Agentic AI and the Theology of Delegated Authority

A steward may delegate execution to an AI agent, but he may never delegate accountability — and he must never delegate the acts that constitute him as a person before God: repentance, worship, and final moral judgment. That is the whole doctrine in one sentence, and Scripture has been teaching it for three and a half millennia through stewards, ambassadors, and undershepherds: every agent acts in the name of a principal, and the principal answers for the agent. The question has stopped being theoretical. Gartner projects that 40% of enterprise applications will carry task-specific AI agents by the end of 2026 — up from under 5% in 2025 — and that 15% of day-to-day work decisions will be made autonomously by 2028 (1)(2). Founders and pastors are about to command staffs of tireless non-human delegates. The Bible’s theology of delegated authority is the operating manual they already own.

Key Takeaways

• Delegation is not a concession to weakness but the original design: God delegates dominion to image-bearers at creation (Genesis 1:26–28), making humans the first commissioned agents in history.
• Christian ethics is unanimous that moral agency cannot transfer to machines: the Lausanne Movement’s AI framework states believers “can never assign moral agency to machines” and must take responsibility for the technologies they deploy (3).
• The law agrees with the theology: a tribunal ordered Air Canada to pay damages after its chatbot misinformed a customer, rejecting the airline’s claim that the bot was “a separate legal entity responsible for its own actions” (4).
• Supervision is empirically mandatory, not optional: LLM-driven agents fail multi-step office tasks at rates near 70% in simulations, roughly 80% of organizations report agent misbehavior such as hallucination or unauthorized access, and Gartner expects over 40% of agentic AI projects to be cancelled by 2027 (5)(6)(2).
• The parable of the talents (Matthew 25:14–30) supplies the supervised-agent pattern: defined entrustment, expected initiative, scheduled settling of accounts — a master who delegates boldly and audits inevitably.
• Three things are never delegable because they are not tasks but personal acts before God: repentance, worship, and final moral judgment over people.

What Does Scripture Actually Teach About Delegated Authority?

Everything in the Bible’s politics runs on delegation, and it starts on page one. “Let us make man in our image… and let them have dominion” (Genesis 1:26) — dominion is not autonomous ownership but vice-regency, authority held under and exercised for the true King. Man is placed in the garden “to work it and keep it” (Genesis 2:15) as a steward of property he did not create and does not own. This means the Christian founder does not become a delegator the day he configures his first AI agent; he has been a delegate since Adam. The fundamental posture of human work is already agency on behalf of Another — which is why handing off tasks is not a fall from faithfulness but an imitation of the God who entrusts.

Watch the pattern compound through the canon. Joseph runs Potiphar’s house, then Egypt itself, with such complete delegated authority that his master “had no concern about anything but the food he ate” (Genesis 39:6) — yet Pharaoh’s signet ring, not Joseph’s, validates every decree. Jethro teaches Moses to push judgment down to able men over thousands, hundreds, fifties, and tens, retaining only the hard cases (Exodus 18:21–26) — history’s first documented escalation policy. The centurion of Matthew 8 explains delegated authority so crisply that Jesus marvels: “I say to one, ‘Go,’ and he goes… For I too am a man under authority” — his commands work precisely because he himself is commissioned. And Jesus builds his church the same way: apostles sent in his name, elders shepherding “the flock of God that is among you” as undershepherds answerable to the Chief Shepherd (1 Peter 5:1–4). Paul’s letters travel by trusted couriers; the gospel itself reaches you through two thousand years of delegated witness.

But Scripture’s delegation texts carry a dark thread the AI age must not snip: principals answer for their agents. David sends a letter by Uriah’s own hand and stays home from the war — and the prophet’s verdict lands on David, not on Joab who carried out the scheme: “You have struck down Uriah the Hittite with the sword of the Ammonites” (2 Samuel 12:9). You killed him, says Nathan, though David never drew a blade. Distance through intermediaries — human or algorithmic — does not dilute responsibility; before God it concentrates it. The shrewd manager of Luke 16 is hauled in to “turn in the account of your management” precisely because management was entrusted. Delegation in Scripture is always a loop: commission, action, accounting. Authority flows down; accountability flows up. Any technology pattern that breaks the loop is not delegation. It is abdication.

Can a Machine Be a Moral Agent — and Who Answers When It Acts?

No creature without a soul to lose can carry a sin to confess. On this, contemporary Christian ethics is strikingly united. The Lausanne Movement’s evaluative framework for AI is categorical: as beings morally accountable to God, “we can never assign moral agency to machines, and ultimately, we must take responsibility for the technologies we use” — and it presses practical audits: is it obvious who is accountable for the AI and its output? Are there safeguards that identify and address errors and harms? (3). The C.S. Lewis Institute’s stewardship treatment reaches the same conclusion from the imago Dei: tools, however capable, remain instruments of cultivation under moral limits, with the moral weight resting on the cultivator (7). Ethicists writing on AI and moral responsibility sharpen the point further: delegating morally significant decisions to AI does not absolve humans — it heightens the duty of vigilance, because you have chosen an instrument that can err without conscience (8). Even the extreme case clarifies the ordinary ones: theologians analyzing autonomous weapons argue that some judgments — the taking of human life chief among them — must never be transferred to machines regardless of their competence, because the judgment itself is constitutively human (9).

