Imagine presenting your agentic AI program to the CFO. You say that sourcing is faster and your team is more productive. The CFO asks the obvious follow-up: compared with what?
For many procurement teams, that is an uncomfortable question. The Hackett Group’s The Autonomous Sourcing Index finds that reported improvements are more widespread than consistent measurement practices. That gap does not mean the reported gains are false; it means they are harder to verify systematically.
What leaders report
Start with the good news. Among respondents, 30% report significant or major improvement in sourcing productivity: 22% significant and 8% major. A further 27% report moderate gains and 20% slight ones. On cycle time, 27% report significant or major improvement (15% significant, 12% major), 27% moderate and 23% slight.
There is also a cautionary note in the same data. In both cases, 23% of organizations report no improvement at all.
The business-performance picture is more mixed. When the Index asks about category, supplier and market intelligence, 54% report moderate-or-better gains. The same is true for business partnering and stakeholder satisfaction. On spend under management, 45% report moderate-or-better gains, and on spend savings, 42% do.
Respondents report value, particularly in insight-led outcomes, although these are self-reported improvements rather than independently verified results. That deserves to be taken seriously.
What leaders measure
Now compare the claims with the evidence behind them. The Index asked how organizations measure the performance of their agentic AI.
Only 10% measure consistently across deployments. Another 28% measure in selected deployments and 23% are still developing their approach. Meanwhile, 32% have no formal measurement at all, and 7% do not know.
Add up the last two and a plurality of organizations, 39%, cannot show a structured view of what their agents are doing. Another 23% are still building theirs.
The research states the consequence plainly: without a baseline, gains remain anecdotes, and scaling decisions stay harder to defend.
Why this matters more than it seems
It is tempting to treat measurement as administrative housekeeping. For agentic AI, it is closer to a prerequisite for scaling, for three reasons.
- Funding depends on proof. Agentic AI programs compete with other priorities. The Index finds that 15% of respondents cite an unclear business case or ROI as a barrier, and 13% cite funding or competing priorities. An anecdote does not survive a budget review. A baseline and a trend line do.
- You cannot tell which agents deserve expansion. Without consistent metrics, organizations scale what feels successful rather than what is. The Index also found that agent count alone does not predict outcomes, so you need performance data to know which deployments are worth extending.
- Trust requires evidence. Trust, explainability and reliability are cited as a barrier by 20% of respondents. Measurement is part of how teams earn that trust, because it shows stakeholders what the agent did, and whether it was right.
Why the gap exists
Why would organizations report gains they have not formally measured?
Part of the answer is that early deployments are exciting and visibly helpful. A category manager sees an RFx drafted in minutes, and the benefit feels obvious. Part of it is timing: many programs started as pilots, and pilots rarely set up baselines in advance. If no one recorded how long a sourcing event took before the agent, there is nothing to compare against afterward.
Part of it is also definitional. Cycle time, productivity and savings can each mean different things across categories and business units. Without agreement on how they are calculated, every deployment reports its own version.
Building a baseline that holds up
The report’s maturity playbook handles measurement stage by stage, which is a useful way to approach it without overbuilding.
- At the nascent stage, the guidance is to establish a cycle-time baseline before the pilot starts, and to track pilot adoption, output quality and policy exceptions. This is the single cheapest step available. Record how long the process takes today.
- At the developing stage, the focus shifts to category coverage, active use cases, cycle time and productivity, and adoption. The aim is repeatable measurement across workflows, rather than one-off pilot reports.
- At the scaling stage, the measures broaden to capability breadth, material outcomes and measurement coverage, meaning the share of deployments that are actually measured. This is where the 10% figure should rise.
- At the highly evolved stage, the playbook ties funding to verified outcomes and tracks enterprise value, governed autonomy, intelligence impact and stakeholder experience.
A simple discipline runs through all four: decide the metric before the deployment, record the baseline before the change and report against both afterward.
Connecting measurement to savings
One more finding is worth noting. Spend savings show the weakest reported gains of the four business outcomes, with 35% of respondents reporting none. The research suggests why: translating insights into savings depends on governed execution, not recommendations alone.
That makes measurement even more important. An agent that surfaces a savings opportunity has not saved anything until someone acts on it, and tracking that conversion from recommendation to realized value is where many programs lose the thread.
Where platform capabilities fit
Technology can make consistent measurement easier or harder. If agents run across disconnected tools, assembling a single view of performance becomes a manual project. Platforms that capture activity inside one governed environment give teams a better chance of producing consistent metrics.
Zycus, the Index’s partner, positions its Merlin Agentic AI Platform around governed execution, with agents such as Merlin Intake, Merlin Autonomous Sourcing, the Autonomous Negotiation Agent (ANA) and Merlin Analytics operating on a common source-to-pay core. When evaluating any platform, ask what data it captures for each agent action, whether you can report outcomes consistently across use cases and how it links recommendations to realized savings. The Hackett Group does not endorse any vendor depicted in its research.
Your next step
If you are scaling agentic AI, or about to, take an honest inventory this quarter. For each deployment, can you state the baseline, the metric and the result? If the answer is no for most of them, your organization may resemble the 90% of respondents that do not report consistent measurement across deployments, and you have an opportunity to get ahead of the CFO’s question.
The full report offers considerably more, including how reported improvement compares across outcomes, how organizations at different maturity stages measure value and the accountable roles and metrics for each stage of the playbook.
Download The Autonomous Sourcing Index and put your measurement approach in context.
The Hackett Group® does not endorse any participant, vendor, product or service depicted in its research. Source: The Hackett Group® in partnership with Zycus, The Autonomous Sourcing Index, 2026.
To explore Zycus’ recognition as a Leader and the only Customer Favorite in a separate industry evaluation, visit The Forrester Wave™: Supplier Value Management Platforms, Q4 2026. This recognition does not validate the reported gains in this study.
For a complementary perspective on procurement adoption, see The Hackett Group Agentic AI in Procurement Adoption Index.
For a wider view of procurement performance measures, explore Ardent Partners’ 2026 Procurement Metrics report. This separate report complements the discussion of baselines without independently verifying the Autonomous Sourcing Index respondents’ reported gains.

