Microsoft Frontier Company: Whose Intelligence Does It Protect?
The short answer
Microsoft Frontier Company is an operating business inside Microsoft, launched on 2 July 2026 with a $2.5 billion investment, that embeds 6,000 industry and engineering experts at customers to build and run AI systems on Microsoft’s platform. Microsoft defines its promise to protect your intelligence two ways: your data and IP will not train models that commoditise what differentiates you, and customers “shouldn’t be locked into a single model any more than they should be locked into a single technology vendor.”
The first is a policy. The second holds at the model layer, but one layer down Microsoft’s documentation shows partial exits: rules, metrics and audit records leave only as Microsoft-specific files or records, and Work IQ’s memory not at all. For a mid-cap on Microsoft 365 that can be an acceptable trade, if it is a decision rather than a default.
Key takeaways
- Frontier Company is Microsoft, not a spin-off. It is not clear whether the $2.5 billion is new money.
- Microsoft promises no training on your data and no vendor lock-in. The second holds for models, not for every layer below them.
- Exportable is not portable. Data, code and an ontology’s vocabulary leave in open formats. The layers that take longest to build do not.
- The Frontier Transformation Engineer badge is Microsoft’s partner FDE credential. It certifies delivery on Microsoft’s stack only.
- The most quoted Copilot ROI is commissioned. Forrester’s 116% models a $6.25 billion company. UK trials found similar or smaller time savings, and none shows the time becoming output.
What is Microsoft Frontier Company?
Microsoft Frontier Company is a Microsoft business unit that staffs AI deployment inside customers, announced on 2 July 2026 by Judson Althoff, CEO of Microsoft’s commercial business. Its 6,000 “industry and engineering experts” are engineers and salespeople, per Directions on Microsoft, and its president, Rodrigo Kede Lima, has been a Microsoft sales leader in the Americas and Asia for six years.
Directions, an independent analyst firm, reports that it is not a spin-off or a separate company, though Microsoft says it has its own leadership and financial accountability. Citing unnamed sources, it says Frontier Company looks like a make-over of Microsoft’s consulting organisation and Cloud Solution Architects program. Whether the budget is new money or recommitted consulting funds, the launch post does not say.
The launch post names LSEG, Land O’Lakes, Unilever and Novo Nordisk as customers, and Accenture, Capgemini, EY, KPMG, PwC “and others” as FDE partners. It arrived two days after AWS committed $1 billion to its own unit. It is one of six labs, hyperscalers and platforms that moved into embedded delivery in 2026, alongside Google Cloud, Anthropic, OpenAI, Meta and AWS; the DeployCo and Ode comparison sets them side by side. It is unrelated to the Microsoft Frontier program of early access to preview AI features.
What does Microsoft Frontier Company actually do?
It embeds Microsoft people at a customer to “co-design, co-innovate, deploy and continuously improve AI systems at scale based on measurable business outcomes” on Microsoft’s platform. The launch post says this “goes beyond what has been labeled as Forward Deployed Engineering (FDE)”. “Continuously improve” means an engagement with no designed end: a commercial model, not a flaw. The clearest reference is on the landing page: LSEG and Microsoft’s FDE team co-engineered Workspace AI Search across research, banking, trading and risk workflows.
How much does Microsoft Frontier Company cost?
Microsoft has not published a price. Directions on Microsoft reads the model as “the Azure migration playbook applied to AI”, and its director of advisory services, Lane Shelton, puts it plainly: “Free deployment is the customer acquisition cost and the consumption meters are the payback.”
If that is right, the engineers are the cheap part and the meter is the bill. Work IQ in Copilot Studio is billed in Copilot Credits, Microsoft Agent Factory offers “a single consumption model” across Microsoft’s agent products, and a Fabric IQ ontology runs on Fabric capacity. Ask for a consumption estimate per workflow before the first sprint; week-two design choices set the meter for years.
What is the Microsoft Frontier Transformation Engineer badge?
The Frontier Transformation Engineer badge is not a certification or a single exam. It is an individual credential for engineers at Microsoft partners. The Microsoft-sponsored learning journey, 11 courses and about 86 hours, requires three Microsoft certifications (GH-300, AI-103 or AI-102, and AB-100), five project-ready assessments on Microsoft IQ, Work IQ, agentic AI, Fabric IQ and Agent 365 with an 80% pass mark, and a Hypervelocity Engineering assessment.
