---
title: Every task can be AI-supported · Falk Uebernickel
url: https://uebernickel.com/writing/every-task-can-be-ai-supported
author: Falk Uebernickel
description: Almost every task can now be AI-supported. A matrix of generation effort and verification effort shows how, and how much human oversight each form of support needs: Autopilot, Sign-Off, Proof Trail, Co-Creation, and Human only by choice.
---

![Painting in greyscale: a figure on a gallery looks down into a vast hall where looms work on their own in shafts of light.](https://uebernickel.com/img/art/every-task-can-be-ai-supported-1200.jpg)

# Every task can be AI-supported. The question is how humans stay in charge.

Oct 2026  •  Written by Falk Uebernickel  •  [Discuss on LinkedIn](https://www.linkedin.com/posts/falkue_aigovernance-itstrategy-humanoversight-activity-7514683548440788995-HBTM)

Human thinking, AI-assisted writing: the ideas and the argument are mine; the text was written together with AI.

Listen · 13 min · read by an AI voice12:54

*Working note, general edition.* Debates about AI at work often ask *whether* a task should be touched by AI at all. This note takes a different position: there is hardly any area of knowledge work or operations left that AI cannot support. What differs between tasks is *how* AI supports them, and how much human oversight that form of support contains. Building on the asymmetry between generating and verifying an output (Wei, 2025; Cobbe et al., 2021) and on design patterns from recent thesis work (Hötter, 2026), I describe a matrix of generation effort and verification effort. Each quadrant is a form of AI support with its own built-in oversight: Autopilot, Sign-Off, Proof Trail and Co-Creation. What remains outside the matrix, Human only, is not a technical limit but a conscious choice. The examples are illustrative. [PDF version](https://uebernickel.com/papers/Uebernickel_2026_Every-task-can-be-AI-supported.pdf) · [German version (PDF)](https://uebernickel.com/papers/Uebernickel_2026_Jede-Aufgabe-laesst-sich-mit-KI-unterstuetzen_DE.pdf).

## 1 · From whether to how

Many AI frameworks start with risk: the higher the risk, the more human involvement. Leaders draw their lines in a similar way. Strategy, people decisions, client relationships and safety stay with them; reporting, scheduling and routine decisions could be handed over. Lines like this are usually drawn along the question *can AI do this, or not?*

My position is that this question has largely been answered. Almost any task can now be supported by AI, from compiling a supplier dossier to preparing a pricing decision. The useful question is *how* AI should support a given task, and how humans stay in charge while it does. A risk class does not answer this: it says how much human involvement a task needs, not what kind.

Consider two tasks from business-to-business sales with similar stakes. A wrong answer in a security questionnaire can cost the deal. Yet the questionnaire is laborious to compile but quick to check, because every answer points to a policy, a certificate or a contract clause. A return-on-investment claim takes minutes to write, yet whether it holds becomes visible only months after go-live. AI can support both. It should support them in very different ways.

## 2 · Two efforts, four ways AI supports the work

The matrix rests on one observation: for many tasks, checking an answer is far cheaper than producing it. Machine-learning research exploits this asymmetry when a separate verifier ranks generated solutions (Cobbe et al., 2021). Wei (2025) states it as a general rule, the verifier’s law: the ease of training AI to solve a task grows with how easily the task can be verified. For organisations, the asymmetry has a second meaning. If AI makes generation cheap, the form of support is decided by how the result can be checked. The matrix therefore places two efforts on two axes:

- **Generation effort:** how hard is it to produce the output? For example, the expertise required, the number of sources, the amount of context.

- **Verification effort:** how hard is it to check the output? For example, the time to check, the number of roles involved and the cost of correcting an error.

In Hötter (2026), each axis is scored with five binary criteria (0–2 low, 3–5 high). Risk is not discarded. It enters through the cost of correction, as part of verification effort. An action that cannot be undone is, by definition, expensive to verify after the fact.

| Variant | Generation effort | Verification effort | Oversight | Examples |
|---|---|---|---|---|
| A · Autopilot (act & undo) | low | low | after the fact: humans monitor the log and can undo any change | CRM updates, data enrichment |
| B · Sign-Off (draft & approve) | high | low | per output: experts approve each item against its source | security questionnaires, NDAs |
| C · Proof Trail (claim & prove) | low | high | at milestones: claims stay assumptions until evidence confirms them | ROI claims, business cases |
| D · Co-Creation (build & gate) | high | high | continuous: humans and AI iterate, nothing is released without approval | implementation plans, pricing |
| Human only | — | — | a conscious choice, not a technical limit | people decisions, key client relationships |

Drift (hypothesis): as verification gets cheaper, tasks move towards Autopilot.

**A · Autopilot (low generation, low verification).** AI acts; people supervise the log and keep the power to undo. This suits high-volume, reversible work such as updating CRM records, enriching master data or refreshing a delivery status. It fails when the undo button is missing. A customer-facing answer, once given, cannot be reversed: in *Moffatt v. Air Canada* (2024 BCCRT 149), an airline was held liable for what its website chatbot told a customer. Irreversibility raises the verification effort of a task, whatever its generation effort.

**B · Sign-Off (high generation, low verification).** AI compiles at volume; an expert approves each item against its source. Typical cases are the documentation trails that business already runs on: security questionnaires and non-disclosure agreements in sales, compliance evidence from suppliers, regulatory paperwork. Here AI removes most of the drafting work, and the risk shifts to the reviewer. People who approve mostly correct items drift into approving everything (Parasuraman & Riley, 1997; Skitka et al., 1999), and rare errors are missed far more often than common ones (Wolfe et al., 2005). A sign-off that never says no is not a control.

