Seventeen capabilities.
What a step of work requires, and what a model can do with it today. The unit behind Coverage in HRPulsar.
Ask which steps of a process an agent can take and you get two answers: “most of them” from vendors, “none” from the person doing the work. Both answer for the whole job. To answer per step you need a unit that fits a person and an agent equally well. This page is that unit: the list itself, and how we built it from 39 published processes.
- published processes
- 39
- steps, as their sources describe them
- 390
- sources, 13 APQC categories
- 123
- capabilities: 13 cognitive, 4 boundary
- 17
Why not a job title. Why not a competency.
A job title is too coarse: “accountant” covers steps a model does better than a person and steps it can't touch. A competency describes a person, not the work, and it never carries the thing that most often locks a process: accountability. In HR models, accountability lives in the grade and the job, not in the competency.
So the unit is a capability: something you can require of a person or an agent, and about which you can say, honestly and with a date, whether a model can do it today. The list of such units wasn't written at a desk. It was derived from the work, and it turned out short.
Thirteen cognitive. Four boundary.
Coverage is counted on the cognitive thirteen. The boundary four are where a model isn't in the picture by the nature of the work; they're listed so the limit stays visible and is never counted as a gain.
- Yes
- A model does the step; a person spot-checks. “Better than a person” means it's also more thorough on the same step.
- Draft
- The model produces, a person checks before anything leaves. The most common case across the library, not the advanced one.
- No
- The act itself stays with a person, even when a model prepared every piece of material for it.
Cognitive: coverage is counted here
| # | Capability | Where it shows up | A model today |
|---|---|---|---|
| P1 | Extracting structure from an unstructured source | Text, speech, a scan, or a form turned into fields and relations | Yes |
| P2 | Checking against a reference and finding discrepancies | The reference is set in advance: a playbook, a regulation, a register, a previous period | Yes, better than a person |
| P3 | Rule-based classification and routing | The rules are known and expressible; the choice is from a finite set | Yes |
| P4 | Investigating a cause across heterogeneous sources | There is no reference; the hypothesis is built from logs, statements, and documents | Draft |
| P5 | Generating text in a given register | A letter, a status update, a report, the wording of a correction | Draft |
| P6 | Judging risk or priority in a business context | The cost of a mistake is known only inside the company's context | No |
| P7 | Open-ended dialogue with an external party | Negotiation, an interview, bargaining; the set of outcomes is unknown before the conversation starts and the counterpart's reaction changes the result | No |
| P8 | Consolidating people's contributions and resolving priority conflicts | Assign a person, reconcile opinions, push a decision through, agree on a priority, hand over to the owner | No |
| P9 | Producing a working artifact from a specification | A drawing, code, a schema, a patent claim, a training course; verified by whether it works, not by the register of its prose | Draft |
| P10 | Designing a measurement and an experiment | What counts as the metric, how to define it operationally, how to set up a pilot so the effect is distinguishable from noise | Draft |
| P11 | Coordination by calendar, triggers, and lists | Notify, collect responses, track readiness, put on the schedule, escalate by rule | Yes |
| P12 | Scripted or regulated dialogue with an external party | The counterpart's reaction picks a branch; the set of outcomes is known before the conversation starts: a question tree, a questionnaire, a closing procedure | Yes |
| P13 | Producing a result accepted by a stakeholder's taste and consent | A creative concept, a visual identity, a design direction; the acceptance criterion cannot be stated before the showing and is revealed only through the accepting party's reaction | Draft |
Boundary: the limit, outside the count
Adding these to the count is the easiest way to overstate the gain: reading a gauge is cheap and frequent, an on-site inspection is rare and expensive. The line between the two moves as sensors get cheaper; at review time a step migrates from one row to the other, the rows stay.
| # | Capability | Where it shows up | Who covers it |
|---|---|---|---|
| B1 | Instrumental reading of a value from an object | Temperature at a critical control point, a scan at goods receipt, weight, vital signs | An instrument, not a model |
| B2 | Physical action on an object | Unloading, placement, repair, withdrawing a batch, fixing defects | A person |
| B3 | On-site sensory assessment of the condition of an object or a person | A dented package, play and noise in an assembly, a patient's appearance, a user's behaviour in a task | A person |
| B4 | Transferring a skill to a person with in-the-moment correction | Demonstration, a supervised attempt, correction before the mistake sets in | A person |
Catalog v1.1. Verdicts issued on 26 and 31 August 2026 on the frontier-2026-08 model generation; reviewed quarterly. The codes change only with a new catalog version and a re-annotation of the whole library.
