Agentic business is an operating model, not a product purchase
Most enterprises already own the models, the data and the platform. What they do not own is a defensible answer to four questions: what work may be delegated, how much authority the delegate holds, who answers when it is wrong, and how the organisation gets better at it. This practice builds that answer, then makes it operational in one workflow, in one quarter, with evidence.
Start from one workflow
Not a portfolio, not a strategy deck. One process with a measurable cost and a visible exception path.
Authority is designed, not assumed
Every delegated unit gets an explicit permission envelope and a named human who owns the consequence.
Measurement is shared
Agent outcomes and human outcomes sit on one table, reviewed by the same committee, on the same cycle.
Capability is left behind
Every engagement ends with internal roles able to build, evaluate and govern the next agent without us.
A machine colleague with known authority, measured like a colleague, revocable by a person
We use one definition across every engagement, and we hold ourselves to it in the deliverable. An enterprise is agentic when a material share of its operating decisions are made by a machine colleague with known authority, measured against the same standards as a human colleague, and reversible by a named person.
Three conditions have to hold together. If any one is missing, the organisation has automation with a chatbot on top — which is where most programmes sit after their first wave of pilots, and why the second wave stalls.
- Authority is explicit — the agent's mandate is written down: what it may decide, what it may commit, and what it must escalate. No written mandate, no delegation.
- Supervision is designed — a human reviews what the agent does, at a frequency set by risk, not by hope. Exceptions route to a person with the authority to decide, and the reason for escalation is recorded.
- Measurement is shared — the agent is scored on the same business measures the human team is scored on, and it also carries a confidence and escalation rate that a human role does not have.
Why this matters commercially. Delegation without a mandate is a liability with API access. Delegation with a mandate, supervision and measurement is an operating asset — it can be scaled, priced, audited and improved. Every governance and capability decision downstream in this site exists to make that second thing possible.
From tool adoption to delegation
Adoption asks
"Will people use this tool?" It is answered by enthusiasm, and it decays the moment the novelty does.
Delegation asks
"What may be handed over, under what limits, and who signs for the outcome?" It is answered by design.
Ownership asks
"Who is accountable when it is wrong, at 02:00 on a Sunday?" It is answered by a name, not a role.
One asset class, two populations
Our internal term for this discipline is Pure Lam: people and AI agents are built, measured, governed and developed as one class of operating asset. The leadership question is not "how many agents do we run" but "what is the return on the capability we hold, in both populations".
Agentic business in numbers
A tiered design decision, and the first thing a board should see
Where your organisation actually is, honestly assessed
Digital employees and human employees, measured differently
Who owns the agent portfolio, and who may decide what
Authority is a tiered design decision, and it is the first thing a board should see
A machine colleague with known authority, measured like a colleague, revocable by a person. Most failed programmes started by choosing the model before choosing the authority tier.
Authority is a tiered design decision, and it is the first thing a board should see
Most organisations have one vague answer to "how much may the agent do", applied to every agent in the estate. That is why escalations are unpredictable and why the incident review is always retrospective. We replace it with an explicit tier, a written mandate, and a supervisory rule attached to each one.
| Tier | What the agent does | Human involvement | What must exist first | Evidence produced |
|---|---|---|---|---|
| T0 — Advisory | Produces a recommendation with sources. Takes no action in any system of record. | Human decides and executes every time. The agent is a second pair of eyes. | Grounded retrieval with traceable citations, and a definition of what a good recommendation looks like. | Recommendation quality scored against the human decision that followed. |
| T1 — Decides within threshold | Decides and acts where the case falls inside a defined band — amount, confidence, queue age, customer segment. | Reviews the band weekly, handles out-of-band cases, owns the band definition. | A threshold model agreed with the business owner, plus a shadow period comparing decisions against human ones. | Threshold band, hit rate, misclassification rate, appeal outcomes. |
| T2 — Executes with post-audit | Completes the workflow end to end inside a bounded scope, including writing to systems of record. | Reviews a sampled percentage of completed work, calibrated to risk, and acts on exceptions. | Tool and write permissions scoped to the workflow, a rollback path, and a regression set that runs on every release. | Full execution trace: inputs, retrieved sources, actions, timestamps, final state. |
| T3 — Monitored autonomy | Runs the workflow with limited human involvement, raising only what its own confidence rules flag. | Holds a kill switch, receives an exception digest, and can revoke the mandate unilaterally. | Everything at T2, plus a mature evaluation set, a proven escalation precision rate and a tested continuity plan. | Continuous control monitoring, exception rate trend, quarterly re-authorisation record. |
Tier is set per work unit
Not per agent, and not per department. The same agent may sit at T0 for one unit and T2 for another, because the consequence of being wrong is not uniform across the work it touches.
