From the State of AI authors: How we reimagined our flagship report with AI
With the recent launch of the 2026 State of AI in the Enterprise: The untapped edge, we didn’t just publish another report—we used AI to reimagine how the report is developed, designed, and experienced. The report represents months of intentional experimentation with AI across the full life cycle, from hypothesis development through publication and marketing, to amplify the capabilities and expertise of our teams.
Our intent was simple: Apply the same AI transformation principles we use in our client work to our own flagship work, then share what we learned. Here’s a behind-the-scenes look at how it went.
Why State of AI was ready for a fresh take
State of AI is one of Deloitte’s most complex recurring research initiatives: It’s a comprehensive survey with multiple global stakeholder groups, parallel report and marketing workstreams, and a fixed external launch deadline. That combination of complexity, repetition, and time pressure made it an ideal candidate to test where AI could actually change outcomes versus simply speeding up tasks.
A high-visibility report demands defensible insights and a compelling narrative. So we set a clear operating principle: human-led, AI-accelerated. AI can accelerate drafting and iteration, but we remain accountable for strategic intent, interpretation, real-world context, and validation.
With leadership support, we treated this as a pilot year, with the freedom to test tools and rethink prior ways of working while maintaining the quality that State of AI readers have come to expect.
Where we applied AI
We looked across the full life cycle—from survey development to the report site experience—and asked: How can we push the boundaries with AI? We used Deloitte’s internal AI platform, Sidekick along with other enterprise-grade large language models (LLMs) and tools, ensuring all outputs were reviewed and validated for accuracy.
Across the workflow, AI helped us move faster on high-friction tasks, including project planning and status reporting; early-theme brainstorming and survey question drafting; theme extraction from survey results; market research and storyline iteration; style/tone and grammar checks; and first drafts of collateral and marketing content.
What “human-led, AI-accelerated” looked like for State of AI
“Human-led, AI-accelerated” isn’t just a catchy phrase—it’s an operating model that requires accountability and defining decision-owners versus draft generators. During our work on State of AI this year, we continued to learn where human expertise truly excels, where AI hits its limits or creates more noise, and what the optimal mix actually looks like.
- Humans lead (decide, validate, sign off): direction and themes, final survey questions, interpretation and contextualization of insights, draft report, narrative validation with credible sources, and brand/risk compliance approvals.
- AI accelerates (suggest, draft, compare): initial hypotheses, pattern surfacing, report improvement suggestions, comparisons to related reports, and draft copy for collateral/comms.
Reimagining the experience: The author avatar/voice assistant pilot
To truly innovate, we also piloted a first-of-its-kind AI avatar and voice assistant to help readers navigate the 2026 report experience.
At a high level, we built a cloud-hosted, API-driven conversational system designed for broad public use, grounded via retrieval against risk-approved report materials so responses stay aligned to approved research and messaging. It was deployed with cloud-native infrastructure for scalability and oversight, with third-party vendors handling real-time avatar streaming and high-quality speech synthesis.
Just as important: Trust and safety were designed in, not bolted on. We implemented multilayer guardrails (intent guardrails, topical rails, content filters, fallback safety nets) and structured prompting to keep answers grounded, executive appropriate, concise, and resistant to off-topic drift.
Lessons we’re taking forward and sharing with our clients
Because we treated this as a pilot year, we optimized for learning while still delivering. This is what we would emphasize to any team attempting to reimagine a high-visibility process:
Start with a blank slate—but align early
Reimagining a flagship initiative requires creativity and collective buy-in: Our avatar/assistant experience moved from “big idea” to production in under six months with strong leadership support to navigate reviews and approvals.
Make room to experiment without breaking delivery.
Building experimentation into timelines allowed us to test more than eight LLMs and AI tools without overpromising. We evaluated each with accuracy and quality as the core criteria.
Design workflows around human judgment and use AI for speed.
Subject-matter experts and account leaders provided interpretation and context AI can’t replicate—and 100% of AI-suggested insights were materially reshaped by expert input.
Engineer trust.
Guardrails, approved sources, testing, and escalation paths are non-negotiable, especially for public experiences. We tested guardrails against more than 400 prompts.
Build AI fluency across the team.
We found AI improved not just speed, but judgment—helping teams recognize when AI adds value versus when it creates noise.
Standardize what works; stay flexible on tools.
We documented workflows, prompts, tools, and tests—plus what didn’t work—so we can automate proven patterns while continuing to evaluate new use cases.
Opens in new window
This was an experimental year by design. The outcome was not just a report, but a set of repeatable patterns underscoring where AI reliably improves speed and exploration and where human expertise and judgment remain indispensable. Our focus now is scaling what proved valuable, automating what’s repeatable, continuing to explore new opportunities as capabilities evolve, and extending these lessons to our teams and clients.
Beena Ammanath
