AI / Wanderings 2026
Loops
AI creates lasting value when action, evidence and adjustment form a measurable loop. Twenty examples show how to build one without losing control of the outcome.
AI systems create lasting value through feedback loops. They act, observe the result and adjust before acting again.
In my current humAIne work, I use this pattern with Claude, automation scripts and business processes. The useful question is how to design a loop that improves the work without losing control of the outcome.
The three parts of a useful loop
A loop contains three things:
- An action. The AI does something.
- A signal. Something in the world tells it how that action went.
- An adjustment. The system changes its behaviour based on the signal.
Action, signal and adjustment repeat.
Your thermostat does this. Your body does this. Every living organism on the planet runs feedback loops to survive. Software can run those loops thousands of times per second across millions of data points without getting tired, bored or distracted.
Many AI projects build the action and neglect the signal or adjustment. They produce output without a dependable way to learn from the result.
The reinforcement learning intuition
Reinforcement learning gives us a useful intuition for this pattern.
A dog learns to sit because it gets a treat. A child learns to avoid touching the stove because it hurts. An RL agent learns to play chess because a mathematical reward favours winning over losing.
The formal version: an agent exists in an environment, takes actions, receives rewards, and updates its policy, its strategy for choosing actions, to maximize future rewards.
Practitioners can use the principle without building a full RL system. Define what “good” looks like, let the system try, measure the result and adjust.
I use this pattern with Claude, automation scripts and business processes, rather than neural networks and reward functions. The implementation is secondary to a loop that produces an honest signal.
How to design a loop that works
Start with the goal, then work backwards to the model.
Step 1: Define the outcome in measurable terms
"Make our emails better" is not a goal. "Increase reply rate from 4% to 12%" is a goal.
"Improve customer support" is not a goal. "Resolve 80% of tickets without human escalation while keeping CSAT above 4.2" is a goal.
If you cannot measure it, you cannot loop on it.
I have sat in rooms with smart people who wanted to “use AI to improve their marketing” and could not say what improvement meant. Without a metric, the system has no dependable signal, the loop cannot learn and the claimed advantage remains a slide.
Step 2: Design the signal
The signal tells your system whether it is getting closer to the goal or farther away, and weak signals ruin otherwise promising projects.
Good signals are:
- Fast. If it takes three months to know whether the action worked, your loop is too slow to learn anything useful.
- Honest. Vanity metrics kill loops. If you optimise for clicks instead of conversions, you'll get clickbait. Congratulations.
- Specific. "Users liked it" tells the system nothing. "Users who saw version B spent 3.2x longer on the page and converted at 1.8x the rate" tells it exactly what to do next.
The best signal is often already being collected: server logs, support tickets, user behaviour data or sales-conversion timestamps. Connect it to the next adjustment.
Step 3: Close the loop
The signal has to flow back into the system and change its behaviour. This is the "learning" part. Without it, you have monitoring.
In a technical RL system, this is the policy update. In a practical business system, this can be as simple as: "Every Monday, the AI reviews last week's results, identifies the top and bottom performers, and adjusts its strategy for the coming week."
The mechanism matters less than whether it is automatic, consistent and able to change what the system does next.
Step 4: Set guardrails
Loops without constraints optimise for the metric and destroy everything else. This is Goodhart's Law: when a measure becomes a target, it ceases to be a good measure.
Your email loop might discover that sending 47 emails per day maximises reply rate, until recipients mark you as spam. A support bot might close tickets quickly and improve the resolution number while customers leave because their problems remain unsolved.
Every loop needs boundaries. Maximum frequency. Quality thresholds. Human review triggers. Ethical constraints. Build them in from day one, not after something breaks.
The compounding effect
Loops matter more than models because they compound.
A better model can deliver a one-time improvement; a 10% better model gives you a 10% improvement. A loop can compound its gains: a 2% improvement per cycle becomes 2.7x over 50 cycles, 7.2x over 100 and 52x over 200.
The mathematics is relentless. Given enough cycles, teams with mediocre models and strong loops will outperform teams with strong models and no loops.
I tell founders to spend less time debating which LLM to use and more time improving the speed and honesty of the feedback loop. That is where a durable advantage can form.
20 feedback loops in practice
These 20 patterns show how the same structure can be applied across business work. Some are established use cases; others are practical designs to test. Each follows the same pattern: action, signal and adjustment.
