THE VERY GOOD GUYS / RESEARCH
Can AI answer the question, complete the task and pass the work to the next person or tool? These papers examine real projects and saved tests to find out where things worked and where they broke.
These papers look back at saved results, project files and recorded work. Each explains the evidence, findings and limits. They have not been peer reviewed.
An AI tool can say the job is done, yet give back an answer the next tool cannot use.
Saved records116 distinct company domains in the retained artifact.
An AI assistant can sound sure even when records are missing or it reads your numbers the wrong way.
Reported agreement exampleAn operator compared an assistant result with manual work; this is not an accuracy distribution.
An AI video can look right at the start and end, yet change the room or building in the middle.
Shots in the saved reviewA selected production sample, not a representative benchmark.
Creating a book file is only one step. Someone still needs to check it, approve it and make sure the printer can use it.
Attempts exhaustedThe dated status describes repeated polling of a terminal failed export.
A guide can read well but leave out steps. The real test is whether someone else can use it to do the job.
Declared working sessionsInputs named by the source manual, alongside existing operating documents.
OUR METHOD
You should be able to see why we reached a conclusion. For new tests, we define a good result and what we will compare it with before seeing the answers.
Choose the business question, the information we can use and what a good result looks like. Name the person who will make the decision.
Save what went in, what came out, what failed and what a person had to fix. Record repeated attempts so they do not look like separate successes.
Can the next tool read it? Are the facts supported? Did the action finish? How much work did a person still do? Check each question separately.
Explain what we directly checked, what someone reported and what we think it means. Include failures and the next test that could change our conclusion.
The current papers use the records available from earlier work. Those projects did not necessarily follow every step of this method. Private source materials remain private.
THE QUESTIONS WE FOLLOW
Can recorded knowledge become instructions another person can reliably use?
We study the distance between knowing a business and documenting a process someone else can execute. The work includes interviews, operating manuals, role handoffs and explicit records of unresolved decisions. The meaningful test is whether a person can complete the work, including its exceptions.
Does the information used to find and serve customers support the decisions that follow?
We examine prospect classification, call intelligence, qualification and customer communication. A useful system needs clear definitions, supported evidence and an observable next step. We distinguish a successful tool call from an accepted answer, and an identified opportunity from a completed customer outcome.
What must be true before an assistant can produce a dependable business result?
We investigate data coverage, business definitions, permissions, delivery and recovery across operational systems. The work connects conversational assistants with deterministic controls. Each capability needs a bounded acceptance test, including what happens when a source is unavailable, a request repeats or an action remains unconfirmed.
Is the available information sufficient for the decision being made?
We study incomplete records, uncertain proxies and the assumptions inside business metrics. Methods include source inventories, explicit fallback rules and sensitivity analysis. The goal is to show what a dataset supports, what it leaves unresolved and which additional observation would make the decision stronger.
How does a generated asset become an accepted, deliverable product?
We examine consistency across generated video, personalized publishing and production workflows. Attractive previews are one stage. Reference fidelity, revision control, technical checks, approval and fulfillment each need their own evidence. We retain failure cases to understand where generation ends and dependable production begins.
Can people understand, correct and use the system in the work that matters?
We treat adoption as part of system performance. Training, decision authority and the customer experience influence whether a technically functioning tool is useful. Our research questions include who reviews an output, how disagreements are resolved and whether the workflow fits the people expected to use it.
A QUESTION IN YOUR BUSINESS?
Bring us a task, a set of numbers or a question. We can run a small test and show you the results, likely costs and a clear next step.
Find your first projectSee how we work together