AI Deputization Audit

One-line job: Run a five-criterion scoring rubric (worth-it / teachability / checkability / stakes / integral-to-process, 0–10) on each recurring process in your work, and decide whether to deputize AI on it, duet on it, or defend it from automation. Audience: Knowledge workers and operators picking where AI agents should run unattended versus where they need a human-in-the-loop gate. Not for: Anyone looking for a one-click automation tutorial. This is a triage framework — the audit produces a decision per process, not an automation script. Last verified: 2026-08-16 Evidence weight: documentation-verified + inferred from public sources. Primary docs (xAI GrokBot, ChatGPT Computer History) cited inline; secondary trade press (The Register, Elephas) provides the privacy critique.

The problem

Two product launches in mid-August 2026 — xAI’s GrokBot “teach-a-task” feature (announced 2026-08-11) and OpenAI’s ChatGPT Computer History (documented 2026-08-13) — moved AI agents from “thing you prompt” to “thing that watches how you work.” Models are capable enough; they just don’t have your context. Both products attack that gap, in two different directions: GrokBot teaches by deliberate demonstration, ChatGPT learns by ambient observation (AI Daily Brief, 2026-08-14 episode; xAI Grok Bot announcement; ChatGPT Computer History docs).

The temptation is to automate everything. That is the wrong move. Nathaniel Whittemore’s framing on the same episode — “deputize, don’t automate” — is the load-bearing distinction: deputizing AI implies an ongoing relationship and a spot-check handoff, not a set-and-forget script (AIDB 2026-08-14 transcript §deputization-not-automation). The audit exists to tell you, per process, which side of that line you are on.

The rule / frame

Score each recurring process 0–2 on five criteria. Total 0–10. 8–10: deputize. 4–7: duet (AI drafts, you confirm). 0–3: defend (no AI; the stakes or the checkability don’t allow it). The five criteria:

1. Worth-it.
2. Teachability.
3. Checkability.
4. Stakes.
5. Integral-to-process.

For each criterion, the score is 0 (no), 1 (partial), or 2 (yes-with-evidence):

• Worth-it — Will the time saved pay back the cost of setting it up, monitoring it, and recovering when it drifts? A process that saves five minutes a week is not worth a chief-of-staff deployment.
• Teachability — Can you actually show the model what success looks like, in a way it can reproduce? “Reply to customer email in our voice” is teachable; “know what the customer will want next quarter” is not.
• Checkability — Can a human spot-check the output cheaply? The cheaper the spot-check, the higher the score. A weekly report is cheap to check; a one-shot real-estate bid is not.
• Stakes — How bad is a wrong answer? Low-stakes processes get higher scores; high-stakes processes get hammered down regardless of the other four.
• Integral-to-process — Is the process core to your value, or peripheral? Peripheral processes can be deputized freely. Core processes should stay human-driven unless the other four scores are high enough to absorb the risk of ceding them.

This is not novel in form. It maps onto Intercom’s RICE scoring model (Reach × Impact × Confidence ÷ Effort, 2018) and Onder’s six-dimension Universal Decision Automation Framework (2026-03-31). It is novel in adapting that lineage to a 2026 “human–AI teammate” framing where the agent’s job is to learn your workflow, not to run a fixed script.

The audit, in practice

Step 1 — Inventory. List the recurring processes you actually do weekly. Aim for 8–15. The wrong granularity produces useless audits: “email” is too coarse, “send the Monday morning standup notes” is right-sized.

Step 2 — Score each. Use the 0–2 scale per criterion. Total per row. Do this in a spreadsheet you can sort by total. The sort order is the prioritization order.

Step 3 — Triage into the three tiers. 8–10 = deputize (agent runs, human spot-checks on a schedule), 4–7 = duet (agent drafts, human approves before action), 0–3 = defend (no AI on this).

Step 4 — Pick the first one. Do not start with the highest total. Start with the highest-checkability item in the deputize tier — the one where a wrong answer is obvious within five seconds. That is your proof of concept.

Step 5 — Plan for several weeks of friction. NLW’s closing observation on the AIDB episode: “even highly AI-forward people struggle here, especially high-productivity workers whose workflows are so dialed in that changing them feels like a short-term waste of time” (AIDB 2026-08-14 transcript §carving-out-the-time). The transcript uses qualitative framing (“a little while,” “short term,” “long term”) rather than a specific week count; the recommendation here is to plan for a few weeks of intentional friction before judging whether the audit is producing net time savings.

Two paradigms you’ll meet

Ambient observation — ChatGPT Computer History watches across apps, building context without effort. The agent sees your clicks, typing, and app switches through macOS Accessibility and summarizes them into memories via a short-lived Codex session. You do not press record; it just runs (ChatGPT Computer History docs).

Deliberate demonstration — GrokBot’s teach-a-task. You press record, click through the workflow in your browser, save the recording as a reusable routine. The bot watches once, by your invitation, and learns from a single take (xAI Grok Bot).

Both have constituencies. Operators who hate being watched lean toward GrokBot. Operators who hate doing the same task twice lean toward Computer History. By the audit’s own scoring, deliberate demonstration scores higher on teachability for most knowledge-work tasks; the framework does not pretend to be neutral on which paradigm fits the rubric more cleanly. It does not claim ambient observation is wrong — it claims ambient observation does not fit the rubric as cleanly. Recommendation: choose by your teachability score, not by hype.

