95% of enterprise AI pilots never scale. TELUS Digital built 53,000 custom copilots, saved 3.8 hours per employee per week, and unlocked $600M in benefits. The blueprint is replicable.
95% of enterprise AI pilots never scale. The average company abandoned 46% of AI proof-of-concepts before production in 2025. Gartner projected that 30% of GenAI projects would be scrapped entirely after the proof-of-concept phase.
These aren't marginal failures. For every 33 AI pilots a company launches, only 4 reach production.
TELUS Digital ran a different playbook. In 2025, they processed over 2 trillion AI tokens, deployed 53,000 custom copilots built by their own employees, gave 70,000 team members access to more than 50 AI models, and logged over $600 million in measurable financial benefits.
The gap between those two stories — 95% failure rate versus 2 trillion tokens in production — is what this post is about.
!Enterprise AI scale: 2 trillion tokens vs. 95% pilot failure rate
What Pilot Purgatory Actually Looks Like
Most enterprise AI programs follow the same arc: a business unit gets excited about an AI tool, runs a successful demo, launches a pilot with a small team, sees promising results — and then stalls.
The handoff from pilot to production is where the model breaks down. IT security reviews block deployment. Procurement cycles slow vendor integration. Legal wants liability clarity that doesn't exist yet. The business unit that ran the pilot gets reorganized. The champion moves on.
Twelve months later, the company has eight AI pilots and zero production deployments.
The failure isn't usually the AI. It's the infrastructure around the AI — or the complete absence of one.
!The anatomy of pilot purgatory: promising demos, stalled deployments
How TELUS Built the Infrastructure First
TELUS Digital took a different approach. Instead of running standalone pilots tied to individual tools, they built a centralized AI platform — Fuel iX — that became the permanent infrastructure through which all AI experimentation ran.
Fuel iX is model-neutral. It gives both developers and non-technical employees access to over 50 large language models from multiple providers: Anthropic, OpenAI, Google, Mistral, and others. You pick the model that fits your task. You don't get locked into a vendor because your pilot was built on one.
This matters more than it sounds. In most enterprises, AI pilots fail not because the use case was wrong, but because the underlying model changed (a new version dropped, costs shifted, capabilities evolved) and nobody had built abstraction into the architecture. TELUS made that abstraction the foundation.
The second structural decision was governance. TELUS established a first-of-its-kind AI Governance Board — a cross-functional body that set the guardrails for responsible AI use before the tools went wide. Security, privacy, ethics, and compliance frameworks were defined centrally, once. Individual teams could then build without re-solving those problems from scratch for every use case.
That combination — a shared platform plus shared governance — is what enabled what happened next.
!Fuel iX: model-neutral platform + AI Governance Board
53,000 Copilots Built by Employees, Not by IT
The most striking number in the TELUS story isn't the 2 trillion tokens. It's the 53,000 custom copilots.
Those copilots weren't built by a central AI team. They were built by employees across the organization — people who identified problems in their own workflows and used Fuel iX's no-code interface to build tools to solve them. No Python. No API keys. No IT ticket.
TELUS leadership set a clear goal: every employee, regardless of technical skill, should be able to benefit from AI. Fuel iX was designed to make that possible. Guided workflows. Intuitive interfaces. Pre-built templates. The kind of tooling that collapses the distance between "I have a problem" and "I built something that helps."
The result was organic adoption at a scale that top-down AI programs rarely achieve. 66,000+ TELUS employees are now using AI tools. The average user saves 3.8 hours per week.
That's not a productivity statistic. That's 250 million hours of recaptured human attention across the organization annually.
!53,000 copilots built by employees, not IT — no-code AI at enterprise scale
The Agent Trainer: From Copilots to Agents
The most concrete production deployment in the TELUS story is Agent Trainer — an AI-powered voice and chat simulation for call center training.
Contact centers have a perennial onboarding problem: getting a new agent ready for live customer calls takes time, costs money, and ties up experienced staff as trainers. Traditional simulations are scripted and brittle. They don't handle the messiness of real conversations.
TELUS built Agent Trainer on ElevenLabs Conversational AI, creating synthetic voice scenarios that simulate real customer interactions across the full range of situations agents encounter. New agents practice with AI customers that behave like real ones — asking off-script questions, escalating, getting frustrated.
The result: onboarding time cut by up to 50%. Over 50,000 customer simulations completed. Agents arrive at live calls better prepared, faster.
This is what "escaping pilot purgatory" looks like in practice. Not a demo that shows what AI could do. A production system that's completing 50,000 simulations and shortening onboarding curves by half.
!Agent Trainer: 50,000 customer simulations, 50% faster onboarding
Try It Yourself: Apply This Blueprint to Your Team
The TELUS infrastructure won't be replicated overnight, but the structural decisions behind it are replicable at any scale. Here's how to start.
Step 1: Audit your current AI tool landscape.
