How to manage AI agents: the delegation playbook for knowledge workers

Why delegating to agents is management, not prompting

An AI agent is a model that takes a goal, chooses steps, calls tools and returns a result without you watching each move. The moment you stop watching, you have delegated, and delegation has rules that predate software: a clear brief, a definition of done, a check before the work leaves the building, and a record of what went wrong.

The demand for this skill is early but real. Our keyword data shows "how to manage ai agents" at only 30 US searches a month, up 200% year on year with an advertiser cost per click of $56 (DataForSEO, retrieved 2026-11-09); "ai agents for work" sits at 90 a month, up 367%. On the employer side, Microsoft's 2025 Work Trend Index reports that 82% of surveyed leaders expect to use "digital labour" to expand capacity within 12 to 18 months.^[1] A McKinsey's November 2025 survey, read through secondary coverage, puts 23% of organisations at the stage of scaling an agentic system and a further 39% experimenting, while only 39% attribute any EBIT impact to AI at all.^[2] B Gartner, an analyst firm, predicts that over 40% of agentic-AI projects are cancelled by the end of 2027 for cost, unclear value or weak risk controls; the figure is an analyst estimate taken from a press-release headline and we grade it C.^[3] C

Put together: most organisations are about to hand work to agents, few have shown a return, and a large share of projects are expected to fail. The gap between those numbers is a management gap, and this page is the playbook for closing it at the level of one knowledge worker and their agents.

In our Future Demand Radar, the "AI agents and delegated work" category scores 59.7 out of 100 in the baseline scenario (p10–p90 band 55.2–64.1), the highest of twelve categories and the only one besides workflow automation classed "build now". Under the conservative scenario it drops to 56.0 and stays "build now". Our baseline scenario expects search demand for the category to reach about 1.7 times today's level by 2028 and 3 times by 2030 (band 2.3–3.9), with a trough risk in 2027–28 if the cancellations Gartner predicts materialise. Forecasts are model outputs with stated uncertainty, not facts.

Writing a task brief an agent can execute

Agents fail on briefs the way junior colleagues do: the goal is implied, the constraints are in your head and "done" was never defined. Anthropic's engineering guide on building agents makes the same point from the vendor side, recommending simple, composable workflows and explicit tool descriptions over clever autonomy.^[4] C

A brief that works has six fields, and every template below uses them.

Field	What to write	Example
Goal	One sentence, outcome not activity	"A one-page summary of the three Q3 customer interviews, with quotes"
Inputs	Where the material is and what is out of scope	"The three transcripts in the /interviews folder; ignore the sales call"
Constraints	Length, format, tone, sources allowed	"Under 400 words, UK spelling, quote only from the transcripts"
Done means	The acceptance test (next section)	"Every quote is traceable to a transcript line; no claims outside them"
Budget	Runs, tokens or minutes, and what to do when exceeded	"One run; if the transcripts exceed the context window, stop and tell me"
Escalate when	Conditions that require a human	"Any mention of a named customer's health or finances"

Write the brief once, save it with the task, and reuse it. A brief that lives only in a chat window is a brief you rewrite every week.

Acceptance tests and spot checks

The acceptance test is the field most people skip, and the evidence says it is the field that matters most.

In a field experiment with 758 consultants, GPT-4 users completed 12.2% more tasks, 25.1% faster and at higher quality for tasks inside the model's competence, but were 19 percentage points less likely to reach the correct answer on a task designed to sit just outside it.^[5] B The consultants did not know which side of the frontier they were on. Neither do you, which is why every delegated task needs a check that does not rely on the output looking right.

Three kinds of acceptance test, from cheapest to most reliable:

1. Structural. Does the output have the required parts, length and format? Automate this; an agent can check another agent's structure.
2. Traceability. Can every claim, number or quote be pointed to a source you supplied? Sample three at random and trace them yourself.
3. Ground truth. For a task with a known answer (a reconciliation, a calculation, a lookup), compare against the answer. Run a known case through the agent periodically even when live work seems fine.

Spot-check rate: start at 100% for a new task type, drop to one in three after ten clean runs, and never below one in ten for anything that leaves your desk. Record the rate in the run log so a future you can see when trust was earned.

Review cadence and escalation

A weekly agent review takes fifteen minutes and has one agenda: the run log. Count runs, failures, corrections and the time you spent checking. If checking takes longer than doing, the task is not ready for delegation; the METR study of experienced open-source developers found them 19% slower when allowed to use early-2025 AI tools while believing they were about 20% faster, a reminder that felt speed is not measured speed.^[6] B That was 16 developers on mature codebases they knew intimately, so it does not generalise to all work, but it does generalise to one habit: measure, do not feel.

