AI readiness: the questions we ask before recommending anything
Before any AI recommendation, we ask the same unglamorous questions about workflows, data, and approval authority. Here they are — use them without us.
When a business asks us where AI fits, the useful work starts with questions, not tools. These are the actual questions from our readiness assessments. You can run most of this exercise without hiring anyone.
1. Which jobs are actually repetitive?
Not "what could AI do" — what do your people do every week that follows the same shape? Drafting the same category of document, triaging the same inbox, reconciling the same reports. Write down ten. The automation candidates are in that list, not in a vendor's demo.
2. Where does the input data live, and is it any good?
Every agent pipeline we run sits on top of data plumbing: a nightly sync, a scraper, an export. If the inputs for a job live in someone's head or a stack of PDFs nobody has parsed, the honest first project is data plumbing, not AI. That project is cheaper and pays off regardless.
3. Who approves the output?
For every candidate job: if this gets done wrong, who notices, and how fast? Jobs with a natural reviewer — an editor, an admin, a dispatcher — automate safely, because you can put the agent's draft in front of that person. Jobs where errors flow silently downstream need gates designed first. In our own systems, nothing public ships without a named approver, and we consider that a feature of the design, not a limitation.
4. What does the failure mode cost?
A bad blog draft costs an editor ten minutes. A bad customer refund costs money and trust. Rank your candidates by cost-of-error, and automate from the cheap end. You will learn how the machinery behaves where mistakes are affordable.
5. Would a cron job do?
Our favourite question. If the job is deterministic — same inputs, same steps, same outputs — you do not need a language model. You need a script and a schedule. A surprising amount of "AI opportunity" dissolves into ordinary automation that is faster, cheaper, and easier to audit. We say this as a company that sells AI work.
What comes out the other end
A ranked list: a few jobs worth piloting with an agent and an approval gate, a few that need data plumbing first, and a few that were cron jobs all along. One pilot, chosen because it can prove value in weeks with a measurement defined before anything is built. That's the whole method. If you want a second pair of eyes on your list, write to us.