From the work, not from people.
The obvious route was to decompose a competency framework. It gives human atoms, the same indicators cut differently, and never gives accountability. So we went the other way: take real processes and see what their steps require. Four rules held throughout.
- 01
Published sources only.
A regulator's document, a standard, a textbook, a vendor guide. Nothing from memory: a process without a source doesn't enter the library. The frame is the APQC Process Classification Framework, so the sample isn't whatever vendors write about most.
- 02
Steps as written.
A process is cut into steps the way its source describes it. We don't improve the process on the way in; that would replace the sample with our opinion.
- 03
Four marks per step.
Which capabilities the step needs, who is accountable for the result, whether the output is a draft for review or a change in the world, and what a model can do with it today.
- 04
The list doesn't grow mid-session.
A step that fits no capability becomes a candidate, reviewed by a person between batches. Without that rule, ten sessions produce two hundred capabilities and the unit is gone.
What happened
We expected around forty capabilities. Five office processes gave eight, and from the third process on, nothing new. So we went as far from an office as we could: a warehouse, a food-safety line, a construction site, an emergency room. Eight such processes added no cognitive capability at all. A triage nurse and a contract lawyer need the same handful. What they added was hands, and that became the boundary list.
The list grew only when the kind of result changed: a drawing instead of a document, a trained person instead of a report, an experiment instead of an analysis. Far from an office is not the same as far from the list.
Two capabilities were split later, after the team pushed back on the first version. Coordination by calendar and triggers (a model does it) came out of consolidating people's contributions (it doesn't). Scripted dialogue (it does) came out of open-ended dialogue (it doesn't). The test for the second split is one question: is the set of outcomes known before the conversation starts.
One verdict we refused to soften. Judging risk or priority appears in 37 of the 39 processes. A recommendation system doesn't turn it into a draft: on the same step it extracts, checks against a reference and builds a hypothesis, and each of those is already counted under its own capability. What isn't counted is the act of deciding, together with the accountability for it. Soften that one row and coverage becomes nearly total, which is exactly the exaggeration this list exists to prevent.
Capabilities aren't the whole step.
Every step in the library carries four more marks. They decide whether a step can move at all, whatever the capability verdict says.
Accountability
None, reputational, formal, or regulatory. It belongs to the step, not to the capability: the same “write a text” weighs differently in a follow-up email and in a public status page during an incident. Regulatory accountability sits on 87 of the 390 steps, and not only in legal and finance: candidate screening, triage, food safety, construction, KYC, clinical trials.
Reversibility
Reversible, costly to reverse, or irreversible. A draft is cheap to throw away. A shipped batch is not.
Weight
How often the step runs and how long it takes. Without it, counting steps overstates the gain: the steps a model is strong at are cheap and frequent, the ones it can't take are rare and expensive.
Output
A draft that a person checks, or a change in the world. A step may need no capability at all and consist of accountability only: signing a contract, releasing to production, deciding a recall. That's a normal step, not a gap in the analysis.
What this list is not.
- It's a snapshot.
- Every verdict carries a date and the model generation it was issued on (frontier-2026-08). We review them quarterly. Verdicts move; the codes don't. The list goes stale one column at a time, not all at once.
- The sources lean vendor.
- 47% of the 123 sources are vendor or consulting material, and a vendor describes a process the way its product covers it. The regulated processes were taken from regulators (FDA, FinCEN, ICH, FCA), and every new batch needs at least a third of its sources to be neither vendor nor consulting.
- Guides describe the ideal process.
- Not the one running in your company. We don't settle that by argument. It's settled in the product, by how many suggested steps a company edits or rejects.
- It's a library, not your company.
- No number on this page says what share of your work an agent can take. The product answers that per step, for your process, with your numbers.
Coverage: your process, step by step.
Shipped · v2.1.1Coverage takes a description of a process, cuts it into steps against this list, and returns three lists: what goes to an agent, what stays with a person, and what nobody covers yet. A step nobody covers becomes a hiring need. A step that goes to an agent comes with a skill file for it. Agents run in your own environment; the platform hands over the file and executes nothing.
Open list. Bring one process.
The list and this page are published under CC BY 4.0: reuse it, cite it, argue with it. If you want to see your own process cut against it, send us one. You get three lists of steps back, and we get one more real process next to the published ones.
Process frame: APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. The PCF is intended to facilitate organizational improvement through process management and benchmarking, regardless of industry, size, or geography. To download the full PCF or industry-specific versions of the PCF, as well as associated measures and benchmarking, please visit www.apqc.org/pcf.