# After Anthropic trains FDEs: How post‑investment PE teams can build a transferable, profitable AI delivery system

> Anthropic announced an investment of USD 100,000,000 to train ten thousand FDEs, while Gartner simultaneously forecasted that companies will abandon many vendor projects due to cost and maintenance capabilities. This report distinguishes talent supply, delivery success and operating returns, and provides post‑investment PE FDE organization design, full‑cost estimation, handover acceptance and a twelve‑week pilot plan.

- Canonical: https://fdepe.com/en/story?slug=fde-exit-anthropic-academy-gartner-operating-model-deep-report-2026
- Author: FDEPE Research Desk
- Published: 2026-10-05T02:48:13.479Z
- Updated: 2026-10-05T02:56:56.170Z
- Language: en
- Original: https://fdepe.com/story?slug=fde-exit-anthropic-academy-gartner-operating-model-deep-report-2026
- Revision: 1
- Translation: automatic; the Chinese original is authoritative.

[Inference | Core conclusion] The value of FDE investment should be measured by a company's ability to continuously improve processes, control costs and realize business benefits after the vendor's on‑site presence ends. Anthropic's new training program is expected to expand the supply of delivery talent; Gartner's forecast highlights weak points in knowledge transfer and cost structure. For post‑investment PE teams, procurement, business operations and engineering handover need to be designed as a single investment decision.

## Executive summary: Validate talent, delivery and profit separately

- [Fact] Anthropic on October 2, 2026 announced an investment of USD 100,000,000, aiming to train 10,000 Frontier Deployed Engineers by the end of 2027. Real projects, simulated deployments, a 12‑week on‑site stay and assessment form the core of the plan; the headcount and investment are commitments and targets, and a complete results cohort is not yet available. [S1]
- [Fact] Gartner on September 29 forecast that by 2028 70% of companies will abandon vendor‑built FDE intelligent AI, mainly due to cost and the inability to evolve independently. It is an analytical forecast and should not be stated as 70% of current FDE projects having failed. [S2]
- [Inference] Both messages point to an organizational issue: enterprises need to possess process knowledge, data access rights, testing assets and internal owners. On‑site engineers can accelerate the development of these capabilities, while prolonged on‑site presence may also mask the lack of knowledge transfer.
- [Recommendation] Decide continued investment based on three gates: business outcomes, total cost and independent operation and maintenance. The fund’s central team costs, business expert time, manual review and exit costs should all be included in the project accounting.

## 1. Fact base: The Academy trains delivery capability, which has not yet demonstrated financial return

[Fact] The first batch of institutions includes Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk and Commonwealth Bank of Australia. Participants are nominated by their institutions, enter training with the designated Claude project, and after on‑site hands‑on work carry out a real project at their original institution for 12 weeks; the first batch of final certifications are expected to be issued in early 2027 [S1].

[Inference] Its mechanism value lies in linking learning with on‑site business and incorporating engineers who already understand the institution’s processes into the program. It reduces the gap between classroom knowledge and production deployment, but cannot automatically solve the lack of business owners, data permissions, legacy system integration, and insufficient budget.

[Scope boundary] Frontier Deployed Engineer is the name used for this program; this article discusses the broader FDE as a Forward Deployed Engineering delivery mechanism. Participation by banks or consulting firms does not imply that their PE‑portfolio companies have already used the program. This report did not obtain project revenue, profit, or post‑graduation independent maintenance data from the participants.

[Research method] Using the two announcements this week as the event start point, supplemented by historical productivity studies and original engineering materials. The organization, contracts, and budget content in the following text are analytical suggestions and are not written as any system already implemented by participating enterprises.

## 2. Organizational value of FDE: translating business problems into runnable, maintainable systems

[Inference | work definition] In this report, FDE‑style delivery must satisfy four conditions simultaneously: enter the specific business site; co‑define outcomes with the business owner; be able to modify production systems and workflows; deliver assets that the enterprise can continuously maintain or reuse. Recruiting engineers, using large models, or increasing consulting headcount alone remains insufficient evidence.

- Business translation: clarify tasks, exceptions, decision authority, and financial outcomes, e.g., breaking down accounts‑receivable collection from “generate email” into identifying disputes, selecting actions, approving sending, and recording cash receipt.
- Engineering closed loop: connect business records, permissions, tool calls, execution logs, and quality tests to form a process usable by real users.
- Organizational collaboration: business experts provide rules and handle exceptions, engineers build the system, finance validates the benefits, and internal maintainers assume subsequent changes.
- Asset retention: deliver data dictionaries, interfaces, configurations, test suites, operation manuals, permissions and runtime logs, and train internal staff with at least redundant maintenance capability.