If you don't know how many AI tools your organization is running, find out. A quick audit:
``bash
Survey your team (adapt as needed for your stack)
Questions to ask every team lead:
Which AI tools is your team actively using?
Which started as a pilot and are now standard?
Which pilots were abandoned — and why?
For engineering teams, check installed extensions and connected integrations:
VS Code / JetBrains: count AI extensions
GitHub: list connected GitHub Apps (look for Copilot, Cursor, etc.)
Slack: check installed apps for AI tools
`
This inventory is your baseline. It tells you where organic adoption is already happening — those are your signals about what people actually value.
Step 2: Pick a model-neutral evaluation framework.
Before committing to any AI platform, run every candidate through the same benchmark on your own data. A minimal framework:
`python
benchmark_models.py — run same prompts across multiple model APIs
import anthropic
import openai
import time
PROMPT = "Summarize this customer complaint in one sentence and suggest a resolution: {complaint}"
def evaluate(model_fn, name, inputs):
results = []
for inp in inputs:
start = time.time()
output = model_fn(inp)
latency = time.time() - start
results.append({"input": inp, "output": output, "latency_ms": round(latency 1000)})
print(f"\n=== {name} ===")
for r in results:
print(f" {r['latency_ms']}ms | {r['output'][:120]}...")
return results
Test with a sample of REAL queries from your domain, not synthetic ones
sample_complaints = [
"My order was supposed to arrive yesterday and nobody contacted me",
"I was charged twice and support keeps asking me to wait",
]
Add model wrappers for each provider you're evaluating
`
The key discipline: test on real prompts from your actual use case. Benchmarks on generic data are noise.
Step 3: Define your governance skeleton before the tools go wide.
TELUS's AI Governance Board sounds big-company. But the underlying checklist is not:
`markdown
AI Use Policy — Minimal Viable Governance
Data Handling
• [ ] Does the tool send data to a third party? Which one?
• [ ] Can it be used with our customer PII? (yes/no/with conditions)
• [ ] What's the retention policy on prompts/completions?
Output Risk
• [ ] What happens if the model is wrong? Who reviews?
• [ ] Is a human in the loop for consequential decisions?
Cost Controls
• [ ] Who can spin up new integrations?
• [ ] Is there a monthly spend threshold that triggers review?
• [ ] Are API keys rotated and access-logged?
Model Change Risk
• [ ] If the model is deprecated, what breaks?
• [ ] Is the integration abstracted behind an interface (easy swap)?
``
Answering these once, for the whole organization, is the structural advantage TELUS built. Every new use case inherits the answers. Pilots don't stall in security review because security already reviewed the platform.
Step 4: Run a no-code copilot sprint.
Take a team that's not technical, give them a no-code AI tool (Claude.ai, Copilot Studio, or similar), and set one constraint: build something that saves 30 minutes a week in your current workflow. Give them a half day.
The outputs won't be polished. That's fine. The exercise surfaces two things: what workflows people actually want to automate (which tells you where to invest in proper tooling), and which employees will become your internal AI champions.
TELUS's 53,000 copilots didn't start as a program. They started as permission. Give your team permission and a platform, and organic adoption follows.
What the $600M Actually Represents
TELUS reported over $600 million in financial benefits from AI since 2023. The number is large enough to be easy to dismiss as a press release statistic.
But look at the mechanics: 3.8 hours saved per employee per week, across 66,000 employees using AI tools, at an average all-in cost per employee. Even at conservative numbers, the math gets large fast.
More useful than the headline is the methodology implied by the number: TELUS was measuring. They had a system for attributing AI-driven efficiency gains to financial outcomes. Most enterprise AI programs don't. They run pilots, see qualitative enthusiasm, and then can't answer the CFO's question about ROI.
The measurement infrastructure is as important as the AI infrastructure. Define your unit economics before you deploy: hours saved, error rates reduced, calls deflected, time to resolution. Then instrument your production systems to capture those numbers.
Without that, you have anecdotes. With that, you have a $600M story.
The Pattern Behind the Blueprint
TELUS's escape from pilot purgatory wasn't a single decision. It was four concurrent structural investments:
Platform before pilots — Build the shared infrastructure first, not use-case by use-case.
Governance before tools — Define the rules once, centrally, so teams can move without asking for permission on every deployment.
Access before sophistication — Give everyone access to the tools, not just technical staff. The best use cases emerge from the people closest to the work.
Measurement before scaling — Know what you're measuring before you deploy, so outcomes are comparable and aggregable.
None of this required TELUS to bet on a specific model or vendor. That's the point. The architecture is model-neutral. The governance is tool-agnostic. What scales is the infrastructure, not the choice of AI tool.
2 trillion tokens later, the pilots are in production.
TELUS Digital showcased their AI transformation at Mobile World Congress 2026 in Barcelona. Their $600M benefits guide is available at telusdigital.com.*