Why we built a research library, and how to use it.

Why it exists

Most leaders we work with are under pressure to "do something with AI". What they are given to decide with is a mix of vendor decks, consultancy surveys and a handful of numbers that travel from slide to slide. Some of those numbers are solid. Many are self-reported, sponsored by someone selling the answer, or forecasts dressed as findings.

When you line the evidence up, a consistent picture appears. Use is high, and measured value is thin.

• Most firms say AI helps; few measure it against a baseline. “Over 80% of firms report no impact of AI on employment or productivity over the last 3 years” (Federal Reserve Bank of Atlanta et al., Mar 2026)
• Evaluation is common; production and scale are rare. “25% of respondents have moved 40% or more of AI experiments into production” (Deloitte AI Institute, Jan 2026)
• High performers are more likely to have redesigned workflows and jobs; most firms have not. “84% have not redesigned jobs around AI” (Deloitte AI Institute, Jan 2026)

That gap is where leaders lose money and credibility, and it is the gap our programmes are built around. So we built an open library: 443 sourced facts from 333 organisations, grouped into 15 patterns that hold across several independent sources, and updated every week. Every fact links to the page it came from. Every score is computed from the data, so anyone can check it.

What is inside

Nine ways in, each answering one question a leader actually asks.

How to use it, by the time you have

Three questions, end to end

1. "Our AI pilots work, but they never reach production."

Ask the library and it matches the pattern The pilot-to-production funnel narrows hard, backed by 8 sources. The headline figure: “25% of respondents have moved 40% or more of AI experiments into production” (Deloitte AI Institute, Jan 2026). The guided tour takes you through the rest, from Deloitte AI Institute, McKinsey & Company, Boston Consulting Group and MIT NANDA, with their samples and caveats.

Then open Use cases and filter to your industry: of 452 entries, 98 are named deployments or case studies, so you can see who crossed from pilot to scale and what they measured. Finally, Frameworks shows what the best evidence supports: firms that redesign whole workflows and put a named owner and KPIs on every use case report much higher returns. Most have not: “84% have not redesigned jobs around AI” (Deloitte AI Institute, Jan 2026).

What you do with it. Before funding the next pilot, pick two or three core workflows, redesign the work rather than adding a tool, and name an owner and a KPI for each use case.

2. "Should we use AI agents yet?"

Start with the Agentic tracker. On web research, the best agent went from 9.9% to 90.4% between Apr 2025 and Jul 2026. The length of task an agent can finish grew from about 60 minutes to 16 hours. Those are lab results.

Now read the pattern Agents can do more than firms have deployed: benchmark task success has risen sharply while enterprise deployment of agents stays in single digits by function, and governance and data readiness lag. For example, “21% report a mature governance model for autonomous agents” (Deloitte AI Institute, Jan 2026). The AI Radar then tells you where a written call exists: coding agents and voice agents are marked "Use now", computer-use & browser agents "Pilot". The Workflow Map shows where activity is densest, led by customer support.

What you do with it. Start with a bounded workflow where named deployments exist, keep a person accountable for the output, and put governance in place before you scale, not after.

3. "How do we prove the return on AI?"

Open the pattern Self-reported benefit outruns measured P&L impact, the best-sourced pattern in the library, with 12 sources. “Over 80% of firms report no impact of AI on employment or productivity over the last 3 years” (Federal Reserve Bank of Atlanta et al., Mar 2026). Then Time saved is not money saved: controlled trials and government pilots show real but modest time savings, mostly self-reported, and an experienced-developer trial found a slowdown.

Our Insight “AI is in the survey. It is not yet in the P&L.” walks through the strongest studies on this, and why the gap is a measurement problem before it is a technology problem.

What you do with it. Set a baseline before you roll anything out, measure in money rather than hours saved, give every use case an owner and a KPI, and treat self-reported gains as a claim to test.

How to read it

• Every fact has a source we opened, linked next to it, with its date. If a source does not say something, the field stays blank.
• Scores are computed, not typed. Signals, momentum and counts are calculated from the data each week, so they can be reproduced.
• Judgements are written down. A call such as "Use now" or "Pilot" carries a written reason. Until someone writes one, the item has no call.
• Weak evidence is flagged, not hidden: self-reported, vendor-funded, preliminary or forecast.

What it is not. It is not complete, and it is not market sizing: counts show where we have found sourced activity, and an empty cell means none found yet, not none exists. Benchmarks are lab tests, not proof that something works in your firm. Data on Vietnam and much of Southeast Asia is still thin and often vendor-sponsored; we say so where it applies. The full rules are in our method.

How our programmes use it

Civico programmes put senior leaders on a real AI problem from their own company. The library is the evidence base they test it against. In the AI Academy and the Venture Residential, participants check their assumptions against the patterns, look at who has already done something similar, and leave with a decision they can defend to their board. Their questions, and yours, set what we research next.

You can shape it too: vote for the next research question, ask the library anything, or get one short Insight a week drawn from it.