# Why 88% of AI Agent Projects Never Reach Production – and What Anthropic Just Did About It > What Anthropic's Managed Agents platform changes for consultancies, how it compares to five alternatives, what it costs, and the three caveats to weigh before committing. **Type:** Breakdown · **Read time:** 9 min · **For:** Managing Partner, COO · **Published:** 22 Apr 2026 **Video companion:** https://www.youtube.com/watch?v=hljjJ_XXvlI [All resources](https://aigenticlab.com/resources) · [View as HTML](https://aigenticlab.com/resources/anthropic-managed-agents-consultancies) --- ### The production gap ## Demos work; deployments don't Research consistently finds that 88% of AI proof-of-concept projects never reach wide-scale deployment. The problem is rarely the demo. Demos work. The problem is everything required to run AI reliably once it leaves the controlled environment and meets a real client brief, real credentials, real data, and a team that cannot afford downtime. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. That is not because the models are inadequate — the underlying AI is more capable than ever. The problem is the scaffolding around the model: sandboxing, state management, security, credential handling, and monitoring. For a 15 to 200-person consultancy without a dedicated DevOps team, building all of that in-house is a losing bet. MIT research found that buying AI tools from specialist vendors succeeds roughly twice as often as building internally. The question is not whether to use AI agents — it is who you buy the infrastructure from. That is exactly where Anthropic's Managed Agents enters the picture. ### Under the hood ## What Managed Agents actually is Managed Agents is Anthropic's hosted platform that runs long-horizon AI agents on your behalf via a cloud API. You do not manage the infrastructure. Anthropic does. To understand why they built it this way, you need a problem Anthropic's own engineers discovered. When you build an agent, you wrap the model in scaffolding — an orchestration harness that makes assumptions about what the model can and cannot do. Those assumptions go stale every time the underlying model upgrades. Anthropic found that Claude Sonnet 4.5 exhibited what they called "context anxiety", rushing to complete tasks as it sensed its context limit approaching. Running the identical harness on Claude 4.5, that behaviour disappeared entirely. The scaffolding was the bottleneck, not the model. Their fix borrows a principle from operating system design: virtualise the components and make them swappable. The architecture splits into three layers. - **Sessions** are an append-only log of everything the agent did, stored outside Claude's context window. The full task history is always recoverable; an agent that needs to resume reads the log and continues from where it stopped. - **Harnesses** are the orchestration loop that calls Claude, routes tool calls, and manages context. If a harness crashes, a new one boots up, fetches the event log, and resumes from the last recorded event. - **Sandboxes** are managed execution environments where the agent runs code and edits files. They are interchangeable — if one fails, the agent retries on a fresh container. The security architecture matters separately for any firm handling confidential client data. In the old design, code Claude generated ran in the same container as your credentials, so a prompt injection attack only needed to convince Claude to read its own environment variables. In the new architecture, API keys and OAuth tokens sit in a secure vault entirely separate from the sandbox where Claude-generated code runs, and Claude calls tools through a proxy that fetches credentials independently. That is a structural fix, not a patch. ### Where the return is ## Three use cases worth measuring The most useful question is not "what can this do" but "what can we measure the return on." Here are three starting points, ranked by how clearly you can quantify the impact. - **Proposal and bid generation.** Complex proposals consume 40 to 80 hours of non-billable senior time for many consultancies. An agent receives the RFP, searches your knowledge base via MCP integration, pulls relevant case studies and methodologies, and assembles a structured first draft — so partners spend time on strategic judgment rather than document assembly. Against the KWA framework, this sits at the intersection of Knowledge (your centralised case-study library) and Workflow (the repeatable bid-response process); the agent layer is the final step and only works reliably once the first two are in order. - **Research and analysis.** McKinsey's internal AI tool handles 500,000 prompts per month across 72% of their staff, reclaiming roughly 30% of research time. For your firm, that means client meeting preparation, competitive intelligence, and market analysis — tasks that currently account for 40 to 60% of professional time — handled in a secure, audited environment. - **Time capture and utilisation.** Consultants under-report billable hours, sometimes by 15 to 25%. An agent monitoring calendars and documented activities can suggest accurate time entries continuously. For a 50-person firm, every 10% improvement in utilisation can add up to £20,000 per consultant per year. The arithmetic compounds quickly. ### The honest comparison ## Five alternatives and what each one trades off Anthropic is not the only option. Here is an honest comparison of the five most relevant alternatives for a professional services firm. - **OpenAI Responses API** has the most mature developer ecosystem and an open-source agent SDK with background mode for long-running tasks. The difference is architectural: OpenAI gives you building blocks; Anthropic gives you the assembled harness. - **Microsoft Copilot Studio** is the path of least resistance if your firm runs on Microsoft 365 — a low-code visual builder with deep integration across Teams, SharePoint, and Outlook. The constraint is a two-minute flow timeout, which rules out the sustained autonomous work Managed Agents is built for. - **AWS Bedrock Agent Core** is the most flexible — fully model-agnostic and framework-agnostic, with per-second billing on actual resource consumption. The challenge is cost modelling: you are billing across orchestration, inference, retrieval, observability, and networking simultaneously, and complexity accumulates fast. - **Salesforce Agentforce** is strong if you are already a Salesforce shop, with pre-built sales and service agents out of the box — but it requires the full Salesforce stack. Implementation for a small firm typically starts at £50,000 in year one, scaling well beyond that for a mid-size firm once you add licensing, training, and ongoing consulting. - **Open-source frameworks** — LangGraph, Kaito, and similar — give you maximum control and zero vendor lock-in. The trade-off is months of engineering time, and for a consultancy without dedicated infrastructure engineers that flexibility rarely justifies the cost. ### Run the numbers ## What it actually costs Managed Agents charges on two dimensions: 8 cents per active session hour for the infrastructure, plus standard Claude API token costs for model usage. Anthropic's own benchmarks put a one-hour coding session at moderate token load using a top-tier model at roughly 70 cents total. Heavier workloads — web searches, longer sessions, complex multi-step tasks — will cost more. For a 50-person consultancy running five agents at four active hours per day, the session fees alone come to under £30 per month. The token costs are where real spend accumulates: depending on model choice and task complexity, total monthly cost for that kind of deployment typically lands between £500 and £2,000. Compare that against the upside. One recoverable billable hour per consultant per week across 40 fee-earning staff, at £150 per hour, produces £24,000 per month in recovered revenue. The ROI calculation is not the hard part. What comes next is. ### Read the fine print ## Three caveats you need to take seriously - **Vendor lock-in is real.** Managed Agents runs exclusively on Anthropic's infrastructure — not through AWS Bedrock, not through Google Vertex. If Anthropic falls behind a competitor or faces business disruption, your migration costs will be significant. You are making a single-provider bet; make sure you are comfortable with that before committing. - **Data sovereignty is unresolved for many firms.** If your consultancy handles client-confidential information — strategy documents, financial data, M&A details — sending that to a third-party AI platform creates legal and contractual risk. Only 51% of UK IT decision-makers are confident their AI-generated data complies with GDPR, and the EU AI Act's high-risk system requirements take full enforcement this August. You need explicit client consent frameworks or anonymisation pipelines before connecting client data to any of these platforms. - **This is a public beta.** Multi-agent orchestration and advanced memory tooling remain in limited preview. Internal testing showed up to 10 percentage points of task success improvement, but production reliability at scale is still being proven. > Human oversight is not optional — it is the governance layer. Only 21% of companies have a mature governance model for autonomous agents. You need written policies for which decisions agents can make autonomously and where a human must step in before you deploy. The alternative is the kind of outcome Deloitte encountered when asked to partially refund a government contract because AI-generated content contained hallucinations. That is a governance failure, not a technology failure. ### Decide from where you sit ## How to think about your next step The right platform depends on where your firm sits. If you are Microsoft 365-heavy and want structured, low-code workflows, Copilot Studio is the pragmatic first step. If you need long-horizon autonomous work — research, proposal drafting, multi-step analysis — Managed Agents is worth testing. If you need model-agnostic flexibility and have engineering resource, Bedrock Agent Core or open-source gives you that. For most consultancies, the recommended approach follows the 5 Steps for AI Leadership framework at the Accelerate stage: pick one high-value use case, run it for 90 days, measure time saved rigorously against cost, and use that data to make the business case for broader adoption. BCG research shows that firms treating AI as a strategic capability investment achieve 1.7 times the revenue growth of those treating it as a technology experiment. Do not experiment. Invest.