Microsoft · Published 2026-08-24

This company has more AI agents than employees

privateReview: unreviewedConfidence: high
AI agentscopilotChange managemententerprise transformationupskillingai-agentsagent-governanceno-codeenterprise-adoptionagent-lifecyclechange-managementagent-factory
Open on YouTube ↗

Summary

Overview

  • Speaker: Andy Doyle & Molly Wood
  • Channel: Microsoft
  • Main topic: Enterprise AI agent transformation, workforce change management, and human-centric technology adoption at Kantar
  • Purpose: To explore how a non-tech enterprise successfully integrated AI agents into everyday operations, managed employee anxiety around workplace change, and fostered a culture of continuous learning and agent building. Andy Doyle, Chief People & Agent Officer at Kantar, discusses how the global marketing data and analytics company transformed its workforce by deploying over 15,000 AI agents. Kantar transitioned from traditional software deployment to a human-centric change management approach, achieving high daily usage rates, significant time savings, and new cultural paradigms where non-technical employees build AI agents to solve business problems.

Topic Map

The Shift from Technology to People in AI Transformation

  • Explanation: Most tech organizations focus heavily on turning technology on rather than developing human skills to experiment and build agents without fear of failure.
  • Key claims:
    • Most tech organizations assume turning technology on equals deployment.
    • Employees are scared of failure and making mistakes with new technology.
    • Using agents is an iterative process, not a binary success or failure.
  • Examples:
    • Kantar rolled out Copilot licenses to employees and encouraged experimentation.
  • Terminology:
    • agent
    • copilot
    • iteration
  • Why it matters: Organizations must focus on human skill development and psychological safety rather than just software deployment.

Kantar's Agent Factory and Three-Tier Agent Model

  • Explanation: Kantar categorized their 15,000+ AI agents into three models—no-code, low-code, and pro-code—managed through an internal Agent Factory.
  • Key claims:
    • Kantar has more AI agents (15,000+) than employees (12,000).
    • Agents span from individual personal assistants to enterprise-grade corporate tools.
    • The Agent Factory allows decentralized idea generation with centralized governance and IP control.
  • Examples:
    • A People Agent that handles HR queries with a goal of 95% automation without human intervention by the end of the year.
  • Terminology:
    • no-code
    • low-code
    • pro-code
    • agent factory
  • Why it matters: Empowering non-technical business professionals to build agents creates practical tools that directly address real business pain points.

Managing Employee Anxiety and Continuous Training

  • Explanation: Addressing employee fears about job security and change requires transparency, weekly training, and focusing on human-centered transformation.
  • Key claims:
    • Around 85% of employees feel there is too much change going on in the world.
    • Kantar committed to not replacing people who leave, instead using agents to pick up slack.
    • Weekly training sessions are held for all employees to keep pace with rapid technological updates.
  • Examples:
    • Weekly Tuesday training sessions on new Copilot features and agent building.
  • Terminology:
    • change management
    • upskilling
    • adoption
  • Why it matters: Sustainable AI adoption requires continuous learning loops and addressing employee anxiety head-on.

Key Points

Agents Outnumber Employees

  • Explanation: Kantar operates with over 15,000 AI agents for a workforce of 12,000 employees, ranging from personal productivity helpers to enterprise tools.
  • Evidence: 15,000+ agents deployed across 12,000 employees with an 85% daily usage rate.
  • Practical implication: Enterprises can scale operational efficiency by enabling employees to build and utilize multiple specialized agents.

Business-Led Agent Building

  • Explanation: Agents were built largely by business professionals rather than dedicated IT teams, ensuring solutions directly target actual operational friction.
  • Evidence: The people agent for HR queries built by non-tech staff achieving high automation rates.
  • Practical implication: Democratizing agent creation bridges the gap between technical capability and business domain expertise.

Continuous Weekly Training vs. One-Off Rollouts

  • Explanation: Kantar conducts weekly training sessions every Tuesday rather than relying on initial onboarding to ensure teams keep pace with fast-evolving AI capabilities.
  • Evidence: Weekly Tuesday training programs covering new AI features and agent building.
  • Practical implication: Ongoing education prevents employees from falling back into legacy work habits.