[Judgment] The scarcity of FDE may stem from the ability to integrate business judgment with engineering execution. At scale, repeatable parts should be productized while domain differences are retained in configurations and process rules. If each deployment relies on an individual engineer’s personal memory, expansion will continue to linearly consume senior talent.

## 3. Counter‑evidence: forecasts have limits, and real measurements also have biases.

[Fact] Gartner requires that contracts explicitly define intellectual property, joint project responsibility, knowledge transfer, and exit responsibilities, and predicts that by 2028 less than 20% of FDE projects will distill recurring client demands into core supplier product capabilities. The public press release does not provide sufficient sample and methodology to recalculate the forecast. [S2]

[Judgment] This information can be used to design due‑diligence questions, but 70% cannot be directly used to calculate a project's failure probability. The supplier's productization ratio is also not equivalent to client benefit: a one‑off client‑specific process may generate high returns yet not become a generic product.

[Fact | Historical Evidence] METR disclosed on July 10, 2025 that 16 experienced developers completed a random experiment of 246 real open‑source tasks, allowing the use of AI tools at the time, which increased task duration by about 19%. The study subjects and model period are clearly defined and cannot be extrapolated to all current enterprises, FDEs, or non‑coding processes. [S3]

[Fact | Subsequent Revision] On February 24, 2026, METR explained that follow‑up experiments suffered participant and task selection bias and difficulty measuring parallel Agent work hours, concluding that the original results are insufficient to reliably estimate the speedup at the time. Therefore, the 19% figure from 2025 cannot be continued as a fixed conclusion for AI productivity in 2026. [S4]

[Inference] The most useful insight for post‑investment PE projects is to record real tasks, quality, total labor time, and full cost simultaneously. When engineers perceive faster speed, more generated code, or stronger model benchmarks, these must be compared against net benefits in the production environment.

## 4. Three delivery models: selected according to problem stage and internal capability

- Supplier‑led: suitable for tasks with new product capabilities that require deep model or platform knowledge and cannot be performed by the internal team at present. The main risks are high rates and dependence. Contracts must specify on‑site termination points, knowledge handover, and interface responsibilities.
- Portfolio‑company‑internal lead: suitable for processes that already have an engineering team, stable business rules, and clear data and product ownership. The main risk is that the team lacks cross‑domain experience or is squeezed by daily demands, requiring protected project labor hours.
- Fund central team co‑delivery with the company: suitable when multiple portfolio companies face similar issues and have legally shared components. The main risks are that headquarters costs are overlooked, portfolio companies are unwilling to cooperate, and cross‑industry reuse is overestimated.

[Recommendation] The selection mechanism can use four judgments: how deep supplier proprietary knowledge the project requires, how quickly business rules change, whether internal maintenance capability is sufficient, and whether there is evidence of cross‑company reuse. When the problem is clear and can be solved by a standard product, compare the total cost of mature products versus conventional integration first; FDE investment should have a clear additional benefit assumption.

[Organizational Boundary] The fund’s central platform can standardize evaluation criteria, connectors, and security components; data involving each portfolio company’s customers, employees, and trade secrets requires separate authorization. Holding control does not mean any data can flow across companies unconditionally.

## 5. Full‑cost model: there must be a bridge between released labor hours and cash realization.

[Inference | Model Definition] Monthly realizable labor value = qualified task count × net minutes saved per task ÷ 60 × full‑load hourly cost × realization rate. Net savings must deduct additional review, rework and exception handling; the realization rate represents the share of saved time converted into verifiable economic benefits such as fewer hires, less outsourcing, and reduced overtime.

[Inference | Cash Model] Monthly net operating profit = realizable labor value + attributable new business contribution – new operating costs. Operating costs include model, computing, tools, internal maintenance and ongoing evaluation. Initial investment also covers on‑site delivery, integration, training and the first migration.

[Illustrative scenario, not actual data for any company] Assume 20,000 qualified tasks per month, full‑load labor cost of CNY 50 per hour, and an initial investment of CNY 600,000 RMB; the three scenarios do not count additional sales to avoid double‑counting the same released capacity.

- Obstructed: net saving of 5 minutes per task, realization rate of 30%, monthly realizable value of CNY 25,000; monthly operating cost of CNY 90,000, net profit negative CNY 65,000, unable to break even under the current model.
- Baseline: net saving of 10 minutes, realization rate of 60%, monthly realizable value of CNY 100,000; operating cost of CNY 70,000, net profit of CNY 30,000; simple payback period calculated from stable earnings of 20 months.
- Favorable: net saving of 15 minutes, realization rate of 70%, monthly realizable value of CNY 175,000; operating cost of CNY 60,000, net profit of CNY 115,000; simple payback period about 5.2 months.

