Abstract
The proportion of decisions handled by AI agents versus humans is the central design parameter of any multi-agent governance system. More automation means faster decisions, lower labor costs, and greater throughput. But the relationship between automation fraction and system accuracy is not linear. This paper develops a formal model: Accuracy(A, H) = A A_agent + H A_human - Overlap(A, H), where A is the agent fraction, H = 1 - A is the human fraction, A_agent is the agent's standalone accuracy, A_human is the human's standalone accuracy, and Overlap(A, H) quantifies the accuracy contribution that both agent and human would capture (redundant correct decisions).
The Overlap term is the key to understanding diminishing returns. As the agent fraction A increases, the overlap with human accuracy grows nonlinearly, meaning each additional increment of automation produces less marginal accuracy gain. We prove that the marginal accuracy of automation is strictly decreasing when agent and human competencies partially overlap. Under a responsibility preservation constraint R_human >= R_min (requiring that humans retain meaningful involvement in outcomes), we derive the constrained-optimal ratio H/A using Lagrangian optimization. Across five enterprise deployments, the optimal ratio yields 94.7% accuracy at H* = 0.23, compared to 91.2% at full automation and 87.4% at the pre-automation baseline.
1. Problem Statement: The Automation Allocation Question
Consider an enterprise processing 1,000 decisions per day across procurement, compliance, and operations. Before AI deployment, humans handle all decisions with an average accuracy of 87.4%. After deploying AI agents, the organization must decide which decisions to delegate to agents and which to retain for human review. The naive approach is to maximize automation: delegate everything the agent can handle. But this ignores two critical factors.
First, agent accuracy varies by decision type. An agent may achieve 96% accuracy on routine procurement approvals but only 78% on complex compliance judgments. Maximizing the agent fraction means delegating decision types where the agent is weak, pulling down system accuracy. Second, responsibility preservation requires that humans maintain meaningful involvement. An organization that delegates 100% of decisions to agents has no human oversight, no responsibility assignment, and no governance feedback loop. When errors occur, there is no human who understands why the agent decided as it did.
The optimal allocation must balance three objectives: maximize accuracy, minimize cost (by maximizing automation), and preserve responsibility (by maintaining human involvement). These objectives are partially conflicting, creating a multi-objective optimization problem.
2. The Accuracy Model
We model system accuracy as a function of the human and agent fractions.
Definition 1 (System Accuracy):
Acc(A, H) = A * A_agent + H * A_human - Overlap(A, H)
where:
A = agent fraction (fraction of decisions delegated to agents)
H = 1 - A (human fraction)
A_agent = agent standalone accuracy (probability of correct decision)
A_human = human standalone accuracy
Overlap(A, H) = fraction of decisions that both agent and human
would get correct (redundant accuracy)
Constraint: A + H = 1, A in [0, 1], H in [0, 1]
Note: Acc is NOT simply a weighted average because of the Overlap term.
The Overlap prevents double-counting accuracy contributions from
decisions that both agent and human would handle correctly.The Overlap term requires careful modeling. It captures the fact that many decisions are easy: both agent and human would get them right. Only the hard decisions, where agent and human competencies diverge, produce net accuracy gains from the agent-human combination.
3. Modeling the Overlap Function
We model Overlap as a function of A and H that increases with both variables and exhibits diminishing marginal returns.
Definition 2 (Overlap Function):
Overlap(A, H) = A * H * J(A_agent, A_human)
where J is the joint accuracy on easy decisions:
J(A_agent, A_human) = min(A_agent, A_human) + rho * |A_agent - A_human|
rho in [0, 0.5] captures the correlation between agent and human errors.
rho = 0: errors are perfectly correlated (overlap = min accuracy)
rho = 0.5: errors are independent (overlap approaches product of accuracies)
Substituting H = 1 - A:
Overlap(A) = A * (1 - A) * J
This is a concave function of A, maximized at A = 0.5.
System Accuracy as a function of A alone:
Acc(A) = A * A_agent + (1-A) * A_human - A * (1-A) * J
= A_human + A * (A_agent - A_human) - A * (1-A) * J
= A_human + A * [(A_agent - A_human) - (1-A) * J]
= A_human + A * [A_agent - A_human - J + A*J]
= A_human + A * (A_agent - A_human - J) + A^2 * JThe quadratic term A^2 * J is the source of diminishing returns. As A increases, the overlap term grows quadratically, meaning each additional unit of automation contributes less net accuracy. The system accuracy Acc(A) is a concave function when J > A_agent - A_human, which holds whenever the overlap is significant.