Frameworks, Models & Processes

The Three-Tier Agent Model

  • How it works: Classifies internal AI agents based on complexity of creation and scope of deployment.
  • Components:
    • No-code agents
    • Low-code agents
    • Pro-code agents
  • When to use: When scaling enterprise AI adoption across diverse technical skill levels.

Examples & Case Studies

Kantar rolled out Copilot to all 200 employees in their longest-standing office as a pilot.

  • Illustrates: Bottom-up experimentation and grassroots discovery of agent utility.
  • Lesson: Broad initial access sparks unexpected grassroots innovation and employee ownership.

Kantar CEO used self-service AI agents for employment verification letters.

  • Illustrates: Executive dogfooding and practical productivity enhancement at all levels.
  • Lesson: Leadership adoption validates tools and builds organizational credibility for new technologies.

Actionable Takeaways

  • Immediate:
    • Implement weekly training sessions to keep employees updated on AI tools.
    • Encourage low-stakes experimentation and treat failures as iterative learning.
  • Strategic:
    • Democratize agent creation by allowing business units to build their own no-code/low-code solutions.
    • Combine change management with transparent communication regarding job security and AI.
  • Questions to investigate:
    • How do we measure true ROI on decentralized enterprise AI agent deployments?
    • How do we balance governance and IP control with employee-led agent creation?

Claims Worth Verifying

  • Kantar achieved an 85% daily Copilot usage rate across their rollout. (statistic)
  • Kantar saved an average of two hours per person per week in the first year of Copilot adoption. (statistic)

Notable Quotes

"Most tech organizations would say we've turned it on. We've deployed it in the ecosystem. It's there. Co-pilot, you've got co-pilot." "We have more agents than people in the organization."

Compressed Summary

  • Kantar deployed over 15,000 AI agents for 12,000 employees.
  • Shifted from IT-driven deployment to business-led, no-code/low-code agent building.
  • Achieved 85% daily usage and saved two hours per person weekly through continuous training and psychological safety.
  • Keywords: ai agents, copilot, change management, enterprise transformation, upskilling
  • Core insight: Successful enterprise AI transformation requires focusing on human mindset, continuous training, and democratizing agent creation across non-technical business teams.

Core insights

6
Architecturemedium noveltymoderate evidence

Kantar decouples agent creation from agent governance: an internal 'Agent Factory' lets business units generate agents freely while centralizing IP control and governance. This is a federated-authoring / centralized-control architecture rather than either fully centralized IT delivery or ungoverned citizen development.

Why it matters

Most agent platforms force a choice between a single centrally-owned assistant and unmanaged sprawl of user-created bots. This pattern suggests the viable middle is a platform boundary: permissive authoring surface plus a central registry/policy layer that owns identity, IP, and lifecycle — the same split that internal developer platforms use for services.

Generalization

Agent platforms should be designed as authoring plane + control plane, not as a monolith or a free-for-all.

The Agent Factory allows decentralized idea generation with centralized governance and IP control.
Open source video
Empirical Resulthigh noveltymoderate evidence

The unit of agent deployment shifts from 'one enterprise assistant' to many narrow, scope-bounded agents: 15,000+ agents across 12,000 employees, spanning personal assistants up to enterprise-grade tools. Agent count exceeding headcount implies a portfolio, not a product.

Why it matters

If agents outnumber people, engineers must solve discovery, naming, deduplication, ownership, versioning, and deprecation for a long tail of small agents — problems that do not exist for a single-assistant architecture. It also implies per-agent evaluation and cost attribution must be lightweight enough to be applied thousands of times.

Generalization

As agent creation becomes cheap, the scarce engineering resource moves from building agents to cataloguing, evaluating, and retiring them.

