In the current excitement over agentic AI, there’s a strong temptation to build fully autonomous solutions immediately. But this autonomy-first approach is difficult to get right and often introduces risk. It delays value delivery until everything works (if it ever does). It can be rigid—when full autonomy isn’t appropriate, your only option is to turn the AI off entirely. And full autonomy may not even be the optimal operating point.
Instead, I recommend taking an incremental, capability-centric AI approach I call the 4A Methodology: staged introduction of AI capabilities (whether packaged as agents or not) over time—from Analysis to Assistance to Augmentation to (in some cases) Autonomy.
4A is based on building human-AI partnerships that help people develop an intuitive understanding of AI’s strengths and limitations. It also helps manage AI risk, limiting what AI is asked and allowed to do until its performance is proven at each stage. At the same time, active supervision provides feedback and data to improve AI performance.
Start with Analysis, using AI to help people understand and discover information. In proposal writing, AI might mine knowledge bases for past performance and related solutions from previous bids.
Next comes Assistance, automating repetitive or uninteresting work. In disability claims, AI might read a claimant’s records, identifying what’s most important and what can be ignored.
Next is Augmentation—helping people do things they couldn’t do without AI. In cyber, AI might identify vulnerabilities in a firewall configuration that are too subtle for humans to reliably recognize.
And finally, where it makes sense, full Autonomy. This is what people often try to jump to with agentic AI. Autonomy can be very powerful, but it’s also hard to build and assure where AI is responsible for consequential decisions. I might be comfortable with an agent revising a working, version-controlled proposal draft. But there’s no way I’m going to let an agent decide to submit proposals, control my firewall, or make decisions on a disability claim without a human in the loop.
Importantly, this isn’t a maturity model where every solution ultimately progresses to Autonomy. In the disability claims example, if Assistance can safely eliminate more than 90% of a doctor’s workload—as it did in one application I know well—it may be time to count your winnings. Further automation may not justify the additional cost and risk.
And even where Autonomy does make sense, the earlier capabilities don’t go away. Different situations call for different levels of human involvement. The goal isn’t maximum autonomy—it’s finding the right human-AI operating point for the task and context.
There’s a lot packed into 4A. In follow-on posts, I’ll unpack how it helps build effective human-AI partnerships, manage risk, improve solutions, and find the operating point that optimizes value.