Enterprises once picked AI tools by running a few test prompts and calling it a day. That approach collapsed the moment agents started making decisions on their own. Now procurement teams face a different reality: they need systems that reason, remember, call tools, and keep running without constant human oversight. With 80% of executives listing agentic AI as a top priority yet 40% unable to track ROI, the gap sits squarely in the procurement process.
Why Agentic AI Vendor Selection Has Changed
Enterprises once picked AI tools by running a few test prompts and calling it a day. That approach collapsed the moment agents started making decisions on their own. Now procurement teams face a different reality: they need systems that reason, remember, call tools, and keep running without constant human oversight.
Pilot projects no longer cut it. Enterprises have moved from small experiments to production environments where agents handle customer service tickets, route internal requests, and execute multi-step workflows across legacy systems. This shift exposed the limits of old evaluation habits.
Accuracy alone means nothing if the agent:
- Costs too much to run at scale
- Slows down under real traffic loads
- Drifts into unmonitored failures after deployment
Survey data from late 2025 reinforces this shift. Teams that deployed agents reported success on simple tasks, followed by silent failures when the context spanned multiple systems or the workload scaled beyond sandbox conditions. Procurement teams learned the hard way that vendor benchmark slides rarely reflect these realities.
The numbers behind the gap
80% of executives rank agentic AI as a top priority
Agentic AI has moved from experimental to strategic across the enterprise landscape.
40% cannot track the return on those deployments
The gap sits squarely in the procurement process — and it is widening.
$7.29B in 2025, projected at $9.14B for 2026
Enterprise-specific estimates range from $6.8B to $10.9B, with a projected CAGR of 40–50% through 2034.
Growth is accelerating, but ROI visibility has not kept pace.
The CLEAR Evaluation Framework
Vendors still lead with benchmark scores. Enterprises now score them on five dimensions instead. The CLEAR framework has replaced single-number accuracy tests with a production-ready scorecard.
CLEAR Defined
| Dimension | What It Measures |
|---|---|
| C — Cost | Inference and memory expenses at scale |
| L — Latency | Response time when the agent chains multiple tool calls or accesses persistent memory |
| E — Efficacy | Whether the agent completes the actual goal, not just one sub-task |
| A — Assurance | Adherence to security policies, data boundaries, and compliance rules |
| R — Reliability | Consistent behavior across thousands of cycles and changing environments |
How Enterprises Apply CLEAR Inside RFPs
Teams assign weights to each dimension based on their specific deployment context.
- A customer-service deployment might weight Latency and Assurance highest, since response speed and policy adherence directly affect customer outcomes.
- A back-office automation project might prioritize Cost and Reliability, since the agent runs unattended for long periods.
The framework forces vendors to supply concrete numbers instead of vague promises. One enterprise evaluation run across six leading agents on 300 tasks found that agents optimized only for accuracy ended up 4.4 to 10.8 times more expensive than alternatives tuned for the full CLEAR scorecard.
Where Vendors Commonly Fail CLEAR Evaluations
Real evaluations surface gaps that demos never reveal.
- Assurance failures appear when agents must decide whether to escalate sensitive financial requests without a hard-coded policy anchor.
- Reliability drops when persistent memory fills or when tool APIs change without notice.
- Cost overruns surface when multi-hop tool calls are not optimized for token efficiency at scale.
Procurement teams now demand live dashboards that surface CLEAR scores in real time, not one-time benchmark reports.
Agents optimized only for accuracy ended up 4.4 to 10.8 times more expensive than alternatives tuned for the full CLEAR scorecard.
Top Procurement Challenges
1. Security and Data Sovereignty
Enterprises refuse to grant autonomous decision rights to models trained on data they do not control. Compliance teams require full audit logs for every action an agent takes. The EU AI Act enforcement schedule adds considerable pressure in 2026, especially for high-risk use cases. Procurement questionnaires now include explicit regulatory mapping requirements.
2. Legacy System Integration
Most organizations run on decades-old ERP, CRM, and custom databases. Agents that work in sandbox environments fail when they encounter authentication layers, rate limits, or inconsistent data formats in production. Internal expertise shortages make debugging worse — few teams have the in-house talent to diagnose agent memory drift or redesign orchestration layers on short notice.