We have more agents than people in the organization.
Open source video
Agents span from individual personal assistants to enterprise-grade corporate tools.
Open source video
Practicemedium noveltymoderate evidence

The primary supply of agents came from non-technical business staff rather than IT, with the cited example being an HR 'People Agent' built by non-tech staff targeting 95% automation without human intervention.

Why it matters

This moves the binding constraint on an agent platform from model capability to the affordances of the authoring harness: if domain experts can build production agents, the platform must expose safe tool access, data boundaries, and fallback-to-human paths through no-code surfaces rather than SDKs.

Generalization

Democratized agent authoring is only as good as the guardrails embedded in the no-code surface; unguarded authoring transfers risk, not capability.

The people agent for HR queries built by non-tech staff achieving high automation rates.
Open source video
A People Agent that handles HR queries with a goal of 95% automation without human intervention by the end of the year.
Open source video
Mental Modelmedium noveltymoderate evidence

Adoption is framed as an iterative loop where using agents is explicitly not binary success/failure, and employee fear of making mistakes is named as the adoption blocker rather than lack of tooling.

Why it matters

Psychological safety is treated as an engineering-relevant variable: it determines whether users explore agent capabilities or fall back to legacy workflows, which in turn determines whether usage metrics reflect real integration or license compliance.

Generalization

Agent rollout success metrics should measure behavioral change under low-stakes experimentation, not activation or seat count.

Using agents is an iterative process, not a binary success or failure.
Open source video
Employees are scared of failure and making mistakes with new technology.
Open source video
Practicelow noveltymoderate evidence

Enablement runs on a fixed weekly cadence (every Tuesday) covering new Copilot features and agent building, rather than one-off onboarding — an enablement loop coupled to product release velocity.

Why it matters

Capability drift is the predictable failure mode when a platform ships faster than users learn it; a recurring, high-frequency training loop is an operational answer that also doubles as a feedback channel from users back into platform priorities.

Generalization

Treat enablement as a scheduled service with its own cadence and backlog, not a launch event.

Weekly training sessions are held for all employees to keep pace with rapid technological updates.
Open source video
Ongoing education prevents employees from falling back into legacy work habits.
Open source video
Practicemedium noveltyweak evidence

Kantar rolled out Copilot to the entire population of one long-standing office (200 employees) at once, framing broad initial access as the mechanism that produced unexpected grassroots use cases rather than a staged pilot.

Why it matters

This is a testable counter-position to phased pilot-then-scale rollouts: it trades control over early usage patterns for a wider search over possible applications, and it argues discovery value comes from breadth of access, not depth of a curated pilot group.

Generalization

For high-variance exploration tasks, access breadth may beat pilot depth as a discovery strategy.

Kantar rolled out Copilot to all 200 employees in their longest-standing office as a pilot.
Open source video
Broad initial access sparks unexpected grassroots innovation and employee ownership.
Open source video

Deep dives

4

Federated authoring with centralized control: the Agent Factory as authoring plane + control plane

Research question

How can an internal Agent Factory allow decentralized agent creation while enforcing centralized governance, IP control, and lifecycle policy?

Why

Most agent platforms force a choice between a single centrally owned assistant and unmanaged citizen development. The Kantar pattern suggests the viable middle is the internal developer platform split: a permissive authoring surface plus a central registry and policy layer that owns identity, IP, and lifecycle.

The Agent Factory allows decentralized idea generation with centralized governance and IP control.
Open source video
Empowering non-technical business professionals to build agents creates practical tools that directly address real business pain points.
Open source video
Source video

Agent portfolio lifecycle: cataloguing, deduplication, and deprecation when agents outnumber employees

Research question

What platform primitives are required to discover, deduplicate, attribute cost, and retire thousands of narrow agents that outnumber their human owners?

Why

Agent count exceeding headcount turns discovery, ownership, and deprecation into first-class engineering problems. Without them, a portfolio becomes unmanageable sprawl and ROI cannot be attributed per agent or per portfolio.