Failure modes

• Total-only sorting. A process can hit 9 because worth-it and integral-to-process are both 2, while stakes is 0. Sorting by total alone ships that to the deputize tier and burns you. Always look at stakes and checkability individually before promoting.
• Audit without a recovery plan. Deputized processes drift. The agent learns a wrong pattern, the rubric shifts, an upstream input changes. Every deputize-tier process needs a named “this is what to do when the output looks off” path before you turn it on.
• Skipping the 4–8 week ramp. The biggest barrier to net time savings is not capability, it is the operator’s own calendar (AIDB 2026-08-14 §carving-out-the-time). If you do not carve the time, the audit produces a spreadsheet and zero outcomes.
• Treating “duet” as a half-measure. Duet means agent drafts, human approves before action. If you let the agent act and the human reads the result later, you have deputized with a check that is too late. The tier name is load-bearing.
• Ignoring the privacy overlay. Recording your screen creates a new thing to secure. OpenAI’s Computer History ships unencrypted event files on the user’s Mac for up to 48 hours and depends on macOS Accessibility — the same macOS Accessibility permission that security researchers document as a standing target for keylogger and stalkerware malware (Elephas teardown, 2026-08). OpenAI’s own docs warn that “Computer History increases the risk of prompt injection from content in apps and websites” (ChatGPT Computer History docs). The Register frames Computer History’s design choice as “ditches Recall-style screenshot surveillance for friendly keylogging” (The Register, 2026-08-14) — broader industry critique, not the stalkerware-framing source. The same screen-recording surface area applies to GrokBot’s teach-a-task — every workflow you demonstrate is captured and stored. The cited critique below is for Computer History because Elephas and The Register published primary teardowns of that product; no comparable public teardown of GrokBot’s recording pipeline exists at time of writing, so the asymmetry is documentation, not necessarily risk. Add a sixth criterion if your work involves anything you would not hand to a stranger.

Done means

• You have a list of 8–15 recurring processes, each with a 0–10 audit score.
• The deputize tier (8–10) contains at least one process, and you’ve named its spot-check + recovery path.
• The duet tier (4–7) processes have a defined “human approves before action” step that is not deferred.
• The defend tier (0–3) processes are explicitly off-limits, not “we’ll get to that later.”
• You have a date on the calendar four weeks from now for the first review: is this producing net time savings, or is the spreadsheet drifting?

What this article does NOT cover

• Specific prompt-engineering tactics for any of the deputize-tier processes. The audit decides what to deputize; how to wire it is a separate guide per tool.
• Pricing comparison across GrokBot vs. ChatGPT Computer History vs. competitors. The audit is tool-agnostic; the choice of ambient vs. deliberate depends on your teachability score and your privacy tolerance, not vendor pricing.
• Compliance or regulated-industry gating (HIPAA, financial advice, legal documents). The privacy overlay is named; the regulated-industry rulebook is not.
• Team-level or org-level audits. This is per-operator. Scaling this to a team requires its own framework and is out of scope.

Related guides

• Why every agent needs a cost cap on day one — once the audit identifies what to deputize, the cost question becomes “what does each deputized process cost per month at realistic usage?”
• Your first agent on a Mac mini in 90 minutes — if the privacy overlay pushes you toward self-hosted models, this is the lowest-friction path to standing one up.
• Agent self-review is not a check — the “checkability” criterion only works if your spot-check is a real check. Same-model self-review does not satisfy this.
• Start here: local AI and agents — if you have not picked a base agent yet, walk this hub first.
• Hermes vs. OpenClaw: pick a path — a hosted-vs-self-hosted comparison relevant to the privacy-overlay decision.

Sources / what we ran

• AI Daily Brief, 2026-08-14 episode “How to Decide What Work AI Should Do for You” — the lead source for the audit rubric and the deputize/duet/defend framing.
• AIDB 2026-08-14 transcript (timecoded) — anchors every quoted phrase.
• xAI Grok Bot announcement, 2026-08-11 — primary docs for the teach-a-task / deliberate-demonstration pattern.
• ChatGPT Computer History docs, 2026-08-13 — primary docs for the ambient-observation pattern and the prompt-injection warning.
• The Register, Claburn, 2026-08-14 — independent critique of Computer History.
• Elephas teardown of Computer History, 2026-08 — pipeline + 48h unencrypted + Accessibility-permission risk.
• Microsoft Recall privacy/security architecture, 2024-09-27 — prior art for ambient-observation products.
• Intercom RICE scoring model, 2016 — closest prior-art scoring framework.
• **Onder, Medium, 2026-03-31](https://web.archive.org/web/2026/https://medium.com/@omerfonder/what-to-automate-a-universal-scoring-framework-for-automation-decisions-1d6d2d604c36) — six-dimension Universal Decision Automation Framework (Wayback mirror; the live Medium URL is Cloudflare-gated against bare-curl HEAD checks).
• Internal research dossier (operator-side, not customer-citable): Websites/content-strategy/research/ai-deputization-audit/ai-deputization-audit-research-2026-08-16.md — held locally for audit + future refresh; all load-bearing claims in this article are inline-cited above to public primary or secondary sources.