Agentic AI: Orchestrating Intelligent Operations

Enterprises are racing to deploy agentic AI, but a lot of that momentum is stalling out. Gartner projects that a large share of agentic AI initiatives will be scrapped by 2027 — not mainly because the technology fails, but because companies underestimate the organizational lift required to run it well. That's the central theme of a Harvard Business Review Analytic Services white paper, sponsored by Deloitte, which argues the real bottleneck isn't building agents — it's operating them.

Nobody has done this before

Part of the problem is that agentic AI is moving too fast for any single company to build up real experience with it. Alex Bakker of ISG Research points out there's not enough proven market evidence yet, forcing every enterprise to pilot, adapt, and evaluate on its own over an extended period. Large organizations often have well over a thousand applications they'd like to apply agentic thinking to, but internal teams are rarely staffed to deploy — let alone maintain — AI at that scale.

Compounding that is a knowledge gap many companies are just discovering. Siemens' Ronny Hendrych warns that organizations increasingly lack people who understand their own legacy systems well enough to build on them safely. This is where next-generation managed service providers (MSPs) come in — bringing cross-client experience that lets them move faster than any internal team starting from zero.

Four kinds of "debt" standing in the way

The paper identifies four types of organizational debt that need clearing before agentic AI can run reliably at scale:

• Process debt – workflows designed for humans that don't translate to AI without a redesign
• Data debt – fragmented or inconsistent data that undermines reliable decisions
• Technical debt – legacy systems that don't integrate cleanly with AI orchestration layers
• Cultural resistance – friction that surfaces as human roles shift

Everest Group's Anil Vijayan suggests a deliberate sequencing: start automating low-risk, high-volume tasks, then move to more complex multi-system processes, and only later attempt full function-by-function transformation.

Beyond automating tasks — redesigning workflows

A recurring theme is that companies get much more value from agentic AI when they stop automating individual tasks and instead redesign entire workflows around outcomes. Using employee onboarding as an example, the paper contrasts a narrow approach (an agent just reading an offer letter) with a redesigned one, where multiple agents hand off to each other to trigger IT setup, payroll, and training enrollment automatically — organized around the moment that matters to the employee, not the org chart. Getting this right, Vijayan notes, usually requires humans in the loop throughout a transition period, with MSPs often managing that parallel run.

AI's economics don't behave like normal IT spend

One of the paper's more technical points: agentic AI is priced and consumed differently from traditional software. Instead of licenses or fixed infrastructure costs, spend is driven by tokens — the units of data models process — and that consumption can scale unpredictably, breaking traditional cost-of-ownership models.

This unpredictability is also reshaping contracts. Bakker describes the core tension: providers and clients don't know in advance how automatable a given process really is, so both sides carry risk that standard pricing doesn't account for. ISG's proposed fix is "autonomy-level pricing" — tiered pricing set at the start of a contract that rewards providers for pursuing automation without needing a new negotiation every time.

Governance: knowing where humans still need to be in the loop

A September 2025 Gartner survey found only a small fraction of IT leaders felt confident they had the right governance structures in place for agentic AI. The paper stresses that in regulated or safety-critical environments — like rail operations — some decisions still require full explainability, which limits which AI approaches are even appropriate. As Hendrych puts it, in operational technology, mistakes can't simply be undone.

The human side matters just as much. At Mankind Pharma, field managers initially worried that agent-driven recommendations would replace their judgment rather than support it. Chief digital officer Kunal Basal says the system was deliberately built as decision support, not a replacement, with recommendations designed to stay transparent and explainable — and over time, skepticism gave way to genuine buy-in as managers saw tangible time savings.

Vijayan describes a typical maturity curve for oversight: starting with humans reviewing every AI decision, moving to spot-checks as trust builds, and eventually intervening only when something breaks.

How companies are actually choosing MSP partners

The paper argues that evaluating MSPs the old way — cost, labor arbitrage, SLAs — misses what actually matters now. Vijayan points to two better signals: whether a provider has genuinely differentiated intellectual property, and whether they can show real production deployments, not just pilots.

At Mankind Pharma, Basal says the evaluation looked well past technical skills, prioritizing partners who showed strong systems thinking and a real point of view on governance. Deep industry knowledge matters too — Hendrych notes that some MSPs are excellent in one sector and out of their depth in another, so it's worth asking whether you're new territory for them.

Once a partner is chosen, the relationship model shifts as well. Basal describes his company's MSP partnerships as built around joint teams and shared responsibility for outcomes, not just delivery milestones.

The takeaway

The paper's closing argument is that MSPs willing to invest their own capital into agentic AI — rather than waiting for guaranteed ROI, as Hendrych observes was once standard — are a sign of how seriously the industry now takes this shift. The organizations that succeed, per Basal, will be the ones that treat agentic AI as a long-term operating model change, not a project with an end date, paired with partners able to evolve alongside them.