Build vs buy vs embed: choosing an enterprise AI delivery model
The short answer
Buy when the workflow is a commodity and you hold no data advantage in it, build when the workflow is specific to you and you have an engineering bench you can permanently assign, and embed when the workflow is yours but the bench is not. Build, buy and embed are three different answers to one question: who ends up owning the intelligence.
That third case is the common one in a European mid-cap, and it is the one the standard comparison has no column for. The same comparison answers whether to build or buy AI agents, because the agent is the last layer of the system, and the delivery question is who builds and owns what sits underneath it.
The same rule applies one layer down, to the context layer every agent reads. Use the layer that comes with the data platform you already run, treat its contents as a build, and own the cross-system part no platform sells.
Key takeaways
- Build vs buy is an operating-model decision, not a procurement one. The question is not what a system costs, it is who is inside your stack while it is built and who owns it afterwards.
- Buy is correct more often than build-side vendors admit. Three conditions have to hold together: commodity workflow, no proprietary data advantage, and acceptance of the vendor’s roadmap.
- Build estimates price the first version. The integration surface, the governance layer and the maintenance of both are the actual cost, and they arrive in year two.
- An agent platform sold without engineering attached is a licence, not a deployment. Somebody still has to reach the legacy database, the identity provider and the data-residency rule.
- Decide it per workflow, not once for the company. The same company should correctly buy, build and embed in the same year.
- The context layer is bought as a container and built as content. Palantir, Microsoft, Databricks and Snowflake each sell the container, and SAP sells one on the ERP side through Business Data Cloud. None writes your definitions, and none sees the decisions taken across its boundary.
The two-column table is the wrong table
Every build-vs-buy guide compares the same three rows: cost, time to value, control. Both columns quietly assume the same delivery shape, that someone finishes a system and walks away from it. Build hands it from your project team to your operations team. Buy hands it from a vendor to your administrators. The handover is treated as an event rather than as the place where the value is decided.
It is the place where the value is decided. Enterprise AI does not usually fail at the model, it fails at the seam between the system and the company, which is where most pilots stop. A comparison that leaves out who owns that seam is comparing invoices, not delivery models.
So the useful table has a third column, and its rows are different: who works inside your systems during the build, what exists at the end that you own, and who is accountable in month thirteen.
| Buy | Build | Embed | |
|---|---|---|---|
| Who works inside your systems during the build | The vendor’s implementation team, or your administrators configuring their product | Your own engineers | External engineers inside your stack, next to the people who hold the process knowledge |
| What you own at the end | A licence and your configuration of it | The system, its code and its evaluation sets | The system, its code, prompts and evaluation sets, and the written-down process knowledge, named in the contract |
| Where the context layer lives | In the AI product’s own schema; a data platform’s layer is covered below | In your stack, maintained by your team | In your stack, named in the contract, with an internal owner before handover |
| Who chooses the model | The vendor | You | You, and the system is built to run on a different one |
| Who carries the year-two cost | The vendor, priced into the licence | You, indefinitely | You, with an internal team that has run the system before handover |
| Accountable in month thirteen | The vendor for the product, you for the outcome | You | You, after a named handover |
| Right when | The workflow is a commodity and you hold no data advantage in it | The workflow is yours and the bench is permanently assignable | The workflow is yours and the bench is not |
| Wrong when | The process is not generic | The engineers are borrowed from another product | The workflow is a commodity, or nobody internal is assigned to receive the system |
The sections below take each column in turn.
What the 76% “purchased” figure says, and what it does not
Menlo Ventures found that 76% of enterprise AI use cases are purchased rather than built internally, up from a near split a year earlier, when 47% were built and 53% purchased. Read on its own, it looks like a verdict: the market has decided, buy.
Read the method before reading the verdict. The survey covered 495 US enterprise AI decision-makers in November 2025. It counts use cases, so a meeting-notes tool and a pricing engine weigh the same. And it is a US sample, not a European mid-cap one, where data-residency rules and a thinner engineering bench change both sides of the trade.
The figure does not rebut this post’s rule. It is what the rule predicts. Most AI use cases in any company are commodity workflows, and commodity workflows should be bought. What the number cannot tell you is whether the few workflows your margin depends on were among the purchased ones, and that is the only question worth asking of it.
