Why 86 Percent of Enterprise AI Agents Stay Stuck in Pilot Mode
A joint SAP-Infosys framework treats AI agents as digital coworkers rather than scripts, exposing why most enterprise pilots stall and what a defensible, scalable deployment actually requires.
Independent UpShaqo analysis built from fresh, attributed sources. We explain the impact instead of repeating the announcement.
Read for leverage: focus on the workflow change, the customer problem, and the next action—not only the product announcement.
Most enterprise AI programs don't fail because the model is weak. They fail because nobody built the plumbing underneath it. That's the uncomfortable diagnosis offered by executives from SAP and Infosys in a recent MIT Technology Review Business Lab conversation, which doubles as a teardown of what it actually takes to move AI agents from isolated demos into the operating core of a company. The numbers they cite are stark: an Infosys and HFS Research survey of more than 500 global enterprises found only 14% have successfully scaled AI, while 86% remain stuck in early exploration, proof-of-concept, or fragmented pilots, according to the same report.
That gap between pilot and platform is the real story. Here's how the system is supposed to work, where it breaks, and why the companies that get it right may be building an advantage competitors can't simply buy back.
The User Job: From Tool to Coworker
The job enterprise AI is being asked to do has quietly changed. Early deployments were productivity add-ons — meeting summaries, travel assistants, email drafts — where a human stayed in the loop and 80% accuracy was tolerable. Satinder Pal Singh, associate VP for SAP Business AI Platform strategy at Infosys, reframes the ambition: agents as "digital coworkers" that execute discrete tasks inside a business process, not just assistants that speed up a task a human still owns, per the report.
That reframing matters because it changes the acceptance bar. A coworker closing the books or managing a supply chain can't be right 80% of the time. Thorsten Leiduck, SVP and global head of SAP's Business AI Platform Ecosystem, is blunt about it: in mission-critical workflows, 80% accuracy is "a very terrible result," according to the transcript. The job has moved from assisting a human to standing in for one — and that raises the stakes on every input feeding the system.
The Inputs Nobody Wants to Fund
Ask any enterprise AI team what they need and they'll say more compute, better models, faster iteration. Leiduck's answer is less exciting and more foundational: a connected, semantically rich data layer that reflects how the business actually runs, built before agents are deployed, not alongside them, per the source.
This is the unglamorous part of the stack, and it's where most programs quietly die. An Anthropic survey cited in the conversation found 42% of organizations struggle simply to get access to data of sufficient quality to scale AI capability across the business — not a model problem, a plumbing problem, according to the report. A separate BCG finding puts 74% of organizations in pilot mode with no plan to scale, per the same source.
Analysis: these figures point to the same root cause from different angles — companies are deploying agents on top of fragmented, inconsistent data estates and discovering, only after the pilot works in isolation, that it can't generalize. An agent grounded in one team's spreadsheet logic isn't grounded in the company's reality; it's grounded in a corner of it.
Orchestration: A Three-Lens Filter Before You Build
The most exportable idea in the conversation is Infosys's triage framework for deciding which use cases deserve investment at all, built around three questions applied simultaneously, per the report:
- Impact — does deploying an agent move a number that matters: revenue, cost, working capital, risk?
- Feasibility — given the current data foundation and talent, can this use case actually be executed now, or does it need sequencing behind other fixes?
- Scalability — once proven, is there an architecture that lets the win replicate across teams, geographies, or business units?
Singh's framing of a failed pilot is memorable: "a pilot that can't scale is just an experiment," he says, per the source. The concrete payoff of running use cases through this filter shows up in the numbers the two companies cite from joint engagements: one effort cut SAP-specification documentation work from 86 person-hours to 6, a 92% reduction, and a separate S/4HANA transformation completed more than 480 delivery objects in five months using AI-assisted tooling — a pace the report says conventional delivery wouldn't sustain, according to the transcript. A third deployment auto-classifies over 120,000 SAP incidents monthly with no analyst involvement, per the same report.
Mini-framework, for operators: treat the AI portfolio the way you'd treat an investment portfolio — a handful of quick wins to fund the program, a tranche of medium-term bets, and one or two bold long-term positions, as Singh recommends in the source. This keeps stakeholders engaged in year one instead of waiting for a payoff in year three.
Where the System Breaks
The conversation is candid about failure modes, and they cluster around two things: governance gaps and organizational silos.
On governance, Singh's line is the sharpest in the piece: "the absence of governance doesn't speed up AI adoption. It creates a risk backlog," per the report — a backlog that eventually forces a disruptive stop-and-audit cycle rather than the smooth scaling teams hoped for. Leiduck connects this directly to agents "going rogue" in the news, arguing that clear access controls, monitoring, and policy enforcement aren't optional overhead but preconditions for trustworthy scale, according to the source.
On organization, Singh names a specific trap: a company stands up a dedicated AI team or center of excellence, walls it off from the business, and the team builds "technically excellent pilots the business doesn't recognize itself in" — solutions searching for problems rather than the reverse, per the report.
Analysis: both failure modes have the same cure — AI has to be an operating-model change owned jointly by IT and the business, not a technology initiative IT runs for the business.
The Adoption Barrier Is Human, Not Technical
Singh's estimate is specific and worth sitting with: the mindset shift required for teams to treat agents as digital coworkers, rather than automations, accounts for roughly 70% of the implementation challenge — more than the platform itself, according to the transcript. That's a claim worth taking seriously for any leader budgeting an AI rollout as a technology line item rather than a change-management program with technology attached.
The Defensible Opportunity
The strategic argument underneath all of this is about compounding advantage. Only 8% of companies qualify as "AI native" by Infosys's own maturity index, yet that group captures four times the revenue contribution from AI compared with laggards, per the report. Singh's point is that this gap isn't easily closed by purchasing the same platform later — eighteen months of accumulated agent-behavior data, a calibrated knowledge graph, and a workforce that has genuinely internalized how to work with agents are structural advantages that compound, according to the source.
Recommendation for operators: before approving another AI pilot, audit the data foundation it depends on and name the specific business metric it's meant to move. If you can't answer both, you're funding an experiment, not building the platform that compounds.