AI GUIDEA tool is not a business case
An agency hears that AI can improve sales, delivery and support. Buying a platform before selecting a task makes it hard to tell whether anything improved. Start with work that already exists: who performs it, what they use and what a satisfactory result looks like.
Consider three fictional candidates: drafting a weekly internal report, automatically promising client deadlines and helping search approved project documentation. All involve language, but they have different consequences and evidence needs. The first may tolerate review; the second creates a commitment; the third depends on source access and currentness.
Google's Generative AI Leader programme includes business strategy and techniques to improve outputs. AWS's AI Practitioner guide includes selecting suitable applications and responsible use. We use those public outlines to check our coverage, not as evidence that any particular agency will benefit.
Compare candidates on the work
For each candidate, write the frequency, current effort, available source material, cost of a wrong answer and required reviewer. Do not assign impressive savings before measuring the existing process.
For the internal report, the input might be approved task summaries, the output a one-page draft and the reviewer a project manager. Missing status must stay unknown. For deadline promises, a responsible person must still approve commitments. For documentation search, permissions must follow the user and outdated documents must remain identifiable.
A good first pilot has a clear output, accessible non-sensitive test material and a person who can judge whether it helped. That is a practical selection rule, not a claim that reports are universally the best first AI project.
Measure a small pilot
Create a fictional trial with five report inputs. First produce the reports using the existing method. Then try AI assistance with the same information. Record preparation, review and correction time separately, along with important omissions or invented facts.
Suppose manual preparation took 30 minutes. The AI-assisted draft took 5 minutes to request and 20 minutes to review and fix. Total effort is 25 minutes, not 5. If one draft invents a delivery date, the time comparison alone cannot approve the process. These numbers illustrate the calculation; they are not measured results.
Also count tool subscriptions or API usage, setup, access administration and maintenance. Keep the trial narrow enough that a failed approach can be stopped without disrupting client work. Agree in advance who decides to continue, revise or stop.
Write a one-page project brief
Exercise: choose one of the three tasks. Write its user, input, expected output, forbidden actions, reviewer, test cases and stopping condition. Then decide whether you need a chat, a source notebook, a fixed workflow or an agent. Use CA01 for the architecture distinction.
A reasonable report pilot might produce unsent drafts from approved notes, require review of every date and budget, and stop for redesign if invented facts cannot be controlled. A documentation pilot might first test retrieval on a small approved collection before connecting the whole company archive.
The outcome of the exercise is a decision you can assess, not a purchase list. Expand only after observing useful results on the actual work. Cloud certifications offer provider-specific depth; our Academy's common concepts should remain useful even if you later choose another provider.
Sources
Read the original sources to explore the topic and check specific claims.