Escalation rules belong in the brief, not in your memory. Typical triggers: the agent cannot find an input; the output contradicts a source; the task touches money, legal wording, health, a named person or anything external-facing; or the budget is exceeded. On a trigger the agent stops and reports; it does not retry creatively.

Budgets: tokens, runs and time

Three budgets, all written down:

• Runs. How many attempts before the task comes back to you. Default one; two for exploratory work.
• Tokens or cost. For API-driven agents, a hard ceiling per task and per month. Automation platforms such as n8n, Zapier and Make expose run counts and let you cap executions; model providers expose spend limits. Set both.
• Your time. The minutes you expect to spend briefing and checking. If the sum across the month exceeds the time the task used to take you, stop delegating it.

A delegated task is only cheaper if all three budgets hold. Most "agent ROI" disappointments in the surveys above are a time budget that nobody tracked.

Permissions and data boundaries

Give an agent the narrowest access that lets it complete the brief. Read access to one folder, not the drive. A draft mailbox, not send rights. A sandbox account, not production. Any action that sends, pays, publishes or deletes gets a human approval step; the automation platforms above all support an approval node, and it should be the default, not the exception.

Two external constraints shape the boundaries. The EU AI Act's AI-literacy duty (Article 4) has applied since 2 February 2025 to providers and deployers, which in practice means employers must ensure staff who operate AI systems have sufficient understanding of them; the Act's high-risk obligations for workplace uses now start on 2 December 2027 after the 2026 Omnibus amendment.^[7] A And in Deloitte's survey of 2,773 senior leaders, regulatory compliance was the top barrier to agent deployment for 38%, up from 28% a wave earlier.^[8] A A personal run log and brief archive is the lightest possible version of the documentation these rules point toward. This page is not legal advice; your organisation's policy comes first.

What not to delegate, and why

The automation literature is older than large language models and unusually consistent. In a flight-simulation experiment, participants given an automated aid made more errors when the aid was wrong than participants with no aid at all, even with contradicting information on screen.^[9] B A systematic review of 74 studies of clinical decision support found users following incorrect advice at non-trivial rates, with trust, task complexity and time pressure as the mediators.^[10] A An integrative review in Human Factors describes complacency and automation bias as attentional effects of trust and workload that fall with accountability and training, not as traits of careless people.^[11] A A meta-analysis of 106 human–AI experiments found that combinations under-performed the best of human or AI alone on average, with the losses concentrated in decision tasks and the gains in content creation.^[12] A And a CHI 2025 survey of 319 knowledge workers found that higher confidence in the AI was associated with lower self-reported critical-thinking effort, a self-report finding we grade C but which matches the pattern above.^[13] C

From that evidence, a short list of things to keep:

• Final decisions with consequences. Hiring, firing, pricing, medical, legal and financial calls. The decision-task losses in the meta-analysis land here.
• Anything you cannot check. If no acceptance test exists, you are trusting, not delegating. Automation bias is strongest exactly where checking is hardest.
• Tasks where you are the expert. Gains concentrate among novices; experts gained little in the support-agent study and lost time in the developer study.^{[6] [14]} A
• Relationships. Feedback, difficult conversations, apologies, anything a person remembers receiving from you.
• Work that keeps a skill alive. If you never draft, you lose the ability to judge drafts. Keep a share of each task type for yourself, especially the kind you want to stay good at.
• Time-pressured judgement. The clinical review names time pressure as a mediator of following bad advice. Under pressure, slow down or decide yourself.

Keeping the skills you still need

The skills a manager of agents needs are the ones the evidence says protect against the failures above: writing a precise brief, designing an acceptance test, sampling and tracing, reading a run log, and knowing your own frontier. Our guide to AI skills for 2030 ranks these against tool-specific skills by durability, and human skills in an AI economy covers the judgement, taste and coordination skills that complement delegation. Both pages separate what the labour-market data measures from what our forecast expects.

Practise on low-stakes work first. A weekly reading digest, meeting-note clean-up, data reformatting and first-pass research are tasks where the acceptance test is cheap and the cost of a miss is small.

Templates

Copy these into your notes tool. They are plain text on purpose; the brief should travel with the task, not live in a vendor's interface.

Agent task brief

TASK: <one-line name>
GOAL: <outcome, one sentence>
INPUTS: <where the material is; what is out of scope>
CONSTRAINTS: <length / format / tone / allowed sources>
DONE MEANS: <acceptance test, see below>
BUDGET: <runs> run(s); <tokens or cost cap>; stop and report when exceeded
ESCALATE WHEN: <conditions requiring a human>
APPROVAL NEEDED FOR: <send / pay / publish / delete / none>

Acceptance test

STRUCTURAL: <parts, length, format that must be present>
TRACEABILITY: <N> claims sampled at random, each traced to a supplied source
GROUND TRUTH: <known case to re-run, if any> expected result: <...>
SPOT-CHECK RATE: 100% for first 10 runs → 1 in 3 → never below 1 in 10

Run log (one line per run)

date | task | runs used | passed test? | corrections made | minutes checking | note

Three lines a week in the run log tell you more about whether delegation pays than any vendor dashboard.