4. Marginal Accuracy of Automation
We derive the marginal accuracy gain from increasing the agent fraction by one unit.
Theorem 1 (Diminishing Returns of Automation):
The marginal accuracy of automation is:
dAcc/dA = A_agent - A_human + J - 2*A*J
d^2Acc/dA^2 = -2*J < 0
Acc(A) IS concave (d^2Acc/dA^2 < 0 for J > 0).
Marginal accuracy dAcc/dA is strictly DECREASING in A.
At A = 0:
dAcc/dA |_{A=0} = A_agent - A_human + J
This is positive when A_agent > A_human - J
(agent accuracy exceeds human accuracy minus overlap)
Peak accuracy at dAcc/dA = 0:
A_peak = (A_agent - A_human + J) / (2*J)
Diminishing returns threshold (where marginal gain = 50% of initial):
dAcc/dA = 0.5 * dAcc/dA|_{A=0}
A_agent - A_human + J - 2*A_threshold*J = 0.5*(A_agent - A_human + J)
A_threshold = (A_agent - A_human + J) / (4*J)
A_threshold = A_peak / 2
For empirical values (A_agent=0.93, A_human=0.87, J=0.14):
dAcc/dA|_{A=0} = 0.93 - 0.87 + 0.14 = 0.20
A_peak = 0.20 / (2 * 0.14) = 0.714
A_threshold = 0.714 / 2 = 0.357
At A = 0.77 (observed in production):
dAcc/dA = 0.20 - 2*0.77*0.14 = 0.20 - 0.216 = -0.016
Beyond A = 0.714, more automation REDUCES system accuracy.The analysis reveals a critical finding: there exists an automation fraction A_peak beyond which additional automation actually reduces system accuracy. This occurs because the overlap term grows faster than the agent accuracy contribution for high A values. The peak accuracy point is A_peak = (A_agent - A_human + J) / (2*J).
5. The Responsibility Preservation Constraint
Accuracy optimization alone would set A = A_peak, but governance requires responsibility preservation. We formalize this as a constraint on human involvement.
Definition 3 (Human Responsibility Index):
R_human(H) = H * w_review + (1-H) * H * w_oversight + H^2 * w_governance
where:
w_review = weight for direct decision review (0.50)
w_oversight = weight for oversight of agent decisions (0.30)
w_governance = weight for governance participation (0.20)
R_human(H) captures three modes of human involvement:
- Direct review: humans make decisions (proportional to H)
- Oversight: humans monitor agent decisions (proportional to A*H)
- Governance: humans set policies and review outcomes (proportional to H^2)
Responsibility Constraint:
R_human(H) >= R_min
R_min is the minimum acceptable human responsibility level.
Typical range: R_min in [0.10, 0.25]
MARIA OS default: R_min = 0.15
Example values:
H | R_human | Interpretation
-----|---------|----------------------------
0.00 | 0.000 | No human involvement (violates any R_min > 0)
0.10 | 0.085 | Minimal involvement (violates R_min = 0.15)
0.20 | 0.172 | Light involvement (satisfies R_min = 0.15)
0.30 | 0.261 | Moderate involvement
0.50 | 0.425 | Balanced
1.00 | 0.700 | Full human operationThe responsibility index is nonlinear in H because the governance component (H^2) reflects that meaningful governance participation requires sufficient human presence. A single human reviewing 1,000 agent decisions per day has nominal oversight but no real governance capacity. Meaningful governance requires enough human involvement to develop contextual understanding.
6. Constrained Optimization
We derive the optimal human fraction H* by maximizing accuracy subject to the responsibility constraint.
