Method
Seven disciplines that let AI agents scale.
Our method is designed to enable AI agent autonomy — and to grow that autonomy with each quarter of operation — with humans supervising. None of the seven disciplines is optional. Each seems, at first glance, like a constraint. But in practice, it's what allows clients to grant agents greater autonomy, without hesitation and with problems kept under control.
Operate within defined authority
What the agent may do, up to what limit, and who answers for it.
Measure what happened
Accuracy, human intervention, and cost per task.
Revalidate every quarter
An identified person decides, with the track record on the table.
Revise the authority
Raise, keep, or reduce.
↺ The cycle starts again, with the revised authority.
Discipline
The seven disciplines
- 01
The responsibility boundaries are written before the first line of code
Each project begins by clearly outlining the responsibilities of the customer versus ours for each AI agent. These boundaries are documented in a signed document that specifies liability limits, corrective actions such as re-running tasks or issuing credits, and a list of events that can suspend liability. Chief among them is configuration drift: a third party changing credentials, permissions, limits, integrations, or model versions without going through the governance forum. Telemetry detects such changes and suspends obligations until reconciliation. Additionally, Tech E&O insurance adds a final layer of protection against errors and omissions.
Why it matters
A Canadian court ruled that a company is accountable for the information its chatbot provides, dismissing the notion that the AI system functions independently. Responsibility for AI must be explicitly assigned, and its consequences thoughtfully managed.
- 02
Authority is delegated by positive enumeration
An agent can only carry out actions explicitly listed in the authority matrix. Anything not listed is blocked by system configuration, not by policy. Each authorized action has specific limits regarding value, volume, systems accessed, data provided, escalation conditions, and a designated responsible person. Human gates are not just reassurance but a responsibility layer: when a decision surpasses the AI agent's authority, it must be reviewed by an appointed individual with the proper mandate.
ProcessAgentContract annex, signedAction Value limit Volume limit Systems Data Escalate when Human gate Named person in charge Reconcile an entry Defined for this action Defined for this action Defined for this action Defined for this action Defined for this action Defined for this action Defined for this action Triage a claim Defined for this action Defined for this action Defined for this action Defined for this action Defined for this action Mandatory Defined for this action Check a document Defined for this action Defined for this action Defined for this action Defined for this action Defined for this action Mandatory Defined for this action Any action that is not on the list cannot be carried out — by configuration, not by policy Schematic illustration, not a screenshot. Action proposed by the agent
Reconcile, triage, check.
YesNoIs the action in the Matrix?
Blocked by configuration
Not by policy: the agent cannot do it.
NoYesWithin the authority limits?
Execute
Human gate
A named person decides before the action, never after.
ApproveExecute
DenyDo not execute
Evidence · every path leaves a trail, including the ones that end in no
What is not in the Matrix cannot be done. What goes past a limit stops and waits for a person's decision, before the action. And every path leaves a record, including the one that ends in no. Why it matters
Managed autonomy is already the accepted model for AI agents; what is missing is documenting it action by action, enabling AI agents to scale. With each revalidation, a successful record allows higher authority limits and reduces the need for human gates.
Only 11%automatically block agent actions that fall outside the authorized scopeCloud Security Alliance - 03
No measurement, no promise
No outcome is promised on a process whose baseline has not been measured and accepted in writing. Measuring the baseline means quantifying, in the current state and without an agent, the volume, cycle time, error rate, unit cost, and rework points. Without those five numbers, the contract records only a best-efforts commitment — and we say so in the proposal, not in the renegotiation.
And01A baseline measured and accepted in the contract
Comes from · Agentic Sprint Zero
And02At least one quarter of proven operation under AgentOps
Comes from · Managed AgentOps
03Contractual exposure caps
Comes from · Authority and Delegation Matrix
All three: a results obligation
Within the caps.
One missing: a best-efforts commitment
Stated in the contract, not implied.
All three must be in place at the same time. Each comes out of an earlier offering, which is why a results obligation comes in steps. Why it matters
This discipline sets apart those who really scale from those hesitant to implement AI agents. Moreover, it enables the sponsor to highlight to the executive committee the key message: “this process costs X less and resolves Y more than it did before — measured, not estimated.”
Less than 10%of AI-deployed use cases move beyond the pilot stageMcKinsey - 04
Human gates can't be approved retroactively
AI agents' processes progress through stages, many of them with formal human approval gates. Every stage begins only after the previous one is registered, and approvals cannot be granted retroactively. Post-facto approvals serve only as justifications, rather than controls. In legal contexts, these gates hinder approvers because they reveal that the control was merely formal and did not reflect the actual situation.
A gate Not a gate Blocking gateWithout approval, the action does not happen. Alert that can be ignoredA notification. Approval before the actionControl. Approval after the factJustification, and in litigation it weighs against whoever signed. Only 21%have a formal process for decommissioning agentsCloud Security Alliance - 05
The audit trail is generated straight from the operation, not as a subsequent audit project
The audit trail is generated as the AI agent works: inputs, decisions, authority invoked, data sources queried, human interventions, costs, and outcomes. It is not rebuilt in the future on demand from scattered logs.