3. The Procurement Wall
The procurement wall appears before any contract reaches legal review. Enterprises auto-block generic AI sales outreach and demand governance documentation upfront. Sales teams that arrive with nothing more than a product tour get filtered out immediately.
The expectation in 2026 is clear:
- Provide audit-ready action traces before the first meeting ends
- Demonstrate "Know Your Agent" (KYA) tooling
- Show clear ownership frameworks for agent failures
Challenge Summary
| Challenge | Root Cause | Enterprise Requirement |
|---|---|---|
| Security and sovereignty | Models trained on uncontrolled data | Full audit logs; data residency guarantees |
| Legacy system integration | Authentication layers; inconsistent data formats | Proven connectors for specific stacks |
| Expertise gaps | Limited internal agent-debugging talent | Vendor-provided training and support tiers |
| Regulatory compliance | EU AI Act; sector-specific rules | Pre-mapped compliance documentation |
| Procurement wall | Generic sales outreach | Governance docs, KYA tooling, failure ownership |
"Know Your Agent" is becoming a baseline requirement
KYA tooling assigns traceable ownership to every autonomous action an agent takes — similar in concept to Know Your Customer (KYC) frameworks in financial services. Any agent decision must be traceable back to its originating policy, permission scope, and data inputs. In regulated industries, vendors who cannot demonstrate KYA are increasingly filtered out before the first meeting ends.
Real Selection Criteria and the Competitive Landscape
Observability and Evaluation Tooling
Observability tooling ranks as the top deal maker across agentic AI markets. Buyers insist on platforms that monitor agent behavior after deployment, not just during testing. This includes:
- Real-time dashboards tracking CLEAR metrics
- Anomaly detection for unexpected agent actions
- Full replay capability for auditing specific decisions
Persistent Memory Management
An agent that forgets context after twenty steps or calls the wrong API at the wrong time creates downstream damage that no accuracy metric catches. Vendors must demonstrate how their memory architecture handles:
- Long-horizon multi-step tasks
- Memory pruning without context loss
- Recovery from memory corruption or API state drift
Horizontal vs. Vertical Agent Strategy
Horizontal agents still outnumber vertical ones by nearly two to one, primarily handling broad productivity tasks and knowledge work. Vertical solutions in healthcare, finance, and retail are growing faster because they embed domain rules from day one.
| Dimension | Horizontal Agents | Vertical Agents |
|---|---|---|
| Scope | Broad productivity and knowledge work | Domain-specific (healthcare, finance, retail) |
| Time to value | Faster deployment | Longer implementation cycle |
| Error tolerance | Higher | Lower |
| Growth rate | Steady | Faster than horizontal |
| Best for | Organizations prioritizing speed | Industries where error cost is high |
Incumbent vs. Specialist Trade-offs
Incumbents bundle agents into existing platforms, which reduces integration friction but increases lock-in risk. Startups carve out niches where they have reached $100 million or more in annual recurring revenue — six private companies already sit in that bracket, concentrated in customer service, voice, and coding agents.
Competitive Landscape
Salesforce built Agentforce through ten acquisitions in 2025 alone, including Spindle Analytics, Doti Search, and Qualified Automation. The strategy bundles observability and memory directly into the CRM layer, reducing the number of third-party integrations procurement teams need to manage.
Microsoft and Google push native integrations through Copilot and Vertex AI, but still require third-party add-ons for full cross-platform agent visibility.
Sierra focuses on high-touch customer-service agents that handle interruptions at enterprise scale, with an architecture specifically designed for agents that need to stay coherent across long, unpredictable conversations.
PolyAI optimized its voice architecture for barge-in detection and silence handling across millions of calls — a specialist choice for organizations running high-volume voice channels.
