When a procurement team decides to scale agentic AI, the first instinct is usually to count. How many agents are live? How many categories do they touch? How many more can we add next quarter?
The trouble is that it may be the wrong number to watch.
The Hackett Group’s new research, The Autonomous Sourcing Index, benchmarks how organizations deploy, orchestrate and scale autonomous sourcing. One of its clearest findings is also one of its least comfortable: running more agents does not, on its own, produce better sourcing outcomes.
What the research looked at
The Index draws on responses from senior sourcing and procurement leaders across Europe (45% of respondents), North America (32%) and Asia Pacific (23%). Of those respondents, 78% report $1 billion or more in annual revenue, 47% manage $1 billion or more in annual spend, and 45% are executive-level sourcing leaders.
The Index scores each organization on a 10-point scale across deployment maturity, sourcing and orchestration capability, readiness and success factors. The median score is 3.6 out of 10, which places the typical organization in the “developing” stage: repeatable use, with foundations still forming.
The agent-count myth
Start with the baseline. Among respondents, 40% run three to four AI agents in production. Another 16% run one or two, 12% run five to nine and 8% run ten or more. A further 22% have no agents in production, while 2% do not know.
That sounds like momentum, and in one sense it is. But the same research shows that enterprise reach is narrow. Only 3% describe their deployment as enterprise-wide, and 45% say agentic AI covers just 1% to 10% of their sourcing categories. In other words, many organizations have several agents, but those agents touch a small slice of the spend.
Now compare outcomes. The Index asked respondents whether they had seen moderate-or-better improvement (defined as a 5% or greater gain) across six outcomes. It then split the results between organizations running one to four agents and those running five or more.
The gap is smaller than most leaders would expect. On five of the six outcomes, the two groups sit within ten percentage points of each other. On productivity, 71% of the smaller group report gains versus 67% of the larger. On category and supplier intelligence it is 65% versus 67%. On spend savings, both groups come in at 50%. The largest difference in this comparison is spend under management, at 59% versus 42%, and it is the lower number that belongs to the group with more agents.
The conclusion in the report is direct: agent count alone does not predict value.
What does predict value
If the number of agents isn’t the lever, what is? The Index points to the way those agents are connected.
Respondents were also grouped by the architecture they use for agentic AI in sourcing, and the distribution is telling. Standalone point tools are used by 23%. Connected point tools account for 17%. Workflow-enabled AI is the most common architecture at 30%. Cross-functional orchestration sits at 10%, and end-to-end orchestration at just 3%.
When outcomes are broken out by architecture, the pattern is hard to miss. Among organizations running point tools, 42% report moderate-or-better cycle time gains. Among those with workflow-enabled AI, it rises to 72%. For the orchestrated group, it is 100%. A similar pattern appears in productivity (50%, 72%, 100%) and spend savings (29%, 50%, 100%). These are descriptive comparisons, not evidence that architecture alone caused the gains.
A necessary caution
A fair reading of this data requires some humility. Only 3% report end-to-end orchestration, and the combined orchestrated cohort is also small. A 100% reported result should therefore be treated as directional and exploratory, not a guarantee or proof of causation.
That caveat does not weaken the practical message. Even if you set the orchestrated group aside, workflow-enabled organizations outperform those running disconnected point tools on every outcome measured. The direction is consistent. Connected beats isolated.
Why isolated agents stall
Why would that be? Procurement executives interviewed for the research offer a helpful explanation. One put it this way: the larger opportunity is connecting agents across intake, sourcing, contracts and analytics, and automating isolated tasks risks creating another generation of silos.
Think about what a sourcing agent actually needs to be useful. It needs to understand the request that started the event, which is an intake question. It needs enterprise data about suppliers, categories and spend. It needs policy, so its recommendations stay inside the rules. And it needs somewhere to send its output, so a recommendation becomes an award, a contract and a measurable result.
An agent that writes excellent RFx content but cannot see the original business requirement, or cannot hand its evaluation to a contract workflow, still leaves a human to bridge each gap.
This is also where platform choices matter. Zycus, which partnered with The Hackett Group on the Index, positions its approach around this exact issue. Its Merlin Agentic AI Platform combines AI-guided intake, an agentic execution layer and an integrated source-to-pay core, so that agents such as Merlin Intake and the Autonomous Negotiation Agent (ANA) operate on shared data and policy rather than as separate tools. Whichever technology you choose, the architectural question is the same: do your agents share context, or does each one start from scratch? You can learn more about Zycus and its Intake-to-Outcomes approach on its website.
Three questions to ask before you add another agent
- Where does this agent get its inputs? If the answer is “a person copies data in,” the integration gap will cap its value.
- Where do its outputs go? An evaluation that must be manually re-keyed into another system is a handoff, not a workflow.
- How will we know it worked? Agent count tells you what you deployed. Cycle time, productivity and spend outcomes tell you what changed.
The Index also finds that only 10% of organizations measure agentic AI performance consistently across deployments, which makes the third question the one most teams skip.
What to do next
If you are planning the next wave of agentic AI investment, this research suggests shifting the conversation from “how many?” to “how connected?”
The full report goes considerably further than this post. It benchmarks your organization against the 3.6 median and the 4.7 top-quartile score, breaks maturity into four stages with a playbook for each, and details where capability is strongest and weakest across nine sourcing activities.
Download The Autonomous Sourcing Index and see where your organization lands.
Source: The Hackett Group® in partnership with Zycus, The Autonomous Sourcing Index, 2026.
For an additional perspective on procurement platform capabilities and market positioning, explore Zycus in The Forrester Wave™: Supplier Value Management Platforms, Q4 2026. The Forrester evaluation is separate from the Hackett sourcing research.
For a separate perspective on the enterprise platform landscape, explore Zycus’ 2026 Gartner Magic Quadrant for Source-to-Pay Suites. This evaluation should not be treated as evidence for the sourcing outcome comparisons above.