1. Cold outreach optimisation
Action: AI generates and sends personalised cold emails. Signal: Reply rates, meeting-booked rates, unsubscribe rates. Adjustment: Every Friday, the system analyses which subject lines, opening hooks, and CTAs performed best. Worst performers get dropped. Best performers get variations. New experiments get added.
The useful test is meetings booked, not opens. Opens are vanity.
2. Content repurposing engine
Action: AI takes a long-form blog post and generates 15 social media posts, LinkedIn, X, threads. Signal: Engagement rate per post (likes, comments, shares, saves), click-throughs to the original post. Adjustment: The system learns which post formats, hooks, and lengths work for each platform. It stops generating what doesn't work and doubles down on what does.
The test is whether the repurposed posts send qualified readers back to the original piece, not whether they produce an impressive pile of drafts.
3. Customer support triage
Action: AI classifies incoming support tickets by urgency, category, and likely resolution path. Routes them accordingly. Signal: Was the routing correct? Did the ticket get resolved? How long did it take? Did the customer come back with the same issue? Adjustment: Weekly retraining on misrouted tickets. The system gets sharper at distinguishing "annoyed but can wait" from "about to churn."
4. Pricing optimisation
Action: AI adjusts pricing (or suggests adjustments) based on demand signals, competitor pricing, inventory levels. Signal: Conversion rate, revenue per visitor, margin, customer acquisition cost. Adjustment: Daily micro-adjustments. The system learns price elasticity curves that no human analyst could map manually.
This is what Amazon has been doing for a decade. You can do it too, at smaller scale, with an LLM and a spreadsheet.
5. Meeting preparation
Action: Before every meeting, AI pulls context, previous emails, CRM notes, LinkedIn activity, recent company news, and generates a one-page brief. Signal: After the meeting, you rate the brief: "useful / partially useful / missed the point." Add notes on what was missing. Adjustment: The system learns what context you find valuable. After 20 meetings, it knows you care about recent funding rounds and don't care about org charts. After 50, the briefs are surgical.
6. Recruitment screening
Action: AI reviews incoming applications. Scores candidates against role requirements. Flags top matches. Signal: Which AI-flagged candidates made it past the human interview? Which ones got hired? Which ones survived 6 months? Adjustment: The scoring model gets recalibrated quarterly against actual hire outcomes. The system learns what your company values, not what the job description says.
7. Ad copy generation
Action: AI generates 20 ad copy variants per campaign. Signal: Click-through rate, cost per acquisition, ROAS. Adjustment: Kill the bottom 50% every 48 hours. Generate new variants inspired by the top performers. The system converges on what your specific audience responds to.
Performance marketers have been doing this manually for years. The loop runs faster when AI writes the variants.
8. Code review assistant
Action: AI reviews pull requests. Flags potential bugs, style violations, security issues, and suggests improvements. Signal: Did the developer accept or reject the suggestion? Was the flagged issue a real bug? Adjustment: The system learns your team's coding style, your actual bug patterns, and which suggestions are noise. Track the false-positive rate. If it does not fall, the loop is producing noise rather than learning.
9. Inventory forecasting
Action: AI predicts demand for each SKU for the next 30/60/90 days. Signal: Actual sales versus predicted sales. Stockout events. Overstock write-downs. Adjustment: Model retrains monthly on actuals. Incorporates new signals, weather, events, social media trends. Each cycle reduces forecast error.
10. Personalized learning paths
Action: AI recommends the next lesson, exercise, or resource for each student. Signal: Quiz scores, time spent, completion rates, self-reported difficulty. Adjustment: The system learns each student's pace, strengths, and gaps. It stops recommending material that's too easy or too hard. The learning curve flattens.
This points towards AI personalising the learning path while teachers concentrate on the human work.
11. Sales call coaching
Action: AI transcribes sales calls and scores them against a framework, did the rep qualify properly? Handle objections? Ask for the close? Signal: Call scores correlated with deal outcomes. Did scored-high calls close more? Adjustment: The scoring framework evolves based on what predicts success for your product, your market, your buyers. Generic playbooks become specific playbooks.
12. Newsletter subject line optimisation
Action: AI generates 5 subject line options for each newsletter send. Signal: Open rate, click-through rate, unsubscribe rate per subject line. Adjustment: Pattern matching across 50+ sends. The system builds a model of what your specific audience opens. "Numbers in subject lines" might work for your tech audience. "Questions" might work for your executive audience. You don't guess. The loop tells you.