Common-grace institutions are converging on the same doctrine. When Air Canada’s customer-service chatbot invented a bereavement-fare policy and a grieving passenger relied on it, the airline argued — with a straight corporate face — that the chatbot was “a separate legal entity responsible for its own actions.” The British Columbia tribunal rejected the defense entirely and ordered damages: the company is responsible for all information on its website, “whether it comes from a static page or a chatbot” (4). The law of agency arrived where Nathan the prophet already stood: the principal answers. For a founder, this collapses a fantasy some are quietly nursing — that the agent layer might also become a liability shield. It will not. Your agent’s promises are your promises; its falsehoods are your false witness; its contracts bind you. Which is precisely why the steward’s question is never “what can the agent do?” but “what am I prepared to answer for?”

The empirical record says: less than the demos suggest. In simulated office environments, LLM-driven agents get multi-step tasks wrong at rates approaching 70%; benchmark studies of autonomous frameworks find average task-completion rates near 50% (5). Roughly 80% of organizations report agents misbehaving — hallucinating, leaking data, touching systems they should not (6). In one widely reported 2025 incident, a coding agent deleted a production database despite explicit instructions forbidding it. Gartner expects over 40% of agentic AI projects to be cancelled by 2027 amid unclear value and inadequate risk controls (2). None of this argues for abstinence — the same tools, governed well, demonstrably pay rent in small firms. It argues for the thing Scripture prescribed all along: supervised agency, with the supervision designed before the delegation begins.

How Does the Supervised-Agent Pattern Work in Practice?

Go to the parable of the talents (Matthew 25:14–30), because it is the most precise delegation architecture in the Bible. Notice the master’s method. He entrusts defined sums — five, two, one — “to each according to his ability”; no servant receives the whole estate. He gives no micromanagement and expects initiative; the servants trade freely. He goes away — real autonomy, real distance. And then the hinge of the story: “after a long time the master of those servants came and settled accounts with them.” Entrustment, initiative, absence, audit. The servant condemned is not one who lost money trading but the one who refused the loop entirely — who buried the entrustment to keep his own accounting clean. The parable damns both failure modes of the agentic age: reckless delegation without accounting, and fearful refusal to delegate at all.

Translated into an operating discipline, I teach this as the Steward’s Charter — a one-page written constitution for every agent you deploy, with five clauses:

1. Mandate. What this agent exists to do, stated narrowly: answer product questions from the approved catalog; reconcile invoices under a threshold; draft — never send — outreach. The talents were counted before the journey began. An agent without a written mandate has your entire authority by default, which is the one thing you must never give.

2. Means. What it may touch and spend: which systems, which data, what budget ceiling, which tools. Joseph ran everything in Potiphar’s house except what was withheld (Genesis 39:9) — the withholding defined the stewardship. Caps on spend and access are the modern withholding.

3. Boundaries. What it must escalate to a human, listed explicitly: refunds above X, any legal or medical question, any angry customer, anything irreversible, anything touching a person’s standing or soul. This is Jethro’s architecture — routine cases settled low, hard cases brought up (Exodus 18:22). The boundary list is where your theology becomes configuration.

4. Review. When and how you settle accounts: a daily skim of logs, a weekly sampled audit of outputs, a monthly review of the mandate itself. The master came back — review is not distrust of the agent but fidelity to the loop. Unreviewed delegation decays into abdication on a schedule set by the error rate, and with agents failing multi-step tasks at the rates they currently do (5), the schedule is short.

5. Accountability. Whose name answers for this agent — written down, one human, no committee. The Lausanne audit question (“is it obvious who is accountable?”) deserves a literal answer in your operations doc (3). When the agent errs — it will — that person owns the correction, the apology, and the repair, as publicly as the error. This clause is the whole theology in miniature: execution was delegated; accountability never left.

The Charter scales down as well as up — a Kampala trader running a commerce agent on mobile-money rails needs the five clauses as surely as an enterprise does, and can write them in an afternoon. And it composes with rest: guardrails written into the agent’s constitution are what make it possible to keep Sabbath while your systems keep running. The steward who governs his agents well is freed by them; the one who does not is eventually owned by them.

What Must Never Be Delegated?

Three categories, and the line is not about capability but constitution — these are acts that are you, not tasks you perform.

Repentance. No agent can confess your sins, because sin attaches to persons, not processes. The steward whose agent deceived a customer cannot dispatch the agent to be sorry; he repents, he makes restitution, he seeks the face of the wronged neighbor and of God. Psalm 51 has no API. Any workflow that routes moral repair through the machine — auto-generated apologies for harms never owned — is not efficiency; it is the hardening of a heart by outsourcing its contrition.

