AI’s Search-Bar Problem
Twenty-five years of search taught us to ask. Agents need to be given the job.
Most teams still use AI like a search bar. This paper shows what changes when you hand agents whole jobs and give them the official sources to do them, and the five steps leaders can take this quarter.
We ran 581 legal research requests and 845 agent tasks, and pulled in research from McKinsey, Gartner and NBER.
Headline numbers
49%
of consumer ChatGPT messages ask for information or advice.
40%
or more of agentic AI projects will be canceled by the end of 2027, Gartner predicts, over rising costs and unclear value.
+21%
higher answer quality when agents could reach the right sources. Same model, same prompts.
The model isn't the bottleneck anymore. What the agent can reach is.
What's inside this research white paper, in four parts.
- [ Part 01 · The gap ]
Most companies use AI. Few see it on the bottom line.
~90%of function-specific AI use cases are still stuck in pilot.
Adoption isn't the problem. We look at why so much AI work stalls, and what the teams seeing results do differently.
- [ Part 02 · The shift ]
Half of what people send to AI is still a question.
Search taught us to ask and take the top answer. The paper shows what changes when you hand an agent the whole job instead.
- [ Part 03 · The evidence ]
Reach the official source and the right answer comes first.
Same model, same prompts, different reach. Our benchmarks show how much the sources an agent can read change the work it hands back.
- [ Part 04 · The playbook ]
Five moves leaders can make this quarter.
Start with one team and one job. The paper shows how to measure the shift as you go.