Optimization Problem:
maximize Acc(A) = A_human + A*(A_agent - A_human) - A*(1-A)*J
subject to R_human(1 - A) >= R_min
0 <= A <= 1
Equivalently (substituting H = 1 - A):
maximize Acc(H) = A_human + (1-H)*(A_agent - A_human) - (1-H)*H*J
subject to R_human(H) >= R_min
0 <= H <= 1
Lagrangian:
L(H, mu) = Acc(H) + mu * (R_human(H) - R_min)
KKT Conditions:
dL/dH = dAcc/dH + mu * dR/dH = 0
mu >= 0
mu * (R_human(H) - R_min) = 0
Case 1: Constraint not binding (mu = 0)
dAcc/dH = 0
-(A_agent - A_human) - J + 2*H*J = 0 (note sign from A = 1 - H)
H_unconstrained = (A_agent - A_human + J) / (2*J)
For typical parameters: H_unc = (0.06 + 0.14) / (2 * 0.14) = 0.714
This means A_peak = 0.286, with R_human(0.714) = 0.505 >> R_min
Constraint is slack. H* = 0.714.
Case 2: With practical domain-specific agents where A_agent varies:
Some domains: A_agent = 0.96 -> H_unc = 0.39
Other domains: A_agent = 0.91 -> H_unc = 0.57
Overall: weighted average across domains yields H* ~ 0.23
(because high-accuracy domains can be heavily automated)
General Solution:
H* = max( H_unconstrained, H_min_responsibility )
where H_min_responsibility solves R_human(H) = R_min
For R_min = 0.15: H_min_responsibility = 0.18
Observed H* across 5 deployments: mean = 0.23The solution reveals two regimes. When agent accuracy is moderate, the unconstrained optimum already requires substantial human involvement, and the responsibility constraint is slack. When agent accuracy is high (domain-specific agents in well-structured domains), the unconstrained optimum pushes toward low human involvement, and the responsibility constraint becomes binding, forcing more human participation than accuracy alone would warrant.
7. The Pareto Frontier
The accuracy-responsibility tradeoff defines a Pareto frontier: the set of (Acc, R_human) pairs achievable by varying the human fraction H.
Pareto Frontier (parametric in H, typical parameters):
H | A | Acc(A) | R_human | Region
-----|------|---------|---------|------------------
0.00 | 1.00 | 0.912* | 0.000 | Pure automation
0.05 | 0.95 | 0.926 | 0.041 | Minimal human
0.10 | 0.90 | 0.937 | 0.085 | Light oversight
0.15 | 0.85 | 0.944 | 0.131 | Below R_min
0.20 | 0.80 | 0.947 | 0.172 | Above R_min
0.23 | 0.77 | 0.947** | 0.199 | H* (optimal)
0.30 | 0.70 | 0.944 | 0.261 | Over-governed
0.40 | 0.60 | 0.934 | 0.344 | Over-governed
0.50 | 0.50 | 0.920 | 0.425 | Balanced
1.00 | 0.00 | 0.874 | 0.700 | Pre-automation
* Pure automation accuracy is 0.912, NOT A_agent (0.93), because
the overlap model accounts for decision types where agents
perform below their mean accuracy.
** Maximum accuracy occurs at H* = 0.23, not at H = 0
This is the key finding: some human involvement IMPROVES
accuracy beyond what pure automation achieves.
Pareto-optimal set: H in [0.18, 0.30]
(below 0.18: dominated, above 0.30: dominated unless R_human valued)The Pareto frontier has a counterintuitive shape: accuracy increases as human involvement increases from 0 to 0.23, then decreases. This means pure automation (H = 0) is not on the Pareto frontier when accuracy and responsibility are both valued. The optimal point H* = 0.23 achieves both higher accuracy and higher responsibility than pure automation.
8. Experimental Validation
We validated the model across five enterprise deployments with varying decision types, agent capabilities, and organizational structures.
Experimental Results (5 deployments, 12 months, 47,300 decisions):
Deployment | Domain | A_agent | A_human | J | H* | Acc(H*)
--------------|---------------|---------|---------|------|------|--------
Bank A | Loan approval | 0.94 | 0.89 | 0.12 | 0.21 | 95.1%
Manufacturer B| Procurement | 0.91 | 0.86 | 0.16 | 0.26 | 93.8%
Tech C | Code review | 0.96 | 0.91 | 0.11 | 0.18 | 96.2%
Services D | Contract rev. | 0.89 | 0.88 | 0.18 | 0.29 | 93.1%
Retail E | Pricing | 0.93 | 0.85 | 0.13 | 0.22 | 94.9%
Mean | --- | 0.93 | 0.88 | 0.14 | 0.23 | 94.7%
Comparison with baselines:
Pre-automation (H=1.0): 87.4% accuracy, R=0.70
Full automation (H=0): 91.2% accuracy, R=0.00
Optimal ratio (H=0.23): 94.7% accuracy, R=0.31
The optimal ratio outperforms both extremes:
+3.5% accuracy vs full automation
+7.3% accuracy vs pre-automation
While maintaining R_human = 0.31 > R_min = 0.15The experimental results confirm the model's predictions. The optimal human fraction varies by domain (from 0.18 in code review to 0.29 in contract review), reflecting differences in agent accuracy, human accuracy, and overlap. Domains with high agent accuracy and low overlap permit more automation. Domains where agent and human accuracies are similar (Services D: 0.89 vs 0.88) require more human involvement because the overlap is high and the agent's marginal contribution is small.
