Where to invest in AI agents: start with the verifiable bottleneck
By Equipo Quantum Developers
Summarize:
An initial AI-agent investment is defensible only when it releases an observed system constraint and the end outcome improves afterward; automating work away from the bottleneck can create activity without value. If a team processes more documents while final approval, the dock, or the exception queue still limits flow, the automation increased local output rather than system capacity.
Industry is not a prioritization unit
“Agents for finance” and “agents for logistics” are marketing categories that are too broad for investment decisions. Within one company, finance may be constrained by master data, logistics by late decisions, and procurement by approval. Copying a popular case from another organization does not establish the local constraint.
The useful unit is the journey of an object to an outcome: an invoice paid correctly, an order released, or a shipment exception resolved. The GOV.UK Service Standard advises teams to design around a user’s whole problem rather than a preselected technology and to improve it through connected increments. That logic prevents optimization of an isolated task merely because it “uses AI.”
Draw queues, not only activities
The minimum flow map includes:
- object and final outcome;
- steps that change its state;
- input and output by period;
- queue before each step and its age;
- cases returning or requiring exception;
- resource or decision limiting the step;
- owner and closure evidence.
Use a representative window and segment the data. A daily average can hide one currency, customer, or exception type that dominates delay. Do not call the most annoying step the bottleneck; identify the step constraining the system outcome.
Four-question causal test
Before proposing an agent, answer:
- Constraint: Which queue or decision limits the outcome, and how was it observed?
- Mechanism: Which agent action changes capacity, quality, or time at that point?
- Displacement: If it works, where will the next constraint appear?
- Outcome: Which end metric should change in which comparable population?
A proposal that says only “reduce manual work” does not pass. It must explain why that work constrains the result. If released people cannot operate the bottleneck, their hours do not become flow.
A bottleneck-first portfolio method
1. Select a journey, not a department
Choose an owned outcome: close a reconciliation, release an order, or resolve a shipment exception. Define start, finish, and population.
2. Locate the constraint with evidence
Observe queue, age, capacity, rework, and blocked time. Validate findings with process operators. An integration failing once a month may have high impact without limiting the daily flow.
3. Generate three alternatives
Compare a rule or data improvement, deterministic automation, and an assistive agent. Add “no change” as baseline. The NIST AI RMF Core asks organizations to manage risk according to priorities and define responsibilities; it does not require AI.
4. Design the smallest experiment
Restrict population and authority. If the constraint is ambiguous classification, the agent can propose with evidence while a person decides. Measure the journey outcome, not generated-response count.
5. Recalculate after the change
When the queue falls, another step may become limiting. Pause automatic expansion and observe again. An agent that was useful can stop being the priority.
Candidate comparison card
| Field | Required evidence |
|---|---|
| Constrained outcome | final metric, population, and owner |
| Queue or decision | volume, age, and segment variation |
| Controllable cause | data, rule, capacity, or coordination |
| Proposed action | classify, retrieve, recommend, or execute |
| Authority | boundary, approval, and reverse path |
| Expected change | causal link to the outcome |
| Full cost | change, operation, review, and exception |
| Kill criterion | signal that invalidates the hypothesis |
The GAO Cost Estimating and Assessment Guide emphasizes a technical baseline, assumptions, data, analysis of alternatives, sensitivity, and updates with actual costs. Record these before comparing cases; otherwise each sponsor uses a different definition of benefit.
Illustrative example: three ideas, one constraint
This example is illustrative and not a customer case. A distributor observes that orders wait primarily in credit review. Three ideas compete: draft sales emails, forecast demand, and prepare credit-review files.
The third touches the constraint, but that does not authorize automatic approval. The experiment gathers permitted documents, checks completeness, and proposes a reason while the analyst keeps the decision. The primary metric is eligible-order age to decision at constant quality. Email drafts or forecasts may be useful, but they cannot claim the same outcome while review still limits release.
If the queue falls and dispatch becomes the next wait, the next budget is evaluated there. Continuing to expand the credit agent by inertia would again optimize outside the constraint.
Sustaining the portfolio in Quantum
Quantum Automation Center can catalog the candidate, owner, object, and state; relate executions to a timeline; retain permitted artifacts and logs; and display operational or financial analytics. Use those surfaces to maintain hypothesis, evidence, and continuation decision. Permissions and human approval constrain authority during the test. See Automation Center and AI agents.
Do not turn active-agent count into a KPI. The dashboard should show which constraint was targeted, what changed, and where the constraint moved.
Counterargument: the present bottleneck can trap strategy
Companies also need options for future change. A new channel, regulation, or product may require capability that does not constrain today’s flow. If every budget follows the current queue, those options never mature.
Separate operating and exploration lanes. The operating lane requires an observed constraint and outcome. The exploration lane funds learning with scope, budget, and stop condition; it promises no current ROI. When an option finds a journey and evidence, it competes in the operating portfolio.
When not to use this method
Do not use it to decide whether to fulfill a legal, security, or continuity obligation. Such work may be necessary even when it is not the bottleneck. Do not use it when only one task is visible rather than the full journey; cost may simply move to another team.
When the underlying problem is demand, strategy, or authority, an agent cannot create the outcome. Resolve the organizational decision first. When no baseline exists, invest in instrumentation before ranking opportunities.
The right investment question
Do not ask, “Which agent does this industry need?” Ask, “What prevents completion of the outcome today, what evidence proves it, and what is the smallest reversible intervention?” Remap after every change. A useful portfolio follows movement in the system rather than popularity of a technology.
Sources
Article topics