Microsoft’s partner blog says the path is built on how Microsoft readies its own Forward Deployed Engineering teams. In practice, then, it is Microsoft’s partner forward-deployed credential. Five badge holders, three DP-600 Fabric certifications, four prerequisite specializations and a third-party audit earn a partner the Frontier Partner specialization, open since 30 September 2026, which validates delivery “across the Microsoft tech stack”.
Accenture now runs a separate FDE program or group per platform, among them Microsoft, ServiceNow, SAP, Google Cloud and Workday, and the Google Cloud group’s 1,000 planned forward-deployed engineers will be credentialed by Google Cloud. That is the argument of who really does forward-deployed engineering now: the pod belongs to whichever vendor relationship funds it.
What are Work IQ, Foundry IQ and Fabric IQ?
They are three of the four parts of Microsoft IQ, which the Fabric IQ overview calls the enterprise intelligence layer of the Microsoft stack. The fourth, Web IQ, provides context from the web; the Foundry IQ page still lists three.
Work IQ is “a contextual intelligence layer for Microsoft 365”, built on Data, Memory and Inference layers. Foundry IQ is “a managed knowledge layer for enterprise data”, retrieved through Azure AI Search. Fabric IQ is “a semantic intelligence layer for Microsoft Fabric”, whose ontology, in preview, binds entity types, relationships, rules and metrics to live data. The partner badge assesses Microsoft IQ, Work IQ and Fabric IQ by name.
What “protects your intelligence” covers, and what it does not
Microsoft defines protection two ways. The launch post promises that a customer’s data, IP and competitive advantage are not “used to train models in ways that commoditize what differentiates them in their industry”. Its next paragraph adds: “We protect that intelligence with a model-diverse, open, heterogeneous AI platform. Customers shouldn’t be locked into a single model any more than they should be locked into a single technology vendor.”
The first promise covers leakage: your know-how improving a shared model your competitor also rents. The second covers dependency: the knowledge that makes you hard to copy accumulating in a layer you cannot take with you. At the model layer the claim holds. Althoff’s own June yardstick, a platform “model-diverse, open and heterogeneous at every layer of the stack”, is the right one; the table below applies it one layer down. Lane Shelton of Directions on Microsoft gives the opposing reading: an embedded Frontier engineer is “Microsoft’s roadmap being installed as your architecture”, which customers should understand “before their next renewal, not after”.
What do you own if you leave Microsoft’s AI stack?
You keep your model choice, data, code and the ontology’s vocabulary, in formats another platform can read. Rules, metrics, retrieval configuration and audit records export only in Microsoft formats or as static records, and Work IQ’s memory has no documented export. The table checks each layer against Microsoft’s documentation in October 2026.
| Layer | Microsoft component | Leaves with you? | What the documentation says |
|---|---|---|---|
| Models | Microsoft Foundry | Yes, model choice is open | Over 11,000 models, including OpenAI, Anthropic, Meta, Google, xAI and Hugging Face. |
| Data | OneLake | Yes | Tables in Delta Parquet or Iceberg, “two open standards that any tool can read”. |
| Orchestration code | Microsoft Agent Framework | Yes, the code | Open source under the MIT licence. Code bound to a managed Azure service needs its own answer. |
| Business vocabulary | Fabric IQ ontology | Yes, as RDF | Entity types, properties and relationships export as TTL or RDF, in preview, covering a subset of RDF and OWL. |
| Rules, metrics, bindings | Fabric IQ ontology | Exportable, not portable | Export as TMDL through Git integration and getDefinition, with entity and rule blocks in “ontology-only syntax”. Only explicit metrics round-trip. |
| Retrieval | Foundry IQ | The configuration, not the engine | The knowledge base definition reads out as JSON. Azure AI Search runs the retrieval. |
| How your people work | Work IQ Memory | Not documented | Memory “builds persistent understanding of how people and teams work”. The Work IQ documentation describes no export. |
| Governance evidence | Purview | As records, not a live trail | Copilot interactions export through eDiscovery as PST or message files, audit records as CSV, up to 50,000 rows on Audit Standard. Exports expire after 14 days; retention, holds and the live record stay in Microsoft 365. |
| People | Frontier Company or Frontier Partner staff | Only what your own staff learned | Microsoft employees, or partners credentialed on Microsoft’s stack. |
Exportable is not portable. A TMDL file in Git is a backup you can diff, not a system another platform can run, and a PST of Copilot interactions answers a lawyer’s request without being an audit trail anyone can query or extend. Data and code are the cheapest layers to rebuild. Rules, metrics, the record of how people work and live evidence take longest, and leave as files or not at all. That gap is what owning your intelligence is about, and why audit trails for AI agents must be designed for the day you switch.