Buy: the case for it, stated plainly
Thoughtworks’ build vs buy analysis for generative AI frames the same trade-off from the engineering side, and reaches a similar conclusion about where the line sits.
Buying is the right answer more often than a company like ours has any commercial incentive to say. Three conditions, and all three have to hold at the same time.
The workflow is a commodity. Expense approvals, meeting capture, first-line ticket triage, contract redlining against a standard template. If your version of the process is not meaningfully different from your closest competitor’s version, there is nothing to own. Owning a commodity returns nothing and costs maintenance.
You hold no proprietary data advantage in it. The reason to build is data no vendor has: your pricing history, your claims decisions, your engineering change orders, the way your best account manager actually qualifies a deal. If the process runs on generic inputs, building it reproduces a product that already exists, more slowly.
You accept the vendor’s rate of change. You are buying their roadmap, their model choices, their pricing changes and their deprecations. That is a fair trade for a commodity workflow. It is a poor trade for a process your margin depends on.
If all three hold, buy the product and spend the attention somewhere it earns more. The expensive mistake is not buying. It is buying for a workflow that fails the first condition, then funding two years of configuration to force a generic product onto a process that was never generic, and calling the configuration work a transformation programme.
Build: the cost lands in year two
Internal build estimates are usually honest about the first version and silent about everything after it. The first version is not the expensive part.
The expensive part is the integration surface. Your data sits across systems that were never designed to be read by software, and a working agent has to reach all of them: the ERP, the CRM, the document store, the mainframe nobody wants to touch. That surface does not stabilise after launch, it moves every time one of those systems is upgraded.
Then the governance layer, which is a project of its own rather than a feature: the audit trail, decision traceability, role-based access, retention, and the evidence a regulator will actually accept. Enterprises building their first agent routinely discover this layer after the model works and before anything reaches production.
Then both of those, maintained indefinitely, by engineers who could have been working on the thing your company sells.
Build is the right answer when the workflow is genuinely yours and you have engineers you can assign permanently. The word doing the work there is permanently. If the four engineers who would build this are the same four keeping the ordering platform alive, you do not have a build option. You have a plan to acquire one, which is a different project with a different timeline.
Embed: the third column
The embedded model is no longer niche. Forward-deployed engineering has become the AI industry’s most contested hiring category, and the labs now sell it directly through their own deployment arms. Both are worth knowing before you assume you can staff it internally.
Embed means engineers work forward-deployed inside your stack, on your real data, under your governance, and what they build stays with you when they leave. Three distinctions make it a separate model rather than a synonym for outsourcing.
It differs from buying because you own the artifact, not a licence to use one. It differs from consulting because the deliverable is a system in production rather than a design for one, and the risk of getting it running sits on the delivery side. It differs from staff augmentation because the engineers arrive with the architecture already decided, having built the same layer before.
The model exists because of an asymmetry that buy and build both fail to address. The knowledge required to build this is inside your company, held by people who are too busy to be interviewed and who have never written any of it down. The engineering capacity to turn that knowledge into a system is not inside your company, and hiring it competes against the AI labs. Buying solves neither side. Building solves the second only if you already had it.
What embed should leave behind is specific, and worth naming in the contract: a context engine, the owned intelligence layer where your company’s process knowledge is written down in a form software can read, kept model-agnostic, with the agents and workflows running on top of it. Models change underneath every few months. The layer does not. That is the whole point of re-architecting rather than layering.
Embed is not “boost” or “partner”
Other frameworks offer a third option too, and mean something else by it. MIT Sloan’s framework calls it buy, boost or build: boosting means the vendor still provides, runs and maintains the model, and the company enhances it with its own data through fine-tuning or retrieval. Some guides call the third option “partner”, meaning co-developing on a vendor’s or an integrator’s platform.
Both are variations on buy. The vendor still holds the model choice, the runtime and the roadmap, and your proprietary data ends up shaping a system you rent. Embed differs on the one point that decides the long-term value: the system, the knowledge layer under it and the choice of model stay with you after the engineers leave. If a “partner” arrangement passes that test in the contract, it is embed under another name. If it does not, it is a licence with a services line attached.
Where embed fails
Embed is also wrong sometimes. Ask a forward-deployed team to build expense approval and you are paying senior engineering rates to rebuild something you could have licensed. It fails just as reliably when nobody internal is assigned to receive the system, because a system with no owner on your side becomes shelfware on a slower schedule.