[Inference] The payback period is counted from the point of stable operation, excluding ramp‑up, time value and taxes, and is not an IRR. The three scenarios are stress tests and do not constitute probabilistic forecasts; formal project approval should input real task volumes, costs and ramp‑up of benefits month by month. In the baseline scenario, if 20,000 tasks cannot be converted into sufficient realizable work volume, the payback period will extend.

[Suggestion | PE Perspective] The fund’s central team salaries and platform fees should be reasonably allocated to the project. Portfolio companies may report their own cash benefits separately, but the fund level must also deduct central costs. If saved time is used for new revenue, only the actual new contribution or cost realization should be counted.

## VI. Handover Acceptance: Demonstrate capability transfer by independently completing a change.

[Suggestion] Independent operation means the enterprise can use legally obtained model services and tools, monitor, diagnose issues, modify business rules, and perform releases and rollbacks on its own. It may continue to procure APIs and clearly bounded support services; the acceptance focus is that daily operations no longer rely on external on‑site personnel.

- Asset Acceptance: Code, configuration, data dictionary, interfaces, evaluation set, deployment steps and licensing boundaries are complete, and internal staff can rebuild the system from a controlled environment; assets have clearly assigned owners and version records.
- Knowledge Acceptance: Internal maintainers independently handle a predefined business change and a fault while the vendor observes without intervening; retain operation video, change logs, and results.
- Quality acceptance: verify correctness, exception escalation, key action permissions and business outcomes simultaneously in the fixed holdout set and live trial run. High‑risk actions require business approval and a rollback path.
- Economic acceptance: include manual review and maintenance in the cost of each successful task; after handover, compare service cost with SLA on the same task basis.
- Continuous acceptance: re‑test maintenance capability, utilization and cost during the recommended follow‑up visits at 30 days and 90 days. The timing is a recommendation of this report and should be aligned with the business cycle.

[Fact and basis] Anthropic's original evaluation guide emphasizes distinguishing execution trajectories from final environment outcomes, and combines multiple scoring and production monitoring. For business agents, an output of “completed” cannot replace a verifiable record in the business system. [S5]

[Recommendation | Metric definition] Independent change success rate = number of internal changes that pass the quality gate without vendor execution ÷ total changes tested; unit successful task cost = (operating cost during the same period + manual maintenance cost) ÷ number of tasks that pass business acceptance. The cost of failed tasks remains in the numerator and must not be removed because of failure.

[Scope] Generic percentages cannot replace business risk standards. Financial payments, client contracts and regulatory processes should adopt stricter thresholds; low‑risk internal documents may retain more manual review. Companies should freeze acceptance criteria before seeing pilot results to reduce post‑hoc selection.

## 7. Contracts and incentives: turn the handover from an ancillary document into a payment condition.

[Recommendation | Commercial terms framework] The statement of work should specify the task scope, production responsibilities, business acceptance, knowledge transfer, existing intellectual property versus new product boundaries, data export, licensing, termination and handover support. The following are commercial design questions; actual terms must be implemented by both parties in line with the transaction and applicable rules.

- Phased payments: link the budget to verifiable business baselines, production usage and handover outcomes; settlement solely by person‑days or number of deployments makes it difficult to create shared incentives for the enterprise’s independent operation.
- Cost transparency: clearly define fees for model calls, tools, manual support and out‑of‑scope change work, and pre‑agree on billing methods for usage growth, to avoid low pilot fees masking scale‑up costs.
- Changes and incidents: define the responsibilities of the business owner, engineering lead and vendor; agree on incident response, data recovery, version rollback and platform upgrade compatibility boundaries.
- Exit delivery: list deliverable assets, training, hands‑on internal staff sessions, transition support period and pricing. For the vendor’s proprietary technology, clearly specify the authorization scope needed for continued operation.
- Personnel incentives: reward net business contribution, knowledge transfer and reusable components; avoid rewarding only usage volume, automation ratio or technical demos.

[Inference] Suppliers whose revenue relies entirely on on‑site labor may lack economic incentives to reduce client dependence. Buyers can use milestone payments and separate pricing for maintenance services and new projects, allowing suppliers to benefit from successful handovers and subsequent replication.

## VIII. Twelve‑week pilot: fund, business, engineering and finance share responsibility

- Weeks 1–2: select scenarios with sufficient task volume, clear process boundaries and financial relevance; record quality, labor hours, costs and anomalies. The business lead signs off on issues and baselines, and the CFO confirms the revenue measurement.
- Weeks 3–4: develop a minimal workflow, set up data permissions, execution logs, retain evaluation sets and a manual takeover mechanism. Internal maintainers co‑build from day one.
- Weeks 5–8: onboard real users in batches, compare task quality against unit cost; record complex exceptions and failure costs. If quality or cost thresholds are not met, do not expand the task scope.
- Weeks 9–10: conduct internal independent maintenance and change drills; sample‑check whether benefits translate into reduced cash expenditures or additional contributions; differentiate the impacts of models, process standardization and personnel changes.
- Weeks 11–12: business, finance and engineering jointly decide to continue, adjust or stop; components that have passed move into a second‑group replication trial.