What to do next

The OS Builder includes an AI-delegation section gated by your experience and privacy answers: it suggests which task types to delegate first, which to keep, and assigns automation recipes to a platform already in your stack, with an approval step before any external action. Start there, copy the three templates above into your notes, and run one low-stakes delegation with a full acceptance test this week. For the wider picture of where delegated work is heading, read our three scenarios for knowledge work to 2035.

Frequently asked questions

What is the difference between an AI assistant and an AI agent?

An assistant answers within a conversation you steer turn by turn. An agent takes a goal, plans steps, calls tools (search, files, email, APIs) and returns a result with less supervision. The management problem starts when supervision drops, which is why the brief, acceptance test and run log matter more for agents than for chat.

How much should I check an agent's work?

Check everything for the first ten runs of a new task type, then sample one in three, and never go below one in ten for anything that leaves your desk. Research on automation bias shows people follow incorrect automated advice at non-trivial rates, with trust and time pressure making it worse, so the check has to be a rule rather than a feeling.^[10] A

Which tasks are safe to delegate first?

Bounded production tasks with a cheap acceptance test and a small cost of error: reading digests, meeting-note clean-up, reformatting data, first-pass research and first drafts you then rewrite. Controlled studies show the largest gains for such tasks and for people who are not already experts in them.^{[5] [14]} B

Do I need an automation platform to manage agents?

Not to start. A chat assistant plus the three text templates on this page is enough for one person. Platforms such as n8n, Zapier or Make become useful when a task should run on a schedule or trigger, because they add run caps, logs and approval steps that a chat window lacks.

Does the EU AI Act apply to me if I just use agents at work?

The Act's AI-literacy duty (Article 4) has applied since 2 February 2025 and is addressed to providers and deployers, so in an employment setting it falls mainly on your organisation, which must ensure staff using AI have sufficient understanding of it. High-risk obligations for workplace uses now start on 2 December 2027.^[7] A This is not legal advice; check your organisation's policy.

How do I know if delegation is actually saving time?

Keep the run log and add up minutes spent briefing and checking each month. If that sum approaches the time the task used to take you, stop delegating it. The METR study found experienced developers who felt faster with AI were measurably slower, so measure rather than estimate.^[6] B

Sources and evidence

1. 2025 Work Trend Index Annual Report: The year the Frontier Firm is born — Microsoft WorkLab (2025), report B www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-fro (retrieved 2026-11-09)
2. The state of AI in 2025: Agents, innovation, and transformation (as reported by ppc.land) — McKinsey & Company (QuantumBlack) (2025), report B ppc.land/most-companies-still-pilot-ai-programs-despite-widespread-ado (retrieved 2026-11-09)
3. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (press release; analyst estimate, headline verified via search only) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
4. Building effective agents — Anthropic (2024), vendor page C www.anthropic.com/research/building-effective-agents (retrieved 2026-11-09)
5. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan R.; Lifshitz-Assaf, Hila; Kellogg, Katherine; Rajendran, Saran; Krayer, Lisa; Candelon, François (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
6. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, Joel; Rush, Nate; Barnes, Elizabeth; Rein, David (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
7. AI Act | Shaping Europe's digital future (Regulation (EU) 2024/1689, application timeline incl. 2026 Omnibus amendment) — European Commission (2026), report A digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (retrieved 2026-11-09)
8. The State of Generative AI in the Enterprise: Generating a new future (Q4 report) — Deloitte (2025), report A www.deloitte.com/us/en/about/press-room/state-of-generative-ai.html (retrieved 2026-11-09)
9. Does automation bias decision-making? — Skitka, Linda J.; Mosier, Kathleen L.; Burdick, Mark (1999), RCT B doi.org/10.1006/ijhc.1999.0252 (retrieved 2026-11-09)
10. Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard, Kate; Roudsari, Abdul; Wyatt, Jeremy C. (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
11. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman, Raja; Manzey, Dietrich H. (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
12. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, Michelle; Almaatouq, Abdullah; Malone, Thomas (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
13. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Hao-Ping (Hank); Sarkar, Advait; Tankelevitch, Lev; Drosos, Ian; Rintel, Sean; Banks, Richard; Wilson, Nicholas (2025), observational C doi.org/10.1145/3706598.3713778 (retrieved 2026-11-09)
14. Generative AI at Work — Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)