A damaged parcel must be recognised in a photograph. Tomorrow's demand must be estimated. A delivery route must be optimised. An email must be understood and a reply drafted. A CRM record must be updated.
All five tasks may appear in one “AI automation” project, but they are not the same technical problem. A language model can help with some of them. It is not automatically the best engine for all of them—and it does not determine how much authority the complete system should receive.
Two decisions must therefore be made separately:
- What kind of work must the system perform? Perceive, predict, choose, generate or act?
- How far may its output travel without a person approving it? From an isolated personal tool to a continuously operating, bounded control loop?
This distinction prevents two common design errors: using an LLM where a rule or specialised model is more reliable, and granting operational authority merely because a model can produce a plausible answer.
The short answer
Start with the smallest combination that solves the work unit:
| The uncertain part of the work | Useful starting component | Keep outside the model |
|---|---|---|
| A known event must trigger a known action | deterministic rule or integration | exceptions and approval thresholds |
| An image, voice recording or document must be recognised | perception model | acceptance rules and exception handling |
| A future value or risk must be estimated | predictive model | the decision threshold and business response |
| The best option must be selected under constraints | optimisation engine | objectives, prohibited options and hard limits |
| Language must be interpreted or created | LLM or another language model | trusted facts, validation and consequential commitments |
| The route through tools changes with context | bounded agent architecture | permissions, budgets, stop conditions and recovery |
Then give the complete workflow the minimum authority required to create a measurable result. A better model is not, by itself, a reason to remove an approval.
One system, two independent axes
The first axis describes the system's functional role. The second describes its operating authority.
They must not be collapsed into one maturity score. A sophisticated forecasting model can have no authority to act. A simple optimisation model can operate in a closed industrial loop. An LLM may be either a private writing aid or one component inside a bounded agent.
Functional role: perceive → predict → select → understand/create → act
Authority: A0 rules → A1 personal → A2 copilot → A3 embedded
→ A4 bounded agent → A5 controlled loop
The arrows do not mean that every project should move to the right. They are a map of different responsibilities, not a universal upgrade path.
Axis one: what kind of intelligence does the work require?
1. Deterministic rules establish the reliable baseline
A rule engine is not AI, but it belongs on the map because many “AI” requests are ordinary automation problems:
- copy an approved value between systems;
- check that required fields exist;
- calculate a price from a fixed tariff;
- route an order by region;
- stop a payment above a threshold;
- send a notification after a known event.
If inputs and rules are stable, code is cheaper to test, easier to audit and more predictable than a probabilistic model.
2. Perception converts the physical or unstructured world into signals
Computer vision, speech recognition and document extraction answer questions such as: what is present, what was said and which fields are on this form?
BMW Group describes a 2025 pilot at its Regensburg plant that prepares tailored quality-inspection recommendations for roughly 1,400 vehicles manufactured each day. This is a company-reported deployment, not independent proof of ROI. Its architectural lesson is more useful than the scale: the model helps decide where to inspect, while trained specialists still perform the final quality inspection.
3. Prediction estimates what may happen next
Forecasting demand, failure probability, delivery time, churn or fraud risk is a predictive task. Its output is an estimate with uncertainty—not a business decision.
The process still needs a threshold and an owner. A 70% late-delivery risk might create an internal task; it should not automatically cancel a customer order unless that policy was separately designed and tested.
4. Selection and optimisation choose under constraints
Routing, scheduling, stock allocation and capacity planning often depend more on mathematical optimisation than on language generation.
UPS's 2014 sustainability report described ORION as combining operational data, routing algorithms and custom maps. At full deployment, UPS expected it to reduce driving by 100 million miles and fuel use by 10 million gallons annually. Those are historical company projections, not a Methodfield result or a current universal benchmark. The pattern remains instructive: define the objective and constraints, then optimise the choices.
5. Language models interpret and create
LLMs are useful when meaning matters and the input or output is language:
- classify an enquiry written in an unfamiliar form;
- extract requirements from a conversation;
- compare a document with an approved policy;
- retrieve relevant knowledge and prepare a cited answer;
- draft a summary, proposal or response;
- translate while preserving defined terminology.