- Input
- Decision
- Action
- Outcome
Unit-of-work record
Key: unit-of-work ID
- Inputs
- Decisions
- Authority invoked
- Data sources queried
- Human interventions
- Costs
- Outcomes
Stored in chained batches, signed with the client's key
Batch 1
Signed · hash of the previous one
Batch 2
Signed · hash of the previous one
Batch 3
Signed · hash of the previous one
- Next batch
- Changed? The chain flags it
The record is created while the agent works, and every item is bound to the same unit of work. It is not rebuilt later from scattered logs. Schematic illustration, not a screenshot. Why it matters
In Brazil, acceptance criteria are dictated by various standards, including: automated decision review as outlined in article 20 of the LGPD (Brazil's data protection law); registration in medical records and ongoing health audits per CFM Resolution 2.454/2026 (from Brazil's Federal Council of Medicine); cybersecurity policy and cloud contracting at financial institutions per CMN Resolution 4.893/2021 (from Brazil's National Monetary Council); and oversight by the ANPD (the national data protection authority). Companies whose operations generate audit trails as a by-product already produce the evidence these requirements call for — and gain an edge.
Only 10%maintain both audit trails for AI agents and links from the decisions back to the data sourcesModern Data Survey 2026 - 06
Portability is contracted, not promised
Open-format artifacts, interoperability protocols instead of proprietary integrations, an inventory of agents and exportable records, and operational documentation as contracted deliverables. The transfer commitment includes a term and a clause. The entire operation can be contracted under the Build-Operate-Transfer regime, in which AInda's exit, with a transition period, is a deliverable defined at the start of the project — not a concession negotiated at the end.
Why it matters
In a market where 71% say switching their main AI vendor or model would be difficult, a written-in-contract exit is an economic differentiator, not a gesture of goodwill.
71%say switching their main AI vendor or model would be difficult; 91% do not fully understand their dependenciesIBM Institute for Business Value - 07
Cost per task is measured and published, not estimated
Each unit of work carries four numbers: token cost per unit, infrastructure cost per unit, human intervention rate, and success rate. They are reported monthly, with a historical series, against the baseline.
Why it matters
In traditional IT, running a system costs a fraction of building it; in AI at scale, if poorly designed, the recurring cost can exceed the initial investment. A company that doesn't measure cost per task is not operating an AI agent team — it is only funding one.
−67% YoY, ~4,500× spreadtoken prices fell 67% in Q1 2026 and agentic cost still rose: an agent makes 5 to 30 calls per task, and pricing across models ranges from USD 0.04 to more than USD 180 per million tokensK2S46%of CFOs name cost uncertainty as their single biggest internal concern with AI useDeloitte CFO Signals Q2 2026
Gate
The project and process gates
Projects with AInda have ten formal gates. The first three are internal — client understanding, proposal design, and preparation for engagement. The seven gates described below include client input, approval, and control.
Before that, three internal gates: lead accepted, proposal released, and engagement opened
G3
Understand
Maturity and inventory known
G4
Specify
Authority matrix and acceptance criteria approved
G5
Plan
Execution plan approved
G6
Production
Evaluations, adversarial testing, human gate, and rollback tested
G7
Steady state
Thirty stable days within the SLA
G8
Revalidate
Every quarter, a person decides the authority
G9
Deactivate
Credentials revoked and records archived
↺ Every quarter: raise, keep, or lower the authority
| Gate | What must be ready |
|---|---|
| Understand | Authority matrices filled in and baselines defined. |
| Specify | Functional and evidence specifications approved. |
| Plan | Execution plans with acceptance criteria approved. |
| Production | Evaluations on target, human gates active, telemetry collecting, and costs per unit within cap. |
| Steady state | Thirty stable days, agreed SLAs, and operating runbooks delivered. |
| Revalidate | Quarterly review of authority, cost, and behavior. |
| Deactivate | Deactivations authorized, credentials revoked, and evidence trails archived. |
Agent deactivations are more important than they seem. Deactivating AI agents should be a routine operation, not an ad-hoc cleanup — and it's what frees up budgets and attention for the next processes to be transformed.
SLA
Behavior SLA — five contracted indicators
- 01Accuracy rate per process
- Correct outcomes over total processed.
- 02Human intervention rate
- Cases that required unplanned human action.
- 03Resolution time
- Average time from intake to case completion.
- 04Maximum cost per unit of work
- Cap per unit, measured monthly.
- 05Incident containment time
- From detection to stopping the effect.
90 days
From day one to day 91
01Know your AI and agent maturityDays 1–21
02Define process boundariesDays 15–45
03Agentic Sprint Zero in productionDays 30–75
04Continuous operation stateDays 60–90
Days 1–21
01Know your AI and agent maturity
Full census, a human owner per agent, sensitive-data exposure map, and the list of what to switch off, formalize, evolve, and scale.
Gate · Committee approves the target processes for Agentic Sprint Zero.
Days 15–45
02Define process boundaries
Authority and Delegation Matrix signed. Process baselines measured and accepted in writing, with metrics and boundaries.
Gate · Framing closed, with matrix and baselines.
Days 30–75
03Agentic Sprint Zero in production
Agentic Sprint Zero with deliberately narrow scope and real volume, human owners active, and telemetry collecting behaviors and costs from day one.
Gate · Production entry, verified against the four criteria.
Days 60–90
04Continuous operation state
Thirty measured days of operation, runbooks delivered, SLAs agreed, and the first evidence packs reviewed with internal audit.
Gate · Steady state. AgentOps can be contracted.
What you have on day 91
A managed team of AI agents, with clearly named human owners, engaged in active processes under delegated authority, handling actual operational volumes. Costs per unit of work are known and compared to previous records. An evidence pack was reviewed by internal audit. In addition, an informed decision on what to scale was made, based on solid numbers, not just expectations.
We do not sell absence of risk. We sell risk that is known, bounded, measured, and owned.