Kantar has more AI agents (15,000+) than employees (12,000).
Open source video
Agents span from individual personal assistants to enterprise-grade corporate tools.
Open source video
Source video

Safety guardrails in no-code agent authoring for high-autonomy domain agents

Research question

What constrained tool surfaces, refusal paths, and sampling-based audit mechanisms are needed when non-technical staff author agents targeting near-full automation of sensitive processes like HR?

Why

The People Agent target of 95% automation without human intervention implies employee-authored logic touches sensitive workflows without described verification. The guardrail boundary in the no-code surface determines whether democratized authoring transfers capability or merely relocates risk.

A People Agent that handles HR queries with a goal of 95% automation without human intervention by the end of the year.
Open source video
Kantar committed to not replacing people who leave, instead using agents to pick up slack.
Open source video
Source video

Adoption telemetry: differentiating usage from value in agent rollouts

Research question

What depth metrics (tasks completed end-to-end, human handoff rate, escalation rate) should replace session or seat counts to determine whether high daily usage reflects real integration or license compliance?

Why

Reported high daily usage and time savings do not attribute value per agent or establish a counterfactual. Without depth metrics, platform investments are justified by access rather than behavior change.

achieving high daily usage rates, significant time savings
Open source video
Using agents is an iterative process, not a binary success or failure.
Open source video
Source video

Article ideas

4

The Agent Factory Must Be an Authoring Plane, Not a Free-for-All

Enterprise agent platforms should adopt the internal developer platform pattern — a permissive self-service creation surface paired with a centrally owned control plane for identity, policy, IP, and lifecycle — rather than choosing between central IT delivery and ungoverned citizen development.

Angle

Argue from the Kantar Agent Factory example that the governance/IP tension is an architecture problem, not a policy problem; show what primitives the control plane must expose at the authoring layer.

Source video

The No-Code Agent Surface Is Where Risk Actually Lives

Democratized agent authoring does not remove risk; it relocates it into the affordances of the no-code surface, so safe tool access, data boundaries, and fallback-to-human paths must be built into authoring harnesses for non-technical domain experts.

Angle

Use the HR People Agent targeting 95% automation without human intervention as the stress case; ask what verification layer replaces engineering review when business staff ship production agents.

Source video

Your Agent Count Will Exceed Your Headcount. Here's the Boring Work That Follows.

Once agent creation is cheap, the scarce engineering resource shifts from building agents to cataloguing, evaluating, attributing cost to, and retiring them; design for discovery, deduplication, ownership, and deprecation before celebrating 15,000 agents.

Angle

Use the 15,000 agents / 12,000 employees ratio as the forcing function; enumerate lifecycle features that do not exist in single-assistant architectures.

Source video

Psychological Safety Is an Engineering Metric

Treating agent usage as iterative rather than binary, and naming fear of mistakes as the adoption blocker, changes what platforms must measure and how rollouts must be instrumented — behavioral change under low-stakes experimentation, not activation or seat count.

Angle

Counter the standard adoption dashboard; propose metrics that detect fallback to legacy workflows and fear-driven non-use.

Source video

Project ideas

4

Promotion Gate Prototype for No-Code to Pro-Code Agents

gatehouse

Automated promotion gates based on data sensitivity, blast radius, and autonomy level can let business-authored agents graduate to governed production without manual review of every agent, while catching high-risk agents before they ship.

Proof of concept

Define a three-tier promotion policy and implement a gate that blocks or requires review when an agent touches sensitive HR data or requests high-autonomy tool access; test against the People Agent scenario.

Measurement

False positive and false negative rate of the gate against expert review; time-to-production for low-risk agents; number of high-risk agents intercepted before production.

Source video

Agent Usage Depth Telemetry

beyond-evals

Seat and session usage metrics overstate agent integration; depth metrics (tasks completed end-to-end, human handoff rate, escalation rate) will reveal a materially lower value-producing usage rate than headline daily usage.

Proof of concept

Instrument a small agent deployment to log task completion, handoff, and escalation events; compare depth metrics against simple daily-active usage over 4–6 weeks.

Measurement

Ratio of depth-active users to daily-active users; end-to-end task completion rate; handoff and escalation rates; correlation with self-reported time savings.

