The Delegation Map: What AI Takes Over in Your Processes and Where the Human Stays

An inquiry comes in, a quotation has to go out. In between lie individual steps: someone reads the inquiry, picks the matching line items, compares the prices with earlier orders, decides on a discount, writes the text and enters the result into the ERP system. Anyone asking whether AI can take over “quotation preparation” is asking one question about six very different activities and therefore gets no usable answer.

The question becomes usable once it is asked per step. For picking the line items the answer differs from the one for the discount, and for sending it differs from the one for reading. We therefore work with a delegation map: a table with one row per step that records what AI takes over, what runs automatically as a fixed rule and at which points a person decides or reviews. The map is small enough to be created in one meeting, and concrete enough to turn into an implementation plan. It starts where the overview in using AI in your business ends: after the question of where AI helps at all comes the question of how far it may go in each step.

Modelling business processes without diagram software

The starting point is a light form of process analysis. Mature notations such as BPMN exist for modelling business processes, with their own tools and seminars; for the delegation map you need none of that. Instead of the official procedure, we replay a real case from trigger to result. Who read what, what was decided and created, and where was the result transferred? The actual workflow almost always deviates from the official one, and the map has to describe the actual one, because otherwise you automate a process that does not really exist.

Every step falls into one of five verbs.

• Read: take information in, for example reading the inquiry or finding the old case.
• Compare: check and evaluate, for example against the price list or against earlier cases.
• Decide: choose a path or commit a value, such as the discount or the rejection.
• Create: produce something new, such as the quotation text or the reply to a complaint.
• Transfer: pass results on, that is entering them into the ERP system, sending the email or creating the order.

Five to fifteen steps is the right level of detail. If the map gets finer, it already describes the implementation; if it gets coarser, exactly the decisions that matter disappear from view. No more modelling is needed at this point; the model is only the means of making the decisions visible.

Three questions per step

For each row of the map we then settle three questions, in this order.

1. Can “correct” be defined? Is there a fixed rule, is there a standard with reference cases, or does the right answer depend on the individual situation?
2. Can a machine decide whether it is right? A schema, a reconciliation or a threshold checks tirelessly; a person with context checks in seconds; some results can only be judged with the full knowledge of the person doing the work.
3. What does an error cost? Who gets to see it, can it be undone, are money, deadlines or sensitive data involved, and how often does the step run?

The first question separates fixed rules from AI and from human judgment. The second decides whether the review can be automated or needs people; Vibe Coding and the Review Bottleneck shows where the bottleneck lands as a result. The check can be automated wherever the correct result can be derived from the source data: a second agent then recomputes every case, and the judgment calls stay with a person. The third determines where the review sits and how thorough it has to be: review effort follows the risk of the individual step, not a blanket quota.

The five levels of delegation

The answers place each step on one of five levels.

• Stays with the human: judgments with legal or significant effect, accountability and relationships, plus every step whose rules are still unclear within the company itself. AI may supply material, with its sources marked.
• AI assists, the human leads: research, analysis and comparisons. The person assembles the result and stands behind every sentence; the review is built in.
• AI drafts, a human releases: the whole step can be delegated, but the result leaves the company or touches money, deadlines or customers. A named person reviews against clear criteria and releases.
• Runs automatically, samples are reviewed: recurring and reversible steps within narrow bounds, whose result is checked by a machine. Control moves downstream and runs through samples, exception reports and a log that records every step.
• Fixed rule, no AI: where a stable rule is enough, no language model is needed. The rule is tested once and then runs predictably.

Two principles speed up the classification. Read broadly, write narrowly: steps that only take in and compare may go to AI early and extensively; steps that write to the outside start at the draft level, regardless of how good the model is. And AI does not invent values that somebody will later be accountable for; an empty, marked field is better than a plausible estimate.

What does human in the loop mean?

Human in the loop literally means that the person is part of the loop: the workflow stops at a defined point and waits for review or release before anything binding happens. It is distinct from human on the loop: the workflow runs through, and the person supervises it, is notified on exceptions and can intervene. In the language of the map, the draft level is human in the loop and the automatic level with samples is human on the loop. The term is also used for people who rate the training data of AI models; this article covers only the operational use.