The decision rule
One question, asked per workflow rather than once for the company:
Would this company be materially different if this workflow ran better than every competitor’s version of it?
- No. Buy it. Stop evaluating.
- Yes, and you have engineers you can assign permanently. Build it.
- Yes, and you do not. Embed, and require that what gets built stays yours.
A single company should reach all three answers in the same year. Buy the expense tool, build the pricing engine, embed for the cross-system processes that no product covers because no product can see across your departments. Treating this as one portfolio-level decision is how organisations end up with a licence for everything and ownership of nothing.
For agent products the delivery question does not change: who builds the integration and governance underneath the agent, and who owns them afterwards. When a vendor platform such as Salesforce’s Agentforce is the candidate, look at what the vendor’s own channel reports before assuming the licence is the deployment. In a TD Cowen survey of Salesforce partners, no partner said Agentforce had influenced any bookings, which we look at in more detail in what Salesforce partners report about Agentforce. Which workflows deserve re-architecting at all is a separate question, covered in re-architecting versus layering.
For a European buyer there is a second question, and it is not a compliance footnote. Where does inference happen, which subprocessors touch the data, and who is accountable when a supervisory authority asks. Under Article 25 of the EU AI Act the line between deployer and provider is not fixed by the purchase order: put your name on a high-risk system, or change substantially what it does, and the provider’s obligations can follow you home. A delivery model that leaves you dependent on a single vendor’s hosting and a single vendor’s model has made that answer for you.
One layer is shared by every workflow, the context layer agents read, so it is split first into container and contents, and the same rule then decides the contents.
Build vs buy the context layer: what do the platforms actually sell?
You can buy a place to keep a context layer, and a mid-cap that already runs Fabric, Databricks or Snowflake has paid for one. You cannot buy its contents.
Four data platforms sell a layer that tells agents what the business means, and SAP sells one on the ERP side. None runs outside the platform that sells it:
| Platform | What its layer holds | Runs outside that platform? |
|---|---|---|
| Palantir | The Ontology: objects, properties and links, plus the actions that change them | No, Foundry only, though virtual tables read OneLake, Databricks or Snowflake data in place |
| Microsoft | The Fabric IQ ontology: entity types, relationships, rules and metrics | No, it binds to Fabric data sources |
| Databricks | Unity Catalog semantics: metric views and domains | No, Unity Catalog only |
| Snowflake | Semantic views: business metrics, entities and their relationships | No, a schema-level object in Snowflake |
| SAP, the ERP-side option | Business Data Cloud and the SAP Knowledge Graph: entities, processes and relationships “across a customer’s SAP landscape” | No, SAP Business AI Platform; non-SAP data comes in through Business Data Cloud |
So the container usually comes with the platform you run; on SAP, check that you have Business Data Cloud, a product of its own. Palantir is the exception: for most mid-caps Foundry would be a second platform, priced in what Palantir AIP and Foundry cost. No layer holds the unwritten procedure, and Palantir’s action log records only decisions taken through its own action types, which leaves most decisions outside every layer, as the layer on top sets out.
Release status, export formats and pricing, platform by platform, are in Fabric IQ vs Palantir Foundry.
Context layer vs semantic layer vs ontology: what is the difference?
A semantic layer defines metrics, an ontology defines business objects and how they relate, and a context layer is everything an agent needs to act the way your company would: both, plus the knowledge outside any warehouse.
| Layer | Answers | Examples | Can you buy it? |
|---|---|---|---|
| Semantic layer | How is gross margin calculated? | Snowflake semantic views, Databricks metric views, Power BI semantic models, dbt’s MetricFlow | The tool, yes; the definitions are yours to write |
| Ontology | What is a customer, an order, a plant, and how do they relate? | Fabric IQ ontology, Palantir Ontology, SAP Knowledge Graph | The tool, yes, tied to one platform |
| Context layer | What would the company do in this case, and why did it decide that way last time? | The two above, plus procedures, exceptions, decision records at handoffs, permissions and provenance | The container, yes; the contents are your data and your decisions |
The test for any layer: does it hold the decisions taken outside the platform that sells it? A context engine, as we use the term, is the third row built as a system you own: in your stack, model-agnostic, kept current by a named person on your side, and reading whichever semantic layer or ontology you already have. That is why an AI-native enterprise treats the contents of its context layer as the one asset it cannot buy.