[Scope] The twelve‑week period is a pilot recommendation aligned with the training program cycle and does not guarantee that all projects will reach production maturity within twelve weeks. Regulated processes, deep ERP integration and long‑term client retention require a longer observation period; staged decisions should be retained rather than forcibly claiming on‑schedule success.

[Responsibility allocation] The business lead is accountable for adoption, exceptions and operating results; the FDE lead is accountable for systems and delivery; internal maintainers are accountable for subsequent changes; the CFO is accountable for net profit; the fund operating partner is accountable for cross‑company replication and central costs. Any vacant role should be filled before scaling deployment.

## IX. Portfolio‑level governance: use a set of metrics to compare different companies

- Business outcomes: select collection cycle, client retention, unit gross margin or delivery time per scenario. Each project maintains a fixed baseline and time window; incomparable scenarios must not be merged into an average efficiency‑gain percentage.
- Real adoption: active users with actual business tasks / target users who meet criteria, while observing the proportion of tasks processed. Training completions and login counts are recorded separately.
- Quality and exceptions: business qualification rate, severe errors, manual escalations and rework rate. When the automation ratio rises, check whether the manual burden shifts to the exception queue.
- Economics: cumulative initial investment, monthly net profit, unit cost per successful task and payback period; actual values, budget and scenarios are listed separately.
- Transfer and replication: internal independent change success rate, external support hours, second deployment cost and reuse component maintenance cost.

[Recommendation | Portfolio Expansion Rules] Prioritize replicating processes that have passed the three validations of business, quality and independent operation. For scenarios where central components can be reused but data cannot be mixed, use permission isolation and independent assessment; if industry differences make reuse costs approach rebuilding, treat it as a new project and re‑initiate.

## X. Unverified Issues and Future Research Design

- Anthropic Academy: track nominations, enrollments, graduations, production launches and a complete queue with 90 days of independent maintenance; split by institution and project complexity to avoid publishing only successful cases.
- Gartner forecast: obtain research methodology, corporate denominator, abandonment definition and observation period, distinguishing stopping a single project, switching vendors and discontinuing use of the Agent.
- PE portfolio companies: interview operating partners, CFOs, front‑line business staff and departing engineers, cross‑checking contracts, production logs, labor hours and financial accounts; the current report has not conducted interviews.
- Failed projects: include samples with no new revenue, cost overruns, staff turnover and procurement termination, and record the reasons; shutting down low‑value pilots can sometimes be a prudent capital allocation and should not be uniformly labeled as technical failure.
- Replication evidence: compare the incremental delivery costs of the first, second and third companies, test whether the shared platform yields economies of scale, and also record central maintenance expenses.

[Decision Conclusion] FDE can become a key organizational capability for post‑investment transformation in PE, but the investment logic should be based on complete costs, genuine business acceptance and sustainable maintenance. Talent programs provide supply, vendor deliveries provide a starting point; whether the company retains an operational system, knowledge and cash returns determines the long‑term value of this capability.

## Research Explanation and Direct Sources

Data verification date: October 5, 2026. US dollars are used for disclosed amounts, RMB only for clearly labeled self‑constructed examples. Historical research is used solely for background and methodology analysis. The report adopts a conclusion‑first, evidence‑layered and business‑model structure, constituting an independent FDEPE study, with no affiliation or sponsorship from McKinsey.

[[S1] Anthropic, 2026-10-02, Claude Frontier Academy announcement; investment commitments, objectives and training mechanisms. ↗](https://www.anthropic.com/news/claude-frontier-academy)

[[S2] Gartner, 2026-09-29, vendor FDE forecasts and delivery recommendations; public materials do not provide a complete forecasting methodology. ↗](https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028)

[[S3] METR, 2025-07-10, early 2025 AI random experiment on experienced open‑source developer productivity; historically limited sample. ↗](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

[[S4] METR, 2026-02-24, subsequent productivity experiment design adjustments; discussion of selection bias and timing difficulties. ↗](https://metr.org/blog/2026-02-24-uplift-update/)

[[S5] Anthropic, 2026-01-09, Demystifying evals for AI agents; evaluation framework, real results and production validation. ↗](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents)

## Source links

- https://www.anthropic.com/news/claude-frontier-academy
- https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028
- https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- https://metr.org/blog/2026-02-24-uplift-update/
- https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents