Named production deployments show AI agents can produce measurable outcomes. At the same time, research from MIT and Gartner shows that many pilots still fail to deliver measurable return on investment (ROI) or get canceled before they scale. Read the production evidence and the failure research together, and one variable outweighs both model quality and orchestration framework: governed access to business data. Governed data shapes which use cases hold up under load, what returns teams can actually measure, and which pilots stall before they reach production. That argument runs through the evidence below.
TL;DR Named production evidence includes Gradient Labs reporting 80–90% resolution rates and Indicium running internal agents across Moodle, GitHub, Bitbucket, internal wikis, and Snowflake. Published ROI achievement rates vary because Deloitte measures reported gains, McKinsey measures earnings before interest and taxes (EBIT) impact, and MIT NANDA measures rapid revenue acceleration. Failure research names poor data quality, ambiguous objectives, cumbersome workflows, inadequate integration, and weak risk controls as common blockers. Pre-materialized context lets an agent reason across systems when teams unify data upstream into one queryable layer. Runtime tool-calling fits single-system lookups. Get started to give production agents governed, current access to your business data before you scale.
Which AI Agent Use Cases Work in Production Today? Most AI agent examples that survive production fall into a few categories. Customer support dominates the public evidence, while IT operations dominates internal deployment. Cross-system reasoning agents, which answer questions across multiple business systems simultaneously, remain the rarest category.
Adoption data is less mature than outcome data. McKinsey's 2025 global survey found that 23% of respondents were scaling an AI agent system, 39% were experimenting, and no more than 10% were scaling AI agents in any individual business function. Three categories account for nearly all of the named production evidence.
Customer-Facing Support Agents Support is often the first production category because the task is bounded. A customer asks a question, the agent retrieves policy and account state, then resolves or escalates. Public support-agent deployments, including Gradient Labs , follow this shape, and it accounts for most of the existing named production evidence. Support also produces a clean denominator, conversations resolved without a human, which makes results easy to defend to a CFO.
Internal Operations and IT Agents Inside the firewall, the distribution looks different. ZDNet's survey data show that 61% of production agents are in IT operations. Performance bots account for 66%, security monitoring agents for 63%, and development assistants for 62%. These agents receive less public coverage, but they account for much of the internal production usage described in the survey.
Cross-System Data Reasoning Agents A difficult category involves agents that answer cross-system questions, such as which customers opened tickets this week and are also up for renewal. Indicium, a global data consultancy , is one of the named companies running internal agents across its learning platform, code repositories, wikis, and data warehouse.
Across these categories, the deployments that publish results share a common trait. The data each agent reads is governed and kept current, which is exactly what the measured outcomes reflect.
What Do Measured Production Outcomes Look Like? Reported outcomes cluster around resolution rate, planning time, and cross-system access to production. These figures come from linked public case studies and reporting.
Company Agent Type Measured Outcome GrupoBoticário Supply chain planning 40% fewer stockouts; planning time cut from 49 hours to 3 Gradient Labs Financial services support 80–90% resolution rates Indicium Internal multi-system agent Production agents querying Moodle, GitHub, Bitbucket, internal wikis, and Snowflake
The evidence is concentrated in customer-facing support, where the task is bounded, and the metric is obvious. Indicium is the only named internal multi-system deployment in the set. Whether outcomes like these count as ROI depends on which bar you measure against.
How Should Businesses Measure AI Agent ROI? Published ROI achievement rates vary because each one measures a different bar.
Source Finding What It Measures Deloitte 2025 ~70% of AI agent investors reported gains Any gain (survey) McKinsey March 2025 39% report EBIT impact Enterprise-level EBIT impact MIT NANDA July 2025 5% achieve rapid revenue acceleration Revenue acceleration (52 interviews + 300+ implementation reviews)
The nearly 70% and 5% findings can both be true because teams often see some return but rarely hit rapid revenue acceleration. Most investments sit between "any return" and "expected return," so the business case should model the difference between reported gains, EBIT impact, and revenue acceleration before funding scale.
How much an agent can act shapes the return as much as how accurately it answers. Read-only agents can only describe work. Agents that can write can actually do work. A read-only agent produces summaries an analyst could have written, while an agent that updates the CRM, creates the ticket, and posts the message replaces the workflow, and that is where measurable ROI comes from. Consolidating sources so a single query resolves across systems compounds that return, with benchmarks showing up to 90% cost savings on multi-source queries.
Projections usually give less attention to cost. Observability frameworks treat token usage as a cost KPI alongside latency, success rate, handling time, tool selection efficiency, and user feedback. If a team prices tokens at chatbot rates, the model will look profitable on paper and run at a loss in production. Cost-control tactics are part of scaling AI agents ; the ROI model has to include cost instrumentation from day one.
Gartner's recommendation for measuring is to track outcomes rather than activity. Useful metrics include cost per resolved claim, percentage increase in finalized policies, and overrides per 1,000 decisions. Those metrics require agent observability good enough to trace what the agent actually did. Most pilots never get far enough to report against bars like these.
Why Do Most AI Agent Pilots Never Reach Production? Three independent research efforts converge on the same picture. The MIT NANDA report drew on 52 organizational interviews and 300 public deployments, and found 95% of generative AI pilots fail to deliver measurable ROI. Gartner projects organizations will cancel more than 40% of agent projects by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
AI agent governance shows up in that evidence as the missing risk-control layer that gets funded projects canceled. Teams that treat enterprise AI governance as a launch requirement rather than a retrofit remove one of the top blockers before it bites.
MIT NANDA attributes pilot failure to "poor data quality, ambiguous objectives, cumbersome workflows, and inadequate integration into daily operations." MAST's poor-specification and system-design category includes failures like loss of conversation history, which often trace back to context management and stale data.
For teams studying why AI projects fail , source access, authentication, freshness, and governance must be in place before autonomy increases. Data access now shapes deployment decisions more than framework choice does, putting agent data access at the center of production planning.
Which Deployment Patterns Fit Which Production Needs? The Model Context Protocol standardized how agents call tools, shifting the differentiator to what lies behind the endpoint. Two patterns dominate, and they behave very differently under production load.
Runtime Tool-Calling Deployments In the default pattern, the agent calls one system's API or MCP server at a time, at query time. Each call adds round-trip latency; each response returns a raw payload that the agent must parse within a context window , and nothing reconciles the three versions of the same account across customer relationship management (CRM), support, and billing. A single-system endpoint supports lookup within a single system; cross-system reasoning requires reconciled context across systems. Runtime calls still fit in narrow cases when the workflow relies on a single system of record, has low query volume, or focuses more on taking action than on analysis.
Pre-Materialized Context Deployments The alternative unifies data upstream. Sources sync into a single indexed, queryable context layer that the agent searches, rather than orchestrating API calls mid-task. The agent asks one question across many systems, gets structured results, and uses its context window to reason rather than parse payloads. Because pre-materialized data is only as current as its last sync, teams should pair indexed context with live API calls for fields that must be up to date.
Build Versus Buy for the Data Layer In the MIT study , 67% of externally partnered deployments succeed, compared with 33% of internal builds. The build-versus-buy decision comes down to ownership. Teams that build the data layer themselves must own OAuth flows for every source, schema drift, rate limits, and freshness checks, work that sits outside the agent product you set out to ship. Buy when the source coverage you need already exists; build when your data lives in proprietary systems no vendor covers. For most teams, buying means adopting an existing context layer and reaching it through tooling such as an agent SDK , rather than building and maintaining that layer in-house.
Whichever pattern a team picks, the constraint is the same. The agent needs governed, current data it can query across systems, and that is where a purpose-built platform does the heavier lifting.
How Does Airbyte's agentic platform Support Production AI Agents? Airbyte's agentic platform gives agents a governed Context Store for business data. Connectors sync business systems into that pre-indexed store, and Managed Auth handles secure credential storage, so an agent authenticates once per source rather than carrying its own secrets. One query can search across 50+ connected systems, separating simple lookup from cross-system reasoning and reducing the number of live calls an agent has to make.
Teams reach that store through whichever interface fits their stack. Agent MCP exposes the store to assistants like Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, and Windsurf through a single endpoint. The Agent SDK gives engineers programmatic access to build retrieval and actions directly into an application. The Agent CLI supports scripting and automation from the command line, which suits pipeline jobs and continuous integration checks. An API provides the same access over HTTP for applications and services that call the store directly. Search for pre-indexed context and Direct for live calls, let a team decide per field whether indexed data is current enough.
How Do You Move an AI Agent From Pilot to Production? The deployments that reached production solved data access before they scaled autonomy. MIT NANDA, Gartner, and the MAST taxonomy all point to the same operational blockers, which cluster around data quality, integration, and governance rather than model choice. The move from pilot to production is mostly a data-readiness decision.
Airbyte's agentic platform grounds production agents in governed, current business data. It keeps a queryable context layer fresh through incremental syncs, enforces source permissions through managed authentication, and lets teams decide per field whether indexed data or a live call answers a question. That combination targets the exact blockers the research names.
Get a demo to see how it works against your own systems and data.
Frequently Asked Questions How Long Until an AI Agent Deployment Pays Back? Payback depends on how bounded the task is, how much human handling time the agent can remove, and how expensive its data, inference, monitoring, and oversight loops become in production. Narrow, high-volume work usually pays back first because the denominator is obvious and the workflow is easier to measure.
Which Industries Are Slowest to Put Agents in Production? Records-heavy, heavily regulated functions move last because teams must log, review, and defend every retrieval and action before launch. The harder threshold for most companies is scaling a pilot, since a working prototype rarely survives contact with messy production data spread across many systems.
Are Custom-Built or Vendor Agents More Likely to Succeed? Vendor-partnered deployments have shown higher success rates in recent studies, largely because the partner owns the integration and upgrade path that in-house teams tend to underestimate. Building your own makes sense mainly when your data lives in proprietary systems no vendor supports, and you can staff the ongoing maintenance. For most teams, the deciding factor is who maintains the data layer over time.
Do Regulations Apply to Production AI Agents? Yes. When an agent operates in a regulated or high-risk workflow, logging, data governance, and human oversight requirements may apply. Meeting those requirements means keeping trace-level records of what the agent did, which tools it called, and where a human intervened.