A European buyer has a lever US guides leave out. Under the EU Data Act, applicable since 12 September 2025, platform and software providers must offer open interfaces and export data in a commonly used, machine-readable format, and switching charges, egress included, end on 12 January 2027. Whether a rule file or a memory store counts as exportable data is a question for counsel, best asked before the design is fixed.
Get six answers in writing first:
- What runs on the TMDL export outside Fabric, and which of our metrics fail to round-trip?
- What does Work IQ’s memory hold about our teams, can we export it, and can we delete it?
- Can the retrieval pipeline be rebuilt on another search engine from the knowledge base definition?
- Where does the agent audit trail live after we change platform, and who can produce it for a regulator?
- Who on our payroll can change the system without a Microsoft engineer, and by what date?
- Which of these layers do you treat as exportable data under the Data Act’s switching rules?
A Nucleo engagement builds a context engine, the owned, model-agnostic intelligence layer the company runs on, inside the client’s own estate, so it stays when we leave. If that estate is Microsoft, it sits there: owning the intelligence means the rules, metrics and memory sit in formats and accounts you control, not leaving Microsoft. The same six questions apply to us.
What Microsoft does well
For a company already on Microsoft, Frontier Company starts with advantages no independent partner has. The contract is signed, identity and security run on Entra, Defender, Purview and Intune, and Agent 365 governs agents “whether they are built on Microsoft AI platforms, delivered by ecosystem partners or introduced through other technology stacks”. Microsoft reported 30 million paid Copilot seats in July 2026.
In Europe, Microsoft completed the EU Data Boundary in February 2025, saying it had invested more than $20 billion in AI and cloud infrastructure across Europe in the preceding 16 months. Its June 2025 sovereign offer adds a Sovereign Public Cloud for all European customers, where Data Guardian has European-resident staff approve and monitor remote access by Microsoft engineers, plus customer-held encryption keys and public-sector partner clouds in France and Germany.
If you run on Microsoft, the workflow lives in Microsoft 365 rather than across other vendors’ systems, and it is not where you differentiate, then Frontier Company or a Frontier Partner is a defensible choice.
Residency is not jurisdiction: the European question
The EU Data Boundary decides where data sits, not which government can compel Microsoft to hand it over. Before a French Senate inquiry in June 2025, Anton Carniaux, Microsoft France’s director of public and legal affairs, said it had “contractually committed to our clients … to resist these requests when they are unfounded”. Asked whether it would have to transmit data under a well-framed US request, he answered “Absolutely, by respecting this process”, adding that no European company or public body had been affected since Microsoft began publishing transparency reports. Microsoft’s own report for July to December 2025 has since recorded content data provided to US law enforcement for 3 non-US enterprise customers whose data was stored outside the US, one of them in the EU/EFTA (Microsoft). Chapter VII of the Data Act protects non-personal data stored in the EU against unlawful foreign government access requests. The risk is rare, and it is not zero.
Buy from Frontier Company directly and the delivery contract is with a Microsoft entity. Buy through a Frontier Partner, the likelier route for a mid-cap, and you hold two contracts: delivery with the integrator, platform with Microsoft. Ask both which entity signs, where inference and retrieval run, who holds the keys, and which of the two owes you the exit.
What is the real ROI of Microsoft 365 Copilot?
The most quoted Microsoft 365 Copilot ROI, 116% over three years, comes from a Forrester Total Economic Impact study Microsoft commissioned. The independent evidence, from UK government trials, mostly measures time saved, and supports Forrester’s time assumption better than its output assumption.