They are poor substitutes for an exact calculator, a permission check or a source-of-truth database. A confident sentence is not a confirmed fact.
6. An agent is an architecture, not another kind of model
An agent combines a model with instructions, tools, state and a loop that can choose the next step. Its defining property is not eloquent text. It is the ability to select and invoke actions in response to context.
That makes agent design an authority problem. The companion analysis Does this process need an AI agent? helps choose between rules, an assistant and an agent. Agent authority and safe control covers identity, permissions, approval, audit and recovery.
Axis two: how much authority should the workflow receive?
The following A0–A5 bands are a Methodfield design aid, not an industry standard or legal classification. The letter “A” means authority. It also keeps this map distinct from product-maturity levels used in other Methodfield materials.
| Band | Operating mode | What the system may do | Required control |
|---|---|---|---|
| A0 | deterministic automation | execute predefined rules | tests, logs and rollback |
| A1 | personal AI tool | produce content outside the operational record | user checks before use |
| A2 | copilot | recommend or draft inside a work interface | a person accepts, edits or rejects |
| A3 | process-embedded AI | move a case within an internal workflow after validation | schemas, thresholds, sampling and exception queues |
| A4 | bounded agent | choose tools and perform permitted, reversible actions | least privilege, budgets, approvals, audit and stop controls |
| A5 | controlled autonomous loop | observe, decide and act continuously within a narrow domain | independent safety constraints, failover, monitoring and human override |
Maximum authority is not the objective. The correct band is the lowest one that removes the measured bottleneck without making the cost of error unacceptable.
Evidence changes with the task, not the label
A peer-reviewed study of 5,172 customer-support agents found that an LLM-based assistant increased issues resolved per hour by 15% on average; gains were larger for less-skilled workers. This supports A2 assistance in that specific task and organisation—not a claim that every copilot improves every worker.
In a different 2025 randomised study, METR found that 16 experienced open-source developers completing 246 tasks took 19% longer with the AI tools available at the time, even though they expected to be faster. The studies do not contradict each other. They show why the task, user, tool and measurement must remain attached to any productivity claim.
A selection matrix for a real workflow
Classify the work unit before selecting a product.
| Question | Low-risk answer | Higher-control answer |
|---|---|---|
| Can the correct result be expressed as a rule? | use deterministic automation | if not, isolate the uncertain step |
| Is the model output easy to verify before use? | A1 or A2 may be sufficient | use stronger validation or keep the case manual |
| Does the output change money, access, commitments or a legal position? | require explicit approval | do not infer authority from model confidence |
| Is the action reversible? | a bounded automatic action may be tested | keep a person in the path and design recovery first |
| Can quality be measured from actual outcomes? | consider staged authority | remain in shadow or recommendation mode |
| Is an independent safety layer available? | a narrow control loop may be possible | do not use A5 |
For a more detailed promotion path, see From controlled automation to adaptive autonomy.
Example: a hybrid insurance-claim workflow
One process can use several technical types without turning one model into the entire system.
- A deterministic integration receives the claim and assigns an immutable ID.
- Document extraction reads the form, invoice and photographs.
- Computer vision marks visible damage; a predictive model estimates anomaly risk.
- An LLM compares the narrative with the extracted evidence and prepares a structured summary with source references.
- Rules verify required fields, policy dates, amounts and thresholds.
- Routine, low-value cases can enter an internal review queue automatically.
- A person decides disputed, high-value or low-confidence cases.
- The payment system executes only an approved instruction and records the final state.
The LLM handles language. It does not calculate the policy, grant itself payment permission or become the source of record.
Seven gates before authority grows
1. Work-unit gate
Name one input, one decision or transformation, one output and one owner. “Use AI in operations” is not a testable scope.
2. Evidence gate
Build a representative set with ordinary cases, rare exceptions and known bad inputs. Evaluate the complete workflow, not a polished demo.
3. Consequence gate
List what the output can change: money, customer promises, access, safety, employment, legal rights or public communication. Controls should follow the consequence, not the model name.
