US BPO Agentic AI Leaders Report — 2026

US BPO market map

The eleven providers collectively run more than 1.1 million US-served agent seats across voice, chat and back-office work. Tier-1 (Teleperformance, Concentrix, TTEC, Foundever, Alorica) own the majority of Fortune 500 CX outsourcing spend. Tier-2 (Sutherland, TaskUs, iQor) specialize by vertical — trust and safety, healthcare, technology. Tier-3 (Conduent, IBEX, Startek) skew back-office, public sector and mid-market.

Every provider's 2026 narrative leads with agentic AI. The differentiation is not the story — it is whether the AI shows up in the SOW as a commercial commitment, a delivery method or a marketing footnote.

• Tier-1 average seat footprint: 120k–260k US-served
• Tier-2 average seat footprint: 35k–90k US-served
• Tier-3 average seat footprint: 15k–45k US-served
• 9 of 11 providers ship at least one production agentic workflow; only 4 price it as an outcome

Agentic AI maturity model

The five-level maturity model scores providers on production agentic workflows, evaluation and safety discipline, first-party knowledge tooling, willingness to price outcomes and depth of platform partnerships across Amazon Connect, Google CCAI, Genesys, NICE, Salesforce Agentforce, Microsoft Copilot Studio and Kore.ai.

Level 1 is copilot pilots. Level 5 is outcome-owned agentic workflows with a client-visible SLA. Most of the market sits at Level 2 to Level 3 today. A small group — two Tier-1s and one Tier-2 — operates credibly at Level 4 across more than one client program.

• Level 1 — Copilot pilots, seat-priced, no evaluation discipline
• Level 2 — Agent assist in production, AHT-linked bonus in select contracts
• Level 3 — Automated QA at 100% coverage, first-party knowledge base, per-tenant isolation
• Level 4 — Agentic workflows in production with pre-action approval HITL, outcome-linked pricing
• Level 5 — Outcome-owned agentic workflows, client-visible SLA, gain-share default

Voice & conversational AI benchmarks

Tier-1 top-quartile providers land 55% to 62% containment on transactional intents in English, dropping to 38% to 48% on complex intents and 32% to 42% in Spanish. Tier-2 leaders match on English transactional but trail by 8 to 12 points on complexity and language mix.

The gap is rarely the model. It is intent design discipline, disambiguation depth and whether the provider owns the containment budget per intent or applies a single global target. Providers with a per-intent budget are the ones landing sustainable numbers without a CSAT cliff.

• Tier-1 English transactional containment: 55%–62% (top-quartile)
• Tier-1 complex intent containment: 38%–48%
• Tier-1 Spanish containment: 32%–42%
• Median CSAT delta on AI-handled vs. human-handled: −2 to +1 point when scoped to right intents

Agent assist & 100% QA

Real-time agent assist delivers a median 17% to 22% AHT reduction across the eleven providers, with the top quartile hitting 24% to 28% on transactional queues. Ramp-time to proficiency compresses 30% to 45% when assist is paired with structured onboarding.

100% automated QA has replaced 2%–5% sample QA at seven of the eleven providers for at least one client program. QA delivery cost drops 45% to 60%; the operational lift is the calibration cadence, not the tooling.

• Median AHT reduction with agent assist: 19%
• Top-quartile AHT reduction on transactional queues: 24%–28%
• Ramp-time compression: 30%–45%
• QA delivery cost reduction at 100% coverage: 45%–60%

Back-office agents (IDP + LLM)

Back-office agentic workflows — claims triage, KYC review, order-to-cash exceptions, document classification — are the largest margin lever in the report. Straight-through processing rates of 55% to 75% are showing up in the top-quartile programs, with human review reserved for exceptions.

The commercial pattern that works is per-transaction pricing with an outcome floor. Seat-based pricing on IDP + LLM workflows leaves margin on the table for the provider and does not align incentives with the client.

Commercial models

Four commercial-model patterns show up repeatedly. Outcome-based (per resolved contact, per approved claim). Gain-share (AI savings split against a baseline). Per-transaction (IDP and back-office). Managed AI retainer (platform + operate fee on top of seat rate).

Providers at Level 4+ are willing to sign at least two of the four. Providers at Level 2 or below default to seat-based with an AI marketing overlay. Contract-clause patterns matter: baseline definition, savings-attribution methodology and change-of-scope triggers are where the money actually lands.

• Outcome-based: per resolved contact, per approved claim, per SLA-met interaction
• Gain-share: 50/50 or 60/40 split against a documented pre-AI baseline
• Per-transaction: IDP, document classification, back-office exceptions
• Managed AI retainer: platform + operate fee, seat rate held or reduced

Buyer scorecard

The 14-question scorecard is designed to be dropped directly into an RFP. Each question is scored 0–3 with evidence requirements. The aggregate score maps to a recommended commercial-model tier — from seat-based for AI-immature providers to outcome-priced with managed AI for Level-4+ providers.

Use it to force the RFP conversation off marketing decks and onto production evidence: named client programs, evaluation cadence, HITL placement by risk class, model registry ownership and change-management gates on model version bumps.

• Q1–Q3: Production agentic workflows with named references
• Q4–Q6: Evaluation, red-teaming and safety discipline
• Q7–Q9: First-party knowledge tooling and per-tenant isolation
• Q10–Q12: Commercial willingness — outcome, gain-share, per-transaction
• Q13–Q14: Platform partnership depth and multi-vendor neutrality