What does building your own context layer involve?
Building your own context layer is mostly reconciling definitions and writing down what people know, on whichever platform you already run, rather than engineering a new database.
Directions on Microsoft, an independent analyst firm, wrote in December 2025 that “Few enterprises have a well-architected enterprise data model” (Directions on Microsoft). Buying a platform layer gives sales and finance a place to settle what a customer is. It does not settle it.
The work, in the order we would sequence it:
- Reconcile the definitions agents read, one workflow at a time, starting with the one whose number you already measure, in the build sequence that holds.
- Write down the procedures and exceptions that live in inboxes and in people’s heads.
- Record the decision taken at each handoff an agent touches, as a decision trace your company owns.
- Keep permissions and provenance on every item, which the audit trail for agents depends on.
- Store it in formats readable without the vendor’s tool, and name the person who keeps it current. Open formats such as Apache Ossie carry metrics, dimensions and their relationships, but no decision record.
The cost of building is the bench: engineers assigned to this layer for good, because definitions drift and records go stale in the month nobody owns them. Without that bench, the usual constraint in Europe, the build column is closed and the embed column is open.
Should a mid-cap build or buy its context layer?
A mid-cap should buy the container, build the contents whoever’s platform holds them, and own the part no platform sells. By situation:
| Your situation | Context layer decision |
|---|---|
| You run Fabric, Databricks, Snowflake or Business Data Cloud, and agents read metrics and entities inside it | Buy, by using the platform’s layer, and fill it with your own definitions |
| Agents must write back to many operational systems in production now, and there is budget for a platform licence | Buy Palantir as a second platform beside the one you run, under the conditions where Palantir wins |
| The workflow crosses departments and vendors, and the knowledge is procedures, exceptions and past decisions | Build an owned context engine beside the platform layer, or embed if the bench is not permanent |
| No data platform chosen yet | Choose it on cost, skills and where your data sits, then apply the rows above |
Two rules sit under the table. Do not choose a data platform for its ontology: start from the one you run. In Fabric you can generate the ontology “directly from semantic models” (Microsoft Learn), so the first version can start from definitions finance already trusts. And test the exit before you sign, because each platform exports something different: see what leaves with you and what a migration involves.
A generic example, not a client case: a machinery maker on SAP and Databricks defines first-pass yield in a Unity Catalog metric view, and keeps why each part outside tolerance was accepted, and who signed the concession, in a context engine it owns, readable after any change of vendor.
An owned context engine also has a running cost: someone on your side keeps its records current after any external team has left.
What to settle before you sign
Whichever column you land in, these five belong in the contract rather than in a kickoff deck.
- The number that has to move, named before the build starts, with the baseline it is measured against and who owns the data that proves it. Our ROI benchmarks are a starting point for setting it, not a substitute for measuring your own.
- Who owns what at the end. The code, the prompts, the evaluation sets, the context engine, and the format each part of the context layer exports in, checked against an exit-clause checklist. Write it down. Silence here defaults to the vendor.
- Whether the system runs on a different model without re-architecting. If swapping the model means rebuilding the system, you did not buy portability.
- Who maintains it in month thirteen, by name and by team, on your side and theirs.
- Where the data goes, including at inference time, and which subprocessors are in the path.
Most day-rate contracts answer none of these, because a day-rate contract sells effort rather than a working system. That is the honest difference between a delivery model and an invoice, and it is worth more attention than the cost comparison that usually takes up the whole evaluation.
FAQ
Should we build or buy AI for the enterprise? Buy when the workflow is a commodity, you hold no data advantage in it, and you accept the vendor’s roadmap and model choices. Build when the workflow is specific to your company and you have engineers you can assign permanently, not borrow for a quarter. If the workflow is yours but the bench is not, neither answer fits, and the third option is to have engineers build it inside your stack and leave it with you.
What is the third option in build vs buy? Embed. Engineers work forward-deployed inside your systems, on your data, under your governance, and the system they build stays yours. It differs from buying because you own the artifact rather than a licence, and from consulting because the deliverable is a running system rather than a design for one. It also differs from “boost”, where the vendor still runs the model and you add your data on top: under embed, the system and the model choice are yours.