Cognition, Harvey, and the Moveworks team (now part of ServiceNow) dominate domain-specific automation in coding, legal, and IT service workflows respectively.
| Company | Type | Core Positioning | Primary Differentiator |
|---|---|---|---|
| Salesforce | Incumbent | Enterprise CRM / Agentforce | 10 AI agent acquisitions in 2025; bundled observability |
| Microsoft | Incumbent | Platform ecosystem (Copilot) | Native Microsoft 365 integration |
| Incumbent | Platform ecosystem (Vertex AI) | Multi-modal agent support; GCP native | |
| Sierra | Disruptor | Customer-service agents | Enterprise-scale interruption handling |
| PolyAI | Disruptor | Voice-first agents | Barge-in detection; latency-optimized architecture |
| Cognition | Disruptor | Coding agents | Autonomous software development task execution |
| Harvey | Disruptor | Legal AI agents | Domain-specific legal reasoning and document generation |
| Moveworks (ServiceNow) | Disruptor | IT service automation | $100M+ ARR; integrated into ServiceNow workflows |
The broader market includes over 400 active startups across 16 categories, with approximately 1,700 tracked in total. Revenue concentration is heaviest in software-development agents. M&A activity reached 10% of all AI acquisitions in 2025 as incumbents moved quickly to close capability gaps.
Step-by-Step Vendor Selection Guide
Use this checklist to move from RFP preparation through post-deployment governance. Each step maps to a distinct phase of the procurement process.
Step 1: Write an RFP that survives internal scrutiny
Include required CLEAR scores with acceptable thresholds, specific audit log formats and retention requirements, data sovereignty guarantees and encryption standards, scenario-based tests that simulate production failures (not just clean-path demos), and proof of integration with your specific technology stack — not generic API documentation.
Step 2: Structure demos around red flags and green signals
Red flags: vendors who dodge questions about memory persistence, no explanation of tool failure handling, benchmark results that cannot be reproduced in your environment.
Green signals: vendors who proactively share real-time observability dashboards, clear KYA governance features built into the platform, and references from similar-scale deployments that still track ROI six months post-launch.
Step 3: Verify references at production scale
Ask reference customers: Does the agent still perform at the CLEAR scores from your RFP response? How did the vendor respond to your first production failure? What did total cost of ownership look like at 6 and 12 months?
Step 4: Implement post-selection governance
Vendor selection is not a one-time event. Deploy KYA tooling so every autonomous action carries traceable ownership. Map the entire agent workflow against applicable EU AI Act risk categories. Schedule quarterly CLEAR re-scoring sessions with the vendor and define clear escalation paths and SLAs for failure scenarios.
Frequently Asked Questions
What is the CLEAR framework for evaluating agentic AI vendors?
CLEAR stands for Cost, Latency, Efficacy, Assurance, and Reliability. It expands evaluation beyond accuracy to include the metrics that actually predict success in production environments. Enterprises use it inside RFPs to require vendors to supply measurable data on expenses, response times, goal completion rates, policy adherence, and long-term consistency — rather than benchmark scores alone.
How large is the agentic AI market projected to be in 2026?
The global market reached $7.29 billion in 2025 and is projected at $9.14 billion for 2026. Enterprise-specific estimates range from $6.8 billion to $10.9 billion. CAGRs sit between 40 and 50 percent depending on the forecast source, with sustained growth projected through the early 2030s.
What are the biggest procurement barriers when selecting agentic AI solutions?
Security and sovereignty requirements top the list, followed by integration challenges with legacy systems and internal expertise shortages. Procurement teams also encounter the "procurement wall," where generic sales outreach is filtered out and strict, audit-ready governance documentation becomes a non-negotiable entry requirement.
Which companies lead in agentic AI capabilities right now?
Salesforce leads through its Agentforce platform and an aggressive acquisition strategy. Microsoft and Google offer strong platform integrations. Disruptors including Sierra, PolyAI, and the Moveworks team (now part of ServiceNow) deliver specialized performance in customer service, voice automation, and domain-specific workflows.
How should enterprises prepare for EU AI Act requirements in agent procurement?
Include explicit EU AI Act risk category mapping in every RFP. Require full audit logs for all autonomous actions and KYA tooling from every shortlisted vendor. Schedule compliance reviews before contract signature and build ongoing monitoring into the governance plan as enforcement ramps through 2026.
What is "Know Your Agent" (KYA) tooling?
KYA tooling refers to governance infrastructure that assigns traceable ownership to every autonomous action an agent takes. Similar in concept to Know Your Customer (KYC) frameworks in financial services, KYA ensures that any agent decision can be traced back to its originating policy, permission scope, and data inputs. It is becoming a baseline requirement for enterprise procurement in regulated industries.
Build with Octopus Builds
Need help turning the article into an actual system?
We design the operating model, product surface, and delivery plan behind AI systems that need to ship cleanly and keep working in production.