project before anything runs. |
| Cost at scale | Grows with usage or seats. Can become the largest line item. | Largely fixed after build. Compute cost of a boosted-tree model at business volume is small. |
| Fit to your data | Good if your schema resembles the vendor's assumptions. Poor if you run a self-written platform. | Built against your actual schema. This is the whole point. |
| Explainability | Varies. Many vendors will not disclose how the score is derived. | Per-feature attribution on every prediction, because you own the pipeline. |
| Ownership and lock-in | Vendor owns model and pipeline. Leaving means losing the capability. | You own code, weights, and pipeline. Nothing stops working if the supplier goes away. |
| Who maintains it | The vendor. Updates arrive whether you asked for them or not. | You, or a retainer. Needs a named owner and a retraining cadence. |
| Data residency | Usually the vendor's cloud. A blocker in regulated verticals. | Your infrastructure or your cloud account, by default. |
| Handles your edge cases | Only if the vendor's roadmap agrees they matter. | Yes - your edge cases are the specification. |
How to decide
Pick by situation, not by preference.
Buy off the shelf when
- Your problem is a standard one - email classification, transcription, generic lead scoring on a standard CRM schema.
- Your data lives in a mainstream platform the vendor already integrates with.
- You need something working this month and the decision is reversible.
- You have no one to own a model in production, and no plan to hire one.
- The decision the model informs is low-value per case. Accuracy gains do not pay for a build.
Build custom when
- Your platform is self-written and no vendor's schema assumptions fit it.
- The score decides where real money goes, so explainability is a requirement rather than a nice-to-have.
- Data cannot leave your infrastructure for regulatory reasons.
- You have six-plus months of clean history and a specific label your business cares about.
- The per-decision value is high enough that a few points of accuracy pays for the engineering.
Our bias, stated
We build custom ML, so treat this page with the appropriate scepticism - and then check it against the list above. We regularly tell prospects on a first call that an existing tool covers their case and a custom build would be a worse use of their money. A build that should have been a subscription is a bad outcome for both sides, and it is the fastest way for an agency to lose a client's trust.
Questions
Follow-up questions.
How long does a custom ML build take?
For a scoped tabular prediction problem with usable history: two to four weeks for the data audit and adapter, then four to eight weeks for model, evaluation, and shadow deployment. The audit phase is where most timeline surprises surface, which is why it is priced and delivered separately.
Can we start with an off-the-shelf tool and move to custom later?
Often the best sequence. The SaaS tool proves whether the workflow gets adopted at all, and the period of using it produces labelled outcome data that makes the eventual custom model better. Just keep your own copy of the raw data from day one.
What if our data is not clean enough?
Then the honest answer is that data work comes first and the model comes second. A data audit that concludes 'not yet' is a real result. We would rather deliver that in phase one than build a model on history that cannot support it.
Still not sure which side you are on?
Thirty minutes, no pitch. We will tell you which of the two your situation actually points to - including when the answer is the one we do not get paid for.
Book a 30-min call