A rule from our own work belongs exactly here: the person is called in when they are needed; they do not watch. A well-designed workflow runs without anyone starting or observing it; drafts collect in a release queue, automatic steps report only with exceptions and samples. As soon as a step is only safe while somebody keeps a dashboard open, the review point sits in the wrong place. This is how we run our own AI agents, and it is the pattern we use when building workflows for clients.

A worked example: sales research for CNC automation

What a completed map looks like is shown by the AI sales assistant we run for the machine builder MAFU-SHERPA. Its task is to find and approach suitable target customers for camera-guided CNC automation: companies showing concrete technical buying signals, not generic industry lists.

• Read (runs automatically): the system searches the web for companies with matching buying signals and reads their websites, across hundreds of companies, with a source noted for each finding.
• Compare (runs automatically): the fit is scored from 0 to 100 following mostly fixed criteria, with AI only assisting; clear exclusions are filtered out automatically.
• Create (AI drafts): the outreach is produced as a draft that names the machines and signals actually found.
• Decide and transfer (a human releases): the sales team reviews every draft together with its reasoning and sends it or discards it.

Nobody watches the system while it researches; the sales team only sees the pre-scored results with their reasoning. The review at the end is not a concession to the technology but part of the design: the system supplies the technical assessment and the reasoning, the person decides. The details are in the case study on targeted sales for CNC automation. What stands out about this map is where the boundary runs. Everything that only reads and evaluates runs automatically; everything that reaches a customer waits for a person. We would distrust a map in which every row says “runs automatically”.

The small-scale counterpart is item capture in XPO Inventory at the trade-fair builder NEO Expo. A warehouse employee photographs an item, and the app suggests name, description, color, category and unit; dimensions, quantity and item number are values the AI deliberately never invents, because they are too critical for inventory processes. This is a single row of the map, set cleanly to “AI drafts, a human releases”. Field-by-field typing turns into a photo, a quick check, done.

A step has to earn more autonomy

The level of a step changes over time, in both directions. Every step starts one level more supervised than its target; what is meant to run automatically one day begins as a draft with a release. Promotion happens on evidence from real operation. A fixed set of reference cases, defined thresholds and logged corrections show whether a step has earned the sampling level; if corrections start piling up again, it moves back the same way. Some things, however, are never promoted: commitments, prices, legal statements and anything that claims customer results leave the company only after human review. In the sales assistant, this is why reading the websites is the part that runs automatically, while the outreach to a potential customer stays at the release level. The patterns by which such releases work in CRM, ERP and ticketing systems are described in our guide to back-office agents.

When the map says “no AI”

Filled in honestly, the map regularly produces uncomfortable results. If a workflow sits close to the market standard, standard software is the better answer. If the workflow itself is unclear, AI only makes it wrong faster. A task that is rare, or whose data base is too thin, does not pay for the effort. And where a stable rule is enough, an AI agent would be the more expensive and less predictable solution. Which workflows are suitable for automation at all is covered in automating business processes. Which technology sits behind a row (a no-code platform or an agent in a coding environment) depends on the shape of the task; that decision is answered in building AI agents: no-code or code.

Where to start

Take a workflow that has been bothering you for months and fill in the map in one meeting, with at least the process owner and one person who does the work today. One row per step, with five entries.

• Step and verb: what happens, and is it reading, comparing, deciding, creating or transferring?
• Level: from “stays with the human” to “fixed rule”, when in doubt one level more cautious.
• Review point: who reviews, and does it happen before release or downstream via samples?
• Notification: how does the reviewing person notice that they are needed, without watching the workflow?
• Evidence for promotion: which reference cases and thresholds would have to be met before the step may run more autonomously?

You can fill in this map without us, and filling it in often answers the question of where to start. For selecting the processes with the greatest leverage, the AI opportunity assessment is the structured path.

If you want to work out for one concrete workflow which steps AI can take over and where the review belongs: book an intro call.