Should we build or buy AI agents? An agent is the last layer of the system, not the first decision. Buy an agent product for a commodity task such as first-line ticket triage. If the agent has to act on your own data across your own systems, the agent is the easy part, and the delivery question is who builds and owns the integration and governance underneath it: your own engineers if you can assign them permanently, an embedded team if you cannot.
Do most enterprises buy AI rather than build it? In the US, yes, by count. Menlo Ventures’ survey of 495 US enterprise AI decision-makers, run in November 2025, found 76% of AI use cases purchased rather than built internally, up from 53% purchased the year before. The figure counts use cases, not value or differentiation, so it is consistent with buying commodity workflows and owning the few that matter.
When is buying an AI product the correct decision? When three conditions hold at once: the workflow is not meaningfully different from your competitor’s version of it, you have no proprietary data that would make your version better, and you are willing to inherit the vendor’s roadmap, pricing and deprecations. If all three hold, buy. Owning a commodity workflow returns nothing.
Why do build estimates for enterprise AI come in low? Because they price the first version. The expensive parts arrive later: the integration surface against systems never designed to be read by software, the governance layer of audit trail, traceability, access control and retention, and the maintenance of both, indefinitely, by engineers who could be working on your own product.
Does build vs buy get decided once for the whole company? No. It is a per-workflow decision. A single company will correctly buy its expense approvals, build the pricing engine that differentiates it, and embed for the cross-system processes that no product covers. Deciding it once at the portfolio level is how organisations end up forcing a generic product onto a process that was never generic.
Should we build or buy a context layer for AI agents? Buy the container, build the contents. Fabric IQ, Databricks Unity Catalog semantics, Snowflake semantic views and Palantir’s Ontology each sell the container, SAP Business Data Cloud sells one on the ERP side, and if you run one of those platforms you already have one. None writes your definitions or procedures, or holds the decisions taken outside the systems connected to it. Build those with engineers you can assign permanently, or embed a team if you cannot, and keep them in your own stack.
What does building your own context layer involve? Mostly reconciling definitions and writing down what people know, on the platform you already run, rather than engineering a new database: the definitions agents read, one workflow at a time, the procedures and exceptions that live in inboxes, and the decision taken at each handoff an agent touches. Keep permissions and provenance on every item, store it in formats readable without the vendor’s tool, and name the person who keeps it current.
What is the difference between a context layer, a semantic layer and an ontology? A semantic layer defines metrics, such as how gross margin is calculated. An ontology defines business objects, their relationships and rules, and in Palantir’s case the actions that change them. A context layer is everything an agent needs to act the way your company would: both of those, plus the procedures, exceptions and decision records that sit outside any warehouse. Platforms sell the first two; the third is your company’s own knowledge.
Should we buy Palantir for its Ontology if we already run Fabric, Databricks or Snowflake? Only if agents must act as well as read, through governed write-back to many operational systems, in production now, and you can carry a platform licence, in which case Palantir runs beside your current platform. Otherwise the layer that comes with your current platform is enough to start, and Fabric IQ vs Palantir Foundry sets out the conditions.
Related reading
- What is a forward-deployed engineer, and why enterprise AI needs one
- Why enterprise AI pilots fail, and what the ones that ship do differently
- Salesforce partners and Agentforce: what the channel reports
- Fabric IQ vs Palantir Foundry: which ontology should a European mid-cap’s agents run on?
- What Palantir AIP and Foundry cost a European mid-cap
- Salesforce fixed the CRM. Workday fixed HR. Nobody owns the handoffs.
If the workflow is yours and the bench is not, that is the case we work on: engineers inside your stack, a system in production, and the intelligence left with you. See how we work with European mid-caps, or tell us which workflow it is.
Sources: Menlo Ventures, 2025: the state of generative AI in the enterprise; Thoughtworks, build vs buy for generative AI; MIT Sloan, buy, boost or build; TechCrunch on forward-deployed engineers; Salesforce Ben on the TD Cowen partner survey; Palantir documentation on the Ontology, virtual tables and the action log; Microsoft Learn, ontology in Fabric IQ; Databricks, Unity Catalog semantics; Snowflake documentation on semantic views; Apache Ossie, incubator status; dbt Labs, open-sourcing MetricFlow; SAP News on SAP Business Data Cloud, the Autonomous Enterprise announcement and the Sapphire 2026 keynote; Directions on Microsoft, CIO Talk: Microsoft Gets IQ; EU AI Act, Article 25.