Source video

Agent Catalog & Staleness Scorer

gatehouse

In an agent portfolio where agents outnumber employees, a lightweight catalog with ownership metadata, staleness scoring, and duplicate detection can reduce orphaned agents and redundant agent creation by a measurable margin within one quarter.

Proof of concept

Build a registry that ingests agent definitions from no-code, low-code, and pro-code tiers, assigns ownership, computes last-run and last-modified staleness, and flags near-duplicate intents; test on a synthetic or small real portfolio.

Measurement

Percentage of agents with identified owner; number of duplicate clusters detected; median time from staleness flag to deprecation or refresh; reduction in new duplicate agents after catalog introduction.

Source video

Weekly Enablement as Release-Coupled Service

movement-lab

Treating enablement as a fixed weekly cadence versioned with platform releases will reduce capability drift and legacy-workflow fallback more than one-off onboarding, measurable through feature adoption and fallback rates.

Proof of concept

Run a weekly Tuesday training loop for one business unit, publish training as release artifacts, and instrument adoption of new agent-building features and fallback to legacy workflows.

Measurement

Time from feature release to first successful employee-authored agent using it; percentage of employees attempting new features within two weeks; decline in legacy-workflow fallback events.

Source video

Architectural implications

4

Agent supply is federated across business units rather than owned by IT.

Before

Central IT team authors and ships a small number of enterprise agents.

After

Business professionals author no-code/low-code agents, with central control limited to governance and IP.

Consequence

The platform must provide policy, identity, data-boundary, and lifecycle primitives at the authoring layer, because reviewers cannot inspect every agent before it ships.

Source video

Agents are tiered by creation complexity (no-code, low-code, pro-code) rather than by business function.

Before

Agents are categorized by team, department, or use case.

After

Agents are categorized by the skill level required to build and maintain them.

Consequence

The tier becomes a promotion path: an agent can graduate from personal no-code prototype to governed pro-code enterprise tool, which requires a defined escalation and ownership handoff.

Source video

Agent population exceeds employee population while spanning personal helpers to enterprise tools.

Before

A single assistant with a single tool surface, memory model, and evaluation suite.

After

Thousands of narrow agents with heterogeneous scopes and owners.

Consequence

Discovery, deduplication, per-agent cost/usage attribution, and deprecation become first-class platform features rather than afterthoughts.

Source video

Adoption is driven by a recurring weekly enablement loop tied to new feature releases.

Before

One-off onboarding at launch; users operate a frozen feature set.

After

Continuous feature ingestion by users on a fixed weekly cadence.

Consequence

Enablement must be versioned alongside the platform, and training content becomes part of the release process.

Source video

Tradeoffs and failure modes

5

Democratized agent creation

Benefit

Solutions target real operational friction because domain experts, not IT, define the problem.

Cost or risk

Central governance and IP control become harder to enforce as authorship spreads to non-engineers.

How do we balance governance and IP control with employee-led agent creation?
Open source video
Source video

Agent sprawl (agents outnumber employees)

Benefit

Enables many narrow, domain-fit agents rather than one generic assistant.

Cost or risk

True ROI of decentralized deployments becomes difficult to measure across thousands of small agents.

How do we measure true ROI on decentralized enterprise AI agent deployments?
Open source video
Source video

Failing forward / low-stakes experimentation

Benefit

Lowers the psychological barrier that stops employees from adopting agents at all.

Cost or risk

A tolerant stance toward failed attempts can coexist with inadequate review of agents that reach production or touch sensitive data.

Encourage low-stakes experimentation and treat failures as iterative learning.
Open source video
Source video

No-replacement headcount policy

Benefit

Addresses job-security anxiety directly, which the summary ties to adoption.

Cost or risk

Agents must absorb the workload of departing staff, raising the reliability bar for agents that were built by non-technical authors.

Kantar committed to not replacing people who leave, instead using agents to pick up slack.
Open source video
Source video

Broad-simultaneous rollout over staged pilot

Benefit

Surfaces unexpected grassroots use cases and builds employee ownership.