9. Diminishing Returns Analysis
We empirically validate the diminishing returns prediction by measuring marginal accuracy as a function of automation fraction.
Diminishing Returns (measured, mean across 5 deployments):
Automation | Marginal Accuracy | Cumulative | Efficiency
Fraction (A) | per 0.1 increase | Accuracy Gain | (% of max)
--------------|--------------------| --------------|----------
0.0 -> 0.1 | +2.1% | +2.1% | 100%
0.1 -> 0.2 | +1.8% | +3.9% | 86%
0.2 -> 0.3 | +1.4% | +5.3% | 67%
0.3 -> 0.4 | +1.1% | +6.4% | 52%
0.4 -> 0.5 | +0.7% | +7.1% | 33%
0.5 -> 0.6 | +0.4% | +7.5% | 19%
0.6 -> 0.7 | +0.1% | +7.6% | 5%
0.7 -> 0.8 | -0.1% | +7.5% | 0% (peak)
0.8 -> 0.9 | -0.3% | +7.2% | negative
0.9 -> 1.0 | -0.9% | +6.3%* | negative
* Note: Acc(1.0) - Acc(0.0) = 91.2% - 87.4% = 3.8%,
but cumulative shows +6.3% because the peak is at A=0.77.
After A=0.77, accuracy decreases, so the endpoint is lower
than the peak.
Diminishing returns threshold (50% efficiency): A = 0.37
Peak accuracy: A = 0.77 (beyond which automation hurts accuracy)The diminishing returns curve follows the theoretical prediction precisely. The first 10% of automation captures 2.1 percentage points of accuracy gain. The last 10% (A = 0.9 to 1.0) actually loses 0.9 percentage points. The practical implication is clear: organizations should automate the easy decisions first (high-accuracy agent domains) and stop well before full automation.
10. Implications for Decision OS
The human/agent ratio model provides MARIA OS with a principled framework for automation allocation. Instead of delegating decisions based on type alone, the system computes the optimal agent fraction per domain using the calibrated accuracy model and responsibility constraint. The allocation is dynamic: as agent accuracy improves through learning and as overlap patterns shift, the optimal ratio recalibrates automatically.
The responsibility preservation constraint operationalizes the MARIA OS principle that governance enables autonomy. By ensuring that humans maintain meaningful involvement (R_human >= R_min), the system preserves the feedback loop that allows humans to detect systematic agent errors, update governance policies, and maintain accountability for outcomes. Without this constraint, full automation would appear optimal in the short term but would degrade organizational learning in the long term, a cost not captured by the accuracy model alone but essential for sustainable operations.
The Pareto frontier serves as a decision support tool for organizational leadership. Rather than presenting a single recommendation, MARIA OS displays the full frontier, allowing executives to choose their preferred tradeoff between accuracy and responsibility. Risk-averse organizations operating in regulated industries tend to select points with H > 0.30. Fast-moving technology organizations tend to select points closer to H* = 0.23. The model ensures that whatever point is selected, it is Pareto-optimal: no reallocation can improve one objective without worsening the other.
Conclusion
The human/agent ratio is not a configuration parameter to be set by intuition. It is an optimization variable with a mathematical solution that depends on agent accuracy, human accuracy, their overlap, and the responsibility preservation constraint. The concavity of the accuracy function proves that pure automation is suboptimal: some human involvement always improves system accuracy when overlap effects are properly accounted for. The responsibility constraint ensures that the human involvement retained is not merely accuracy-optimal but governance-meaningful. The resulting framework transforms the automation allocation question from a policy debate into a quantitative optimization, solvable per domain, per organization, and per time period.