13. Bug prediction
Action: AI analyses code commits and flags which ones are likely to introduce bugs, based on code complexity, change velocity, author history, and file hotspots. Signal: Did the flagged commit produce a bug within 30 days? Adjustment: Model refines its risk heuristics. Learns that changes to the billing module on Fridays are 3x more likely to break. Starts flagging earlier.
14. Document drafting
Action: AI drafts contracts, proposals, or reports based on templates and context. Signal: How many edits did the human make? Which sections got rewritten? Which sections were accepted as-is? Adjustment: The system learns your voice, your legal preferences, your formatting standards. Edit distance shrinks with every document. After 30 iterations, the drafts need almost no human editing.
15. Supply chain risk monitoring
Action: AI continuously scans news, shipping data, weather forecasts, and supplier financials. Flags potential disruptions. Signal: Was the flagged risk real? Did it affect supply? Was the lead time sufficient to respond? Adjustment: The system learns which signals are predictive and which are noise. Geopolitical instability in a supplier's region? Signal. CEO change at a tier-3 supplier? Probably noise. The alert quality improves each quarter.
16. Product recommendation
Action: AI recommends products to users based on browsing history, purchase history, and similar-user behaviour. Signal: Click-through rate, add-to-cart rate, purchase rate, return rate. Adjustment: Continuous. Every user interaction refines the model. The system learns that "people who buy running shoes also buy foam rollers" and that "people who buy running shoes do NOT want to be recommended running shoes again for 6 months."
17. Internal knowledge base curation
Action: AI answers employee questions using the company knowledge base. When it can't answer, it flags the gap. Signal: Was the answer helpful? (Thumbs up/down.) Did the employee still need to ask a human? Adjustment: Gaps get filled. Bad articles get flagged for rewriting. The AI learns which sources are reliable and which are outdated. The knowledge base gets better because people use it, not because someone maintains it.
18. Fraud detection
Action: AI scores transactions in real-time. Flags anomalies for review. Signal: Was the flagged transaction fraudulent? Was a non-flagged transaction later confirmed as fraud? Adjustment: Both false positives and false negatives feed back into the model. The system gets better at distinguishing "unusual but legitimate" from "unusual and fraudulent." The curve tightens.
19. Portfolio rebalancing
Action: AI monitors portfolio allocations against target weights. Suggests rebalancing trades when drift exceeds thresholds. Signal: Post-rebalance performance versus benchmark. Tax efficiency of trades. Transaction costs. Adjustment: The system learns optimal drift thresholds for each asset class. Learns when rebalancing destroys more value in taxes than it adds in alignment. Learns that rebalancing emerging markets monthly is wasteful but rebalancing bonds quarterly is negligent.
20. Customer churn prediction
Action: AI scores every customer on churn risk weekly. Triggers retention workflows for high-risk accounts. Signal: Did high-risk customers churn? Did the retention workflow save them? Which intervention worked? Adjustment: The risk model gets sharper. The retention playbook gets refined. The system learns that "customer hasn't logged in for 14 days" is a weak signal but "customer downgraded their plan and opened 3 support tickets in a week" is a strong one.
The meta-loop
The best teams also run loops on their loops.
They ask: is this loop learning fast enough? Is the signal still reliable? Has the environment changed enough that our adjustment mechanism is optimising for the wrong thing?
That is the meta-loop: a system for checking whether the other loops still work. In practice, review each learning curve quarterly. A flat curve can mean the problem is solved or the loop is broken; examine the evidence before deciding which.
Compounding requires patience
A loop may show useful changes on day 30, day 90 and day 180 rather than on its first day. By day 180, it can be far ahead of a static system.
Companies often want to plug in AI and see magic on Monday morning. Compounding follows a slower timetable. The companies that gain from it invest in the loop, measure against evidence and allow enough cycles for the evidence to accumulate. Investing and fitness follow the same principle.
Build the loop, set the review dates and give it time to run.
Getting started today
Pick one weekly process with a measurable outcome, such as email campaigns, sales calls, customer onboarding or code deployments.
Ask three questions:
- What does "good" look like, in a number?
- What signal tells me whether this week was closer to good or further away?
- How can the system automatically adjust based on that signal?
Connect those three parts, run the loop for 8 weeks and measure the change against the starting point. Loops have always compounded. AI changes how quickly we can run them.
Martin Uetz is the founder of humAIne. He writes about humans, technology and business from Switzerland and Iceland, usually with too much coffee and too little sleep.