**C · Proof Trail (low generation, high verification).** Claims are cheap to write and slow to confirm: the return of a new tool, a forecast, a footprint reduction. AI can draft them in minutes. The oversight lies in what happens next: each claim stays labelled as an assumption, gets an owner and a date by which evidence is due, and is signed off by the accountable role once the evidence arrives. The failure mode is quiet. Assumptions harden into facts because nobody revisits them.

**D · Co-Creation (high generation, high verification).** Even strategic and design work, such as a pricing model, an implementation plan or a sourcing strategy under volatile prices, can be AI-supported. Here, people and AI build the result together, and explicit gates decide what leaves the room. Field evidence shows why the gates matter. In an experiment with consultants, AI raised quality on tasks inside its capability frontier but lowered correctness on a task just outside it, where consultants trusted it too readily (Dell’Acqua et al., 2023).

**Human only: a conscious choice.** If almost every task can be AI-supported, then the work that stays purely human is no longer a remainder that technology has not yet reached. It is a decision. An organisation may decide that decisions about people, its most important client relationships and the final judgement on quality remain human because the human act itself carries the value: craft, accountability, trust. Such a choice should be named, justified and revisited from time to time, rather than left as an unspoken hesitation.

## 3 · Drift: human-in-the-loop as a transition

The matrix is not static. As verification becomes cheaper, through linked sources, structured outputs, automated tests or rule-based checks, tasks drift towards Autopilot. In software engineering, where every change can be checked against an automated test suite, models resolve a growing share of real issues (Jimenez et al., 2024). I therefore treat per-item approval as a transition technology, as a hypothesis rather than a finding. Its consequence is that the human role changes rather than disappears: it moves from checking each output to defining what counts as correct and checking the checkers, the agenda known as scalable oversight (Bowman et al., 2022).

## 4 · Five implications for leaders

1. **Ask how, not whether.** Start from the assumption that every process can be AI-supported. Before shelving a use case as “too risky”, ask how its output could be checked; the answer selects the variant.

2. **Make Human only a deliberate decision.** Agree, per function, which acts remain human, and write down why. An explicit, reviewed choice protects craft and client relationships better than a general hesitation.

3. **Budget expert time as the bottleneck.** Generation is becoming cheap; verification is not. Every Sign-Off use case consumes the hours of the people whose judgement the organisation depends on.

4. **Measure what the oversight catches.** Track what reviewers change and why, and occasionally plant known errors to see whether they are caught. An approval rate close to 100% is a warning sign, not a success.

5. **Write the rules down.** Separate what is never allowed (law, compliance, industry standards) from what leadership wants (company standards, service style). Written, versioned rules are the precondition for scaling oversight. Where the EU AI Act applies to high-risk systems, its Article 14 requires effective human oversight (European Union, 2024).

## 5 · Where to start

I suggest building a portfolio of AI initiatives systematically rather than case by case. A first step can be a small strategy project that selects where to start, places every candidate use case on the matrix, and names the Human only acts for each function. Within that portfolio, the priority should be *augmentation*: use cases in which AI strengthens people’s work while they keep the judgement and the accountability, as in Sign-Off, Proof Trail and Co-Creation. Starting there builds the review routines, written rules and measures of what oversight catches that any later move towards Autopilot depends on.

## 6 · Limitations

That every task can be AI-supported is a position, not an empirical finding, and support does not mean automation. The matrix was derived from enterprise software sales (Hötter, 2026) and has not yet been validated in other domains such as manufacturing or services. The task examples are illustrative and do not describe any particular organisation. A portfolio as proposed above would also test the matrix: placing, running and measuring a few use cases over two to three quarters.

## References

- Bowman, S. R., Hyun, J., Perez, E., et al. (2022). [Measuring progress on scalable oversight for large language models](https://arxiv.org/abs/2211.03540). arXiv:2211.03540.

- Cobbe, K., Kosaraju, V., Bavarian, M., et al. (2021). [Training verifiers to solve math word problems](https://arxiv.org/abs/2110.14168). arXiv:2110.14168.

- Dell’Acqua, F., McFowland III, E., Mollick, E. R., et al. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013.

- European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, 12 July 2024.

- Hötter, J. (2026). Verification asymmetry in enterprise SaaS sales: Human–AI interface design patterns for AI-based workflows. Master’s thesis, Hasso Plattner Institute, University of Potsdam. Supervised by F. Uebernickel and T. Wuttke.

- Jimenez, C. E., Yang, J., Wettig, A., et al. (2024). SWE-bench: Can language models resolve real-world GitHub issues? International Conference on Learning Representations (ICLR).

- Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors*, 39(2), 230–253.

- Skitka, L. J., Mosier, K. L., & Burdick, M. (1999). Does automation bias decision-making? *International Journal of Human-Computer Studies*, 51(5), 991–1006.

- Wei, J. (2025). [Asymmetry of verification and verifier’s law](https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law). Blog post.

- Wolfe, J. M., Horowitz, T. S., & Kenner, N. M. (2005). Rare items often missed in visual searches. *Nature*, 435(7041), 439–440.

[Download the working note (PDF, 4 pages, with counterpoints)](https://uebernickel.com/papers/Uebernickel_2026_Every-task-can-be-AI-supported.pdf) · [Deutsche Fassung (PDF)](https://uebernickel.com/papers/Uebernickel_2026_Jede-Aufgabe-laesst-sich-mit-KI-unterstuetzen_DE.pdf)