Cost or risk

Less control over early usage patterns and weaker ability to intercept bad practices before they spread.

Broad initial access sparks unexpected grassroots innovation and employee ownership.
Open source video
Source video

Open questions

5

How should ROI be measured for a decentralized agent portfolio rather than per-agent?

Why unresolved

The summary presents the question as open and offers only aggregate adoption statistics (85% daily usage, two hours saved per person weekly), which do not attribute value to individual agents.

Research direction

Define per-agent telemetry plus portfolio-level metrics (task deflection, cycle time, escalation rate) that can be aggregated without requiring manual per-agent review.

Source video

Where exactly should the boundary sit between employee-authored agents and centrally governed agents?

Why unresolved

The summary states the tension exists and names IP control as the concern, but gives no rule, threshold, or gate for when an agent must graduate to central ownership.

Research direction

Prototype promotion criteria based on data sensitivity, blast radius, and autonomy level, with automated gates between no-code, low-code, and pro-code tiers.

Source video

What is the lifecycle model for agents that outnumber their human owners?

Why unresolved

Agent counts exceed employee counts, implying many agents exist without a dedicated maintainer, but the summary describes no ownership or deprecation process.

Research direction

Study orphaned-agent detection, staleness scoring, and deprecation policy in agent catalogs.

Source video

Does high daily usage rate actually indicate value creation, or only access?

Why unresolved

The 85% figure is reported without a definition of 'usage', baseline, or control group, and is flagged as a claim worth verifying.

Research direction

Instrument usage with depth metrics (tasks completed end-to-end, human handoff rate) instead of session or seat counts.

Source video

What safety mechanism replaces human review when non-technical staff ship agents toward 95% autonomous handling?

Why unresolved

The HR People Agent targets high automation without human intervention, but the summary describes no verification layer between employee-authored logic and HR outcomes.

Research direction

Evaluate constrained tool surfaces, refusal paths, and sampling-based audit for domain agents authored without engineering review.

Source video

Key claims

7
factualVerification needed

Kantar operates more AI agents than employees (15,000+ agents for 12,000 employees).

Evidence

We have more agents than people in the organization.

Question

How are 'agents' counted — distinct deployed agents, published agent definitions, or per-user instances — and does the count include abandoned drafts?

Source video
causalVerification needed

Decentralized idea generation can be combined with centralized governance and IP control through an internal Agent Factory.

Evidence

The Agent Factory allows decentralized idea generation with centralized governance and IP control.

Question

What specific controls (review gates, policy enforcement, IP scanning) does the Agent Factory apply, and at which stage of the agent lifecycle?

Source video
causalVerification needed

Non-technical business staff built agents such as an HR People Agent with a goal of 95% automation without human intervention.

Evidence

The people agent for HR queries built by non-tech staff achieving high automation rates.

Question

What is the measured automation rate and error/escalation rate of that agent in production?

Source video
factualVerification needed

Kantar achieved an 85% daily Copilot usage rate across their rollout.

Evidence

Kantar achieved an 85% daily Copilot usage rate across their rollout.

Question

What counts as 'daily usage' (any session, minimum interaction), and what population and time window does the rate cover?

Source video
causalVerification needed

Copilot adoption saved an average of two hours per person per week in the first year.

Evidence

Kantar saved an average of two hours per person per week in the first year of Copilot adoption.

Question

How was time savings measured — self-report, time-tracking, or output comparison — and against what counterfactual?

Source video
factualVerification needed

Roughly 85% of employees feel there is too much change going on in the world.

Evidence

Around 85% of employees feel there is too much change going on in the world.

Question

What survey instrument, sample, and wording produced this figure, and what is the employee population it represents?

Source video
causalVerification needed

Rolling out Copilot to an entire 200-person office at once produced grassroots innovation and employee ownership.

Evidence

Broad initial access sparks unexpected grassroots innovation and employee ownership.

Question

Compared to a staged rollout, did broad access produce measurably more distinct use cases or higher sustained usage?

Source video

Connections

5