October 1, 2026
Research in practice
Test the idea before your team relies on it
A useful AI system has to survive contact with the business around it. A lead needs enough evidence for someone to act. An operating answer needs the right records and the right definition. A generated product needs to pass review and reach the person it was made for.
That is the work The Very Good Guys has been exploring across operations, customer communication, knowledge and production. The retained artifacts and recorded discussions contain prototypes, delivered materials, failures and unresolved questions. Together, they give us a starting point for applied research.
We use that work to guide our research and the systems we build with teams. We test a specific question, build the system around a useful capability and examine what happens in use. The business provides the constraints: available information, people, tools, cost, timing and the consequences of an error.
The first research collection looks back at existing work. It includes a classification batch where a success flag concealed inconsistent output structure, operational assistants whose answers depended on source coverage and authority, and production workflows where a generated asset was only one step toward a finished result. Each paper states what the evidence can support.
Looking forward, the research method needs to be explicit before an experiment begins. Define the task and comparison. Freeze the inputs and acceptance rule. Keep failures and human interventions in the record. Measure review and recovery as part of the work. Use the result to decide what to build, revise or stop.
This approach keeps our operating perspective intact. Marketing, design, sales, service and delivery remain connected parts of the same business. AI changes the possibilities within that system. Our job is to investigate those possibilities carefully and turn the useful ones into working capability.
A business can hire us to find the right first step, build it with the team or keep it working. It begins with a practical question: what could work better, and what evidence would help us make the next decision?
AN EXAMPLE FROM THE WORK
A success flag is one observation.
Our output-contract paper examines a saved company-research batch with 116 records. The system marked 112 successful, but only 28 placed the requested target field directly in the expected object. The other 84 contained it deeper in their structure. A person might find that answer; a downstream tool might not. This is a structural finding, not a factual-accuracy score.
- Name the next consumer of the answer, such as a CRM field or a reviewer.
- Write the acceptance rule before changing the system.
- Count the work required to validate, normalize or review the returned answer.
The useful question becomes specific: what must be true for this output to move safely into the next step?
Inspect the records analysis and synthetic exampleDiscuss a similar workflow