The March 2025 study models $36.8 million of benefits against $17.1 million of costs, with a 10-month payback, for a composite company with $6.25 billion in revenue and 25,000 employees. Microsoft supplied the customer names for the interviews but did not take part in them. Each user saves 9 hours a month, valued at $38 an hour, with 50% of it recaptured. Licences cost $30 a month, rolled out to 40% of staff by year three, and the go-to-market benefit adds $14.8 million on revenue up to 2.6% higher. The 353% figure resellers quote is a separate Forrester study of companies with up to 300 employees.
| UK trial | Run by | Sample | Method | Finding |
|---|---|---|---|---|
| GDS cross-government trial | Government Digital Service | 20,000 staff, three months | Self-reported | 26 minutes a day, 82% would not go back |
| Department for Business and Trade | DBT | 1,000 licences, three months | Diary study, interviews | 72% satisfied; no firm evidence of improved productivity, not a key aim |
| HMRC phase 3 | HM Revenue and Customs | 3,000 staff, randomly allocated | Survey answered by 45%, cut by about 20% | About 60 minutes a week, 2 to 3% of the working week |
On time saved, the commissioned model sits inside the trials’ range. GDS’s 26 minutes a day is about 9 hours a month, Forrester’s figure. HMRC’s bias-adjusted hour a week, about 4.3 hours a month, is half that, and close to the 4.5 hours Forrester values after recapture. HMRC’s analysts also estimated around £50 million a year of net productivity benefit from up to 50,000 licences, a capacity estimate rather than cash. The weak links are the step where saved time becomes output, which DBT found no firm evidence of, and the go-to-market benefit, which no trial tests.
Size it for your company. At $30 a seat, Copilot for a 2,000-person mid-cap costs $720,000 a year before any consumption, and SAMexpert, a UK licensing adviser, puts a 5,000-person company at $1.8 million. Do not earn that back in minutes per seat. Measure one workflow with a number on it, such as days per tender, because that is what Forrester’s 50% recapture stands in for, and re-architect it, as the 2026 ROI benchmarks and rearchitecting versus layering explain.
Who is Microsoft Frontier Company built for?
Large enterprises: every reference Microsoft has named is one. Partners are meant to carry it “across all markets and segments globally”, so for a company with EUR 250M to 2B in revenue the realistic choice is a Frontier Partner, often the integrator already on site, or an independent partner.
| Frontier Company (direct) | Frontier Partner | Independent partner | |
|---|---|---|---|
| Delivery contract | A Microsoft entity | The integrator, with the platform contract at Microsoft | The partner |
| Who staffs it | Microsoft’s industry and engineering experts | Partner staff; the specialization needs five badge holders | The partner’s own engineers |
| Credential certifies | Not published | Delivery “across the Microsoft tech stack” | No platform vendor credential |
| How it is paid | Not published; Directions reads it as a consumption payback | Integrator fees; the partner earns co-sell and Agent prepurchase plan credits | Whatever the contract names |
| Named customers | LSEG, Land O’Lakes, Unilever, Novo Nordisk, Siemens, Core42 | Not published per partner | Varies by firm |
| What leaves with you | The layer table above | The layer table, plus the integrator’s contract terms | Whatever the contract names, layer by layer |
Both routes build good systems. They differ in whose platform the knowledge is shaped around, and whether the contract says what leaves with the engineers. Ask the six questions before your next Microsoft renewal, while the answers can still change the price. If the system will not hold what makes you hard to copy, buy the convenience.
FAQ
What is Microsoft Frontier Company? An operating business Microsoft announced on 2 July 2026 to embed industry and engineering experts at customers and co-design, deploy and continuously improve AI systems tied to business outcomes. Microsoft committed $2.5 billion and 6,000 people to it, and Rodrigo Kede Lima leads it as president. It is not a spin-off or a separate company, it builds on Microsoft’s own platform, and Microsoft says it goes beyond what is usually called forward-deployed engineering.
Is Microsoft Frontier Company the same as the Microsoft Frontier program? No. Frontier Company is the deployment business that embeds Microsoft experts at customers. The Frontier program gives Microsoft 365 users early access to preview AI features before general availability. Frontier Firm, a label that runs through Microsoft’s marketing and partner training, is a third thing. They share a word, not a function.
Is the Microsoft Frontier Transformation Engineer a badge or a certification? A badge, which requires certifications. It is an individual credential for engineers at Microsoft partners: three Microsoft certifications (GH-300, AI-103 or AI-102, and AB-100), five project-ready assessments on Microsoft IQ, Work IQ, agentic AI, Fabric IQ and Agent 365 with an 80% pass mark, and a Hypervelocity Engineering assessment. Microsoft says the path is built on how it readies its own Forward Deployed Engineering teams, which makes it Microsoft’s partner forward-deployed credential.