4. Permission gate
Give each component only the data and actions required for the current task. Separate read, draft, approve and execute permissions.
5. Observability gate
Record model and prompt versions, sources, tool calls, validation results, approvals, final actions and actual outcomes. Do not place sensitive content in general product analytics.
6. Recovery gate
Define timeouts, retry limits, idempotency, rollback, a manual route and the person who can stop the workflow.
7. Promotion gate
Increase authority only after agreed quality, exception, cost and recovery measures remain within limits. Reduce it when the process, data or model changes.
What controlled autonomy looks like
Google DeepMind's 2018 account of autonomous data-centre cooling illustrates an A5 pattern in a narrow domain. The cloud model proposed actions every five minutes; internal constraints and a separate local control system checked them before implementation. Operators could exit AI control at any time. Google reported that the system's energy-efficiency improvement grew from 12% to around 30% over nine months.
This is company-reported evidence from specialised infrastructure, not a target for general business automation. The transferable principle is that direct control was surrounded by independent constraints, failover and human override.
Responsibility does not transfer to the model
In Moffatt v Air Canada, a Canadian tribunal found Air Canada liable after its website chatbot provided incorrect information about a bereavement fare. The decision awarded C$812.02 in damages, interest and fees. It is one tribunal decision in one jurisdiction, not universal legal advice. The operational lesson is broader: a business remains responsible for automation presented through its customer channels.
In the EU, Article 50 transparency obligations under the AI Act begin to apply on 2 August 2026, with details and limited transitional provisions described in the European Commission's current guidance. This article's functional map and A0–A5 bands are not a compliance assessment. A production system still needs a case-specific legal, data-protection and risk review.
Measure outcomes by system type
Different components need different evidence:
| Component | Useful measures |
|---|---|
| perception | precision, recall, missed critical cases, extraction correction rate |
| prediction | calibration, false positives and negatives, decision value over baseline |
| optimisation | objective improvement, constraint violations, stability and computation cost |
| language model | factual support, task completion, edit rate, unsafe or unsupported output |
| agent | successful verified outcomes, tool errors, unnecessary steps, cost, intervention and recovery rate |
| complete workflow | cycle time, rework, exceptions, customer outcome, operating cost and incidents |
“Tokens used” and “agents launched” are operating details, not business outcomes.
A practical starting sequence
- Map one frequent workflow and its cost of error.
- Separate deterministic steps from uncertain cognitive work.
- Assign the uncertain step a functional role: perceive, predict, select or understand/create.
- Test that component in shadow mode on representative examples.
- Start at A2 when a person can verify the output efficiently.
- Add deterministic validation and an exception route before A3.
- Consider A4 only when the path must change with context and every tool is permission-bounded.
- Reserve A5 for narrow, continuously measured domains with independent safety controls.
The AI quality layer provides a practical model for regression tests, live monitoring and release gates around these systems.
Final position
The mature question is not “Which LLM should automate this process?” It is:
What must the system perceive, predict, choose, create or do—and what is the minimum authority required for that work?
Rules, specialised models, language models and agents are complementary components. Good architecture gives each one a defined role, keeps the system of record outside probabilistic output and expands authority only when evidence, controls and recovery are ready.
Sources
- NIST, AI Risk Management Framework resources (opens in a new tab), version reviewed 2 August 2026.
- BMW Group, Artificial intelligence as a quality booster (opens in a new tab), 28 April 2025. Company-reported deployment.
- UPS, 2014 Corporate Sustainability Report (opens in a new tab), historical company projections for ORION.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work (opens in a new tab), The Quarterly Journal of Economics, 2025.
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (opens in a new tab), 10 July 2025.
- Google DeepMind, Safety-first AI for autonomous data centre cooling and industrial control (opens in a new tab), 17 August 2018. Company-reported deployment.
- Civil Resolution Tribunal of British Columbia, Moffatt v Air Canada, 2024 BCCRT 149 (opens in a new tab), 14 February 2024.
- European Commission, Guidelines on transparency obligations for providers and deployers of AI systems (opens in a new tab), 20 July 2026.