The question nobody asks the demo
A demo shows you the best version of a tool on a clean account with a friendly dataset. It almost never answers the one question that decides whether you can put it into production: what does it do, by itself, when no one is in the room?
That single question sorts the entire category. A tool that only ever suggests is safe and slow. A tool that acts on its own is fast and, without the right controls, dangerous. Between those poles sit two more useful options. Knowing which one you are buying, and which one you actually need, saves you from both the tool that creates more work and the tool that creates a mess you cannot trace.
The four execution depths
Picture a line from "tells you what to do" to "does it for you." There are four meaningful stops on it.
- 1 · Recommend. The system analyses and suggests. A human reads the recommendation and does the work. Low risk, full control, and no leverage. You still spend the hours; you just spend them better informed.
- 2 · Execute rules. You hand-build conditions and the system runs them. "If cost per acquisition exceeds target for three days, pause the ad group." Fast and predictable, but only as smart as the rules you wrote, and brittle when reality drifts outside them.
- 3 · Autonomous. Given a goal, the system decides and acts without a rule for every case. The most leverage, and the most exposure. If you cannot see what it changed or reverse it, you have handed your spend to a black box.
- 4 · Bounded autonomy. The system acts on its own, but only inside limits you define, and it logs every action with its trigger, the change made, and the measured result. It reverses cleanly. This is the depth that gives you the leverage of autonomy without the loss of control.
Most of the market clusters at depth 1 (dashboards and "insights") and depth 2 (rules engines). A growing set claims depth 3. Very few are honestly at depth 4, because depth 4 is harder to build: it requires the guardrails, the logging, and the reversibility, not just the model.
Why depth is the real buying decision
Two tools can have identical feature lists and sit at different depths. One "optimises budget" by emailing you a recommendation. Another "optimises budget" by moving the money itself. Same words, completely different commitment of trust. The feature comparison hides the difference; the depth reveals it.
Depth also predicts the failure mode you will live with:
- At depth 1, the failure is that nothing changes. The insight arrives, the week gets busy, and the recommendation is never actioned. You paid for analysis you did not use.
- At depth 2, the failure is silent drift. The rule keeps firing on a situation it was not written for, and no one notices until the report looks wrong.
- At depth 3, the failure is the one that ends up in a post-mortem: a change you did not authorise, at a scale you did not expect, that you cannot fully explain or roll back.
Depth 4 is designed to remove all three. It acts (so something changes), it reasons rather than blindly repeats (so drift is caught), and it is bounded and logged (so nothing happens that you cannot see or reverse).
Bounded autonomy, concretely
"Acts within guardrails" sounds like a slogan until you see the mechanics. In practice it is four things wired together: a trigger (a defined condition the system watches for), a guardrail check (hard limits like spend caps, brand exclusions, and approval thresholds that must pass before anything runs), an action (the actual change, made through a real connection to your account), and an impact log (a record of what changed and what it moved). Then the loop closes: every result is fed back to sharpen the next decision.
That is the Trigger / Action / Impact framework, and it is what separates a system you can run in production from a demo. The control layer underneath it is guardrail-driven automation: the limits and approval gates that constrain every action, with each one logged and reversible. Guardrails are not a brake on autonomy. They are the thing that makes autonomy safe enough to deploy at all.
How to place a tool, and yourself, on the spectrum
When you next sit through a pitch, skip the feature grid and ask four questions:
- Does it act, or only recommend? If it acts, how do I set the limits it acts within?
- Can I see every action it took, with the reason and the result, in a log I can export?
- Can I reverse any single change, and what is the largest change it can make before a human has to approve it?
- When it is wrong, how will I know, and how fast?
The answers place the tool on the spectrum honestly. They also tell you something about yourself: a team without clean data, defined goals, or an owner for automation is not ready for depth 3 or 4 no matter how good the tool is. Autonomy applied to a broken foundation just scales the mess faster. The right depth for you depends as much on your readiness as on the software.
See which execution depth fits your team
The free Readiness Score checks whether your data, processes, and governance are ready for automation that acts, not just recommends. Four minutes, no login, and you get your tier and the one constraint to fix first.
Get your free Readiness Score →Keep reading
- How to evaluate an AI marketing automation platform: 9 criteria that actually matter
- How to add AI to your martech stack without rebuilding it
- The Trigger / Action / Impact framework — the accountability layer underneath bounded autonomy