Authority has an expiry
Every mandate carries a review date. A tier that is not re-authorised lapses to the level below it. This is what stops experimental authority from hardening into permanent, unreviewed autonomy.
Revocation is a button, not a process
One named person can suspend an agent's mandate immediately, without a change advisory board, without a ticket, and without the agent's consent. We test that path in the pilot, not after the incident.
Break the job into units that can actually be handed over
You cannot delegate a job description. You can delegate a unit of work with a defined input, a defined action set, a defined permission, a defined acceptance test and a defined escalation path.
Most delegation attempts fail on the third and fourth of those five. The unit is described too broadly, the acceptance test is "looks right", and the escalation path is a person trying to reconstruct intent from a log.
Our work in this step is deliberately unglamorous: two weeks inside the process with the people who do it, capturing what actually happens rather than what the process manual says happens. The output is a set of delegable units and an explicit list of the ones that are not delegable, with reasons.
Not everything is delegable. Deciding who to trust, committing the organisation to an external position, and any action where the consequence is not reversible at a known cost. Naming those explicitly protects the credibility of everything you do delegate.
Five fields, filled in before anything is built
| Field | What it must contain |
|---|---|
| Input contract | The trigger, the data the unit is entitled to see, and the states the unit may encounter. |
| Action set | The specific operations the agent may perform, and the systems it may write to. Enumerated, not described. |
| Permission envelope | Amounts, segments, time windows, regions and record types. The boundary is quantitative wherever the work permits it to be. |
| Acceptance test | How a completed unit is judged correct, expressed so it can be evaluated automatically and argued about by a human. |
| Escalation path | The trigger, the destination, the response expectation, and what the person receives in order to decide in under two minutes. |
Observe the work as performed
Not the documented process. The actual sequence, including the workarounds people invented.
Segment into units of decision
Split where a decision is made, not where a system changes.
Size each unit to one owner
A unit that no single person can explain is a unit nobody can delegate.
Score delegability
Volume, regularity, reversibility, judgment density. Regular and reversible work moves first.
Break the job into units that can actually be handed over
This is the consulting method in six steps. Everything else in the engagement is downstream of getting this decomposition right.
Name the outcome
One sentence, in business terms, that a finance director would accept as a definition of done.
Break it into decisions
Not tasks — decisions. A task can be automated; a decision carries judgement and therefore needs an authority tier.
Mark each decision's authority
Advise, recommend-and-approve, act-within-boundary, or act-on-material. Some workflows will need more than one tier.
Name the approver
A person, not a role. For every decision that requires approval, someone owns it by name.
State the escalation rule
What makes the agent stop and ask. Confidence threshold, boundary breach, disagreement between sources.
Define the metric
What moves, measured against a baseline captured before go-live. If you cannot name it, it is not delegable yet.
Agentic operations is the right first surface
Not because it is the most valuable work in the enterprise, but because it is the work where a first agent can be given real authority, be measured against a real cost, and produce evidence fast enough to earn the second agent. Agentic Operations is covered in full on its own page →
The work is volume-driven
Thousands of cases rather than dozens. A single error is visible in the numbers, which is exactly what a first programme needs.