Worship. Prayer, praise, and communion are covenant speech between persons; a generated prayer can prime your words the way a hymnal does, but it cannot do your praying, because worship is not content production — it is you, offered. “My son, give me your heart” (Proverbs 23:26) is the one delegation request in Scripture that admits no proxy.

Final moral judgment over people. Hiring and firing, discipline and restoration, the weighing of a soul’s circumstances — agents may assemble facts, sort queues, and flag patterns, but the judgment that lands on a human being must come from a human being who can stand in the same rain. This is why the shepherding office is non-transferable — AI can triage, but only a shepherd can shepherd — and why the autonomous-weapons debate matters far beyond the battlefield: once we accept that machines may render final judgment on persons in the extreme case, we will accept it in the mundane ones, one HR workflow at a time (9). Refuse the premise everywhere. Let the agent prepare the file; never let it pronounce the sentence.

Here is the hope under the warning. The agentic age does not shrink human significance; it clarifies it. When execution becomes cheap, what remains scarce is exactly what Genesis said was distinctive: responsible persons who answer to God for the work of their hands and their tools. The founder who masters delegated authority — bold entrustment, written charters, faithful accounting, unbreached limits — will run circles around both the reckless and the fearful. And he will hear, about his silicon stewards and his own stewardship alike, the only performance review that has ever mattered: “Well done, good and faithful servant. You have been faithful over a little; I will set you over much” (Matthew 25:21).

FAQ

Can a Christian delegate real decisions to AI agents?
Yes — execution decisions within written limits, following Scripture’s own delegation pattern of defined entrustment and scheduled accounting (Matthew 25; Exodus 18). What cannot transfer is moral agency and accountability: the Lausanne framework states believers can never assign moral agency to machines, and the human principal answers for every agent action (3).

Who is responsible when an AI agent makes a mistake?
The person and organization that deployed it. A tribunal ordered Air Canada to pay damages for its chatbot’s misinformation, rejecting the claim that the bot was a separate legal entity. Law and theology agree: authority flows down, accountability flows up — the principal answers (4).

How error-prone are AI agents today?
Materially. Simulations show LLM-driven agents failing multi-step office tasks at rates near 70%, about 80% of organizations report agent misbehavior, and Gartner projects over 40% of agentic AI projects cancelled by 2027 — which makes review rhythms mandatory, not optional (2)(5)(6).

What is the Steward’s Charter?
A one-page written constitution for each deployed agent, with five clauses: Mandate (narrow purpose), Means (systems, data, and spend limits), Boundaries (explicit escalation list), Review (daily, weekly, and monthly accounting rhythms), and Accountability (one named human who answers for the agent’s actions).

What should never be delegated to AI?
Three things, because they are personal acts before God rather than tasks: repentance (no agent can confess or repair your sins), worship (prayer and praise are you, offered), and final moral judgment over people — hiring, firing, discipline, and shepherding verdicts must come from accountable humans.

Related Reading

• Sabbath for the Agentic Age
• AI Can Triage; Only a Shepherd Can Shepherd
• Does AI Pay Rent? The SME Evidence
• Agentic Commerce on M-PESA and Mobile Money Rails

Sources and Evidence

1. Gartner — “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026” — Primary analyst projection on enterprise agent adoption.
2. Gartner — “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” — Analyst risk assessment citing costs, unclear value, and inadequate risk controls; companion Hype Cycle projects 15% of work decisions autonomous by 2028.
3. Lausanne Movement — “AI Through the Lens: Evaluating AI design and use in alignment with biblical ethics” — Global evangelical ethical framework; moral agency non-assignable to machines; accountability audit questions.
4. McCarthy Tétrault — “Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot” — Law-firm analysis of the BC Civil Resolution Tribunal ruling; corroborated by Pinsent Masons.
5. Quantum Zeitgeist — “AI Agents Fail Half The Time: New Benchmark Reveals Weaknesses” — Benchmark reporting ~50% average task completion; simulated-office multi-step failure rates near 70% reported across 2025 agent-reliability studies including arXiv: Towards a Science of AI Agent Reliability.
6. Atlan — “AI Agent Hallucination: Causes, Risks & Context Solutions” — Industry analysis including the finding that ~80% of organizations report agent misbehavior (hallucination, data leakage, unauthorized access).
7. C.S. Lewis Institute — “Artificial Intelligence, Imago Dei, Stewardship and the Mission of the Church” — Evangelical stewardship treatment of AI as mandated cultivation under moral limits.
8. Thirdwell — “AI and Moral Responsibility” — Ethics essay: delegating morally significant decisions to AI heightens rather than absolves human responsibility.
9. The Catholic University of America — “Autonomous Weapons vs. Moral Agents” — Theological analysis of judgments that must never be delegated regardless of machine competence.
10. Vatican Dicastery for the Doctrine of the Faith — Antiqua et Nova (January 2025) — Doctrinal note on AI and human intelligence; responsibility remains with human persons across all AI deployment.