What are Work IQ, Foundry IQ and Fabric IQ? Three of the four parts of Microsoft IQ, which Microsoft calls the enterprise intelligence layer of its stack; Web IQ is the fourth. Work IQ is a contextual layer over Microsoft 365, built in data, memory and inference layers. Foundry IQ is a managed knowledge layer whose retrieval runs on Azure AI Search. Fabric IQ is a semantic layer over Fabric whose ontology holds business entities, relationships, rules and metrics.
What do I keep if I leave Microsoft’s AI stack? Four layers leave in a form another platform can read: model choice, OneLake data in Delta Parquet or Iceberg, code on the MIT-licensed Microsoft Agent Framework, and a Fabric IQ ontology’s vocabulary as TTL or RDF. The ontology’s rules, metrics and bindings export only as TMDL with Fabric-only extensions, Foundry IQ’s configuration reads out as JSON but runs on Azure AI Search, and Copilot interactions export as PST or CSV records. Those are exportable, not portable. Microsoft’s Work IQ documentation describes no export for its memory.
How much does Microsoft Frontier Company cost? Microsoft has not published a price. Directions on Microsoft, an independent analyst firm, reads the business model as the Azure migration playbook applied to AI: free deployment is the customer acquisition cost and consumption of Microsoft AI services is the payback. If that is right, budget for the Copilot Credits, Agent Factory consumption and Fabric capacity the design will use, not only for the engineers.
What is the real ROI of Microsoft 365 Copilot? The most quoted figure, a 116% three-year ROI, comes from a Forrester study Microsoft commissioned, modelled on a composite company with $6.25 billion in revenue and 25,000 staff. Its time assumption sits within the range UK government trials found: GDS staff reported saving 26 minutes a day, and HMRC, which allocated 3,000 licences at random, found about 60 minutes a week of self-reported savings after adjusting for bias. No trial shows the time turning into output, and the Department for Business and Trade found no firm evidence that it did, though that was not a key aim.
Is Microsoft Frontier Company right for a European mid-cap? It can be, if you already run on Microsoft 365, Entra and Purview and the workflow is not the one you differentiate on. Its named customers are large enterprises such as LSEG, Unilever, Novo Nordisk and Siemens, and a mid-cap is more likely to buy through a Frontier Partner, which means two contracts: delivery with the integrator, platform with Microsoft. Ask both which entity signs, where inference and retrieval run, and which of the two owes you the exit.
Related reading
- Who really does forward-deployed engineering now
- OpenAI’s DeployCo, Anthropic’s Ode, and the case for an independent forward-deployed partner
- Forward-deployed engineer vs consultant vs system integrator: a decision rule for buyers
- Own your intelligence: what it actually takes
- Audit trails for AI agents: what regulated enterprises actually need
- The ROI of enterprise AI agents: 2026 benchmarks
Nucleo is an AI-native transformation company for European mid-caps. Our forward-deployed engineers build a context engine, the model-agnostic, portable intelligence layer your company owns, governed to EU rules, and put agents into production on top of it. See how an engagement works, or write to us.
Sources: Microsoft Frontier Company launch post; Microsoft Frontier Company; Directions on Microsoft; Microsoft Frontier program; TechCrunch on the launch and AWS; Achieving success with AI; Frontier Transformation Engineer learning journey; Engineering the Frontier partner practice; Frontier Partner specialization, now live and coming soon; Accenture on its Microsoft FDE practice and ServiceNow FDE program; CIO Dive on Accenture’s Gemini Enterprise group; Microsoft Learn on Fabric IQ, Foundry IQ, Work IQ, ontology import and export, the ontology item definition, knowledge base retrieval, OneLake, Copilot data in eDiscovery, eDiscovery export and audit search; Microsoft Agent Framework; Microsoft Foundry; European Commission on the Data Act; Accelerating Frontier Transformation with Microsoft partners; No Jitter on Copilot seats; EU Data Boundary; Microsoft sovereign solutions for Europe; The Register on data sovereignty; Forrester TEI of Microsoft 365 Copilot; Microsoft on the SMB study; GDS cross-government findings; DBT blog and evaluation report; The Register on the DBT trial; HMRC phase 3 evaluation; SAMexpert on Copilot ROI.