Free tool
Agent Readiness Audit 18 questions across the 6 stages of building an agent product. It scores each stage and tells you which to fix first, with the specific reading for each.
No account, no email, nothing stored. Answers stay in your browser and disappear when you reload.
01
Fundamentals Can you state the job your agent does in one sentence, without the words AI, agent, or automation? If the sentence needs those words, you are selling a technology rather than a finished job.
In place Started Not yet
Do you know what that job costs today, in hours, headcount, or outsourced spend? That number is the budget you are displacing, and it is almost always larger than the price you were about to charge.
In place Started Not yet
Have three or more prospects described the same problem in the same words? One enthusiastic conversation is a coincidence. The same sentence three times is a product.
In place Started Not yet
02
Building Do you have a set of real failure cases you run before shipping a prompt or model change? Without it, every change is a guess and every regression is discovered by a customer.
In place Started Not yet
Is there a hard ceiling on steps and cost per task, with defined behaviour when it is hit? An agent without a ceiling can retry a broken call indefinitely and bill you for it.
In place Started Not yet
Can you ship a prompt change in minutes, without a full rebuild and redeploy? Slow iteration forces you to batch changes, which is how a regression becomes impossible to attribute.
In place Started Not yet
03
Architecture Does every run produce a trace with the resolved prompt, each tool call, and timings? Debugging an agent without a trace is guesswork, and the resolved prompt is the part people forget to store.
In place Started Not yet
Are tool arguments constrained with enums, required fields, and explicit formats rather than free strings? Every loose field is an invitation for the model to guess, and it will guess plausibly and wrongly.
In place Started Not yet
Is there an escalation path with a named owner, a timeout, and a tested return route? Handoff is where agents meet reality, and resumption is the part that usually breaks.
In place Started Not yet
04
Operations Do you know your cost per completed task, rather than per call or per token? A cheap-looking per-call figure hides a loop that takes eleven calls to finish one job.
In place Started Not yet
Is there a written answer to where customer data goes, who processes it, and how long you keep it? Every business buyer asks this, and improvising the answer mid-deal costs weeks.
In place Started Not yet
If the agent acted wrongly at scale, could you identify every affected record? Agents fail in a way ordinary software does not: correctly executed, wrong at volume.
In place Started Not yet
05
Go-to-Market Can you name the specific role whose performance is measured by this workflow? That person is the budget holder. An engineering champion opens the door; they sign.
In place Started Not yet
Do you have one reference customer with a number attached to the outcome? In a narrow market a named peer with a measured result outperforms any demo.
In place Started Not yet
Do you know what a buyer's security review will ask, with answers already written? In enterprise the security review is the deal, and it arrives after the champion is convinced.
In place Started Not yet
06
Monetization Does your price attach to a metric the buyer already tracks? A buyer who budgets per seat cannot map tokens to their P&L, however fair the rate is.
In place Started Not yet
Do you know your gross margin per completed task? Inference at twenty percent of revenue is a software business. At sixty it is something else.
In place Started Not yet
Is expansion priced in advance, so more volume is a normal conversation rather than a negotiation? Unpriced expansion turns every increase into a favour the customer expects for free.
In place Started Not yet