The cost is already measured
Handling time, cost per case, backlog age, first-time-right rate. These exist before you arrive, so improvement is provable without building a measurement system first.
The exception path is knowable
Operations work has a natural escalation case, which is where the human-in-the-loop design starts — rather than the whole workflow being automated blind.
The feedback loop is short
Outcome arrives within days, so the evaluation set, the thresholds and the prompts can be corrected while the programme still has momentum and budget.
Typical first surfaces we have taken on: order exception and reconciliation handling, customer case triage and routing, claims intake and document completeness checking, vendor query resolution, service request entitlement checking, and finance month-end matching support. The selection rule is the same in every case.
The selection rule
A first workflow is right when five statements are all true. If one is false, the workflow will produce an interesting demonstration and nothing else.
Cases are similar enough to be recognised and different enough to require judgement. Identical work is a rules engine problem; wholly unique work is a research problem.
You can name what one case costs in handling time, rework and delay. If no one has ever tried, the programme has no baseline and no way to prove value.
The agent can be wrong without a customer, regulator or counterparty being harmed. Reversibility is what makes a controlled first step responsible.
Policy, reference data and precedent are written down and current. Where they are not, the first deliverable is the knowledge work, not the agent.
One named person accountable for the outcome, with the authority to change the process. This is the single strongest predictor of whether a pilot survives its first quarter.
One scorecard, two populations
The temptation is to measure agents by human metrics or humans by agent metrics. Both are wrong, and both produce performance conversations that go nowhere.
- Volume and throughput of completed tasks
- Accuracy against the evaluation threshold you set
- Escalation rate — and whether it is rising or falling
- Override rate by human approvers, per approver
- Cost per task, and the trend
- Failure clusters and whether they are reducing
- Time to resolution on the tasks it owns
- Judgement quality on the decisions they own
- Throughput on the work only they can do
- Escalation judgement — did they catch it
- Approval quality: how often they accepted what they should have questioned
- Capability growth, evidenced in the LMS
- Failure clusters in their own work
- Whether the team around them got faster
Sales is not a chatbot with a product catalogue attached
The interesting work in sales is not answering questions about pricing. It is assembling, checking and maintaining a qualified pipeline, and doing so on a cycle faster than a human team can sustain. Agentic sales means redesigning that cycle around what the work actually requires, then giving the agent real authority over the parts of it that are repetitive and measurable.
Qualification becomes a classification problem with an audit trail
Historically, qualification is a senior person forming an impression from a call. The agentic version is not to remove the judgement but to make the criteria explicit: fit against the defined ICP, urgency signals, budget presence, and a documented disqualifier list. The agent prepares a fully evidenced account — account background, prior interactions, relevant events, open questions — and states its recommendation with a confidence and its reasoning. The human decides, and the decision is recorded against the criteria rather than against a feeling.
- Measured on — precision of the recommendation against the human decision, and the false-negative rate on the leads a human later promoted.
- Escalates on — any account with contract value above the band, any account flagged as a named competitor relationship, any account with an unresolved complaint.
The agent owns the queue, not the relationship
The unit of work here is "prepare a qualified account for human review", not "sell to the customer". The customer-facing conversation stays with a person. What the agent removes is the preparation tax that consumes the first hour of every rep's day and the notes that never get written afterwards.
Research and proposal work is exactly what a work platform is for
Account research, competitor material review, prior proposal retrieval, the technical appendix, the first draft of the deck, the internal compliance check on claims, and the handover pack for the solution architect. Each of these is a bounded unit with an acceptance test, and most of them are currently done from memory, from a shared drive nobody trusts, and from a document that gets rewritten from scratch for every deal.
- Measured on — preparation hours per opportunity, draft reuse rate, number of claims in proposals that require legal review, time from discovery call to first proposal.
- Escalates on — any pricing construct, any non-standard clause, any commitment to a delivery date the agent inferred rather than read.
Retrieval discipline is the whole game
Proposal quality collapses when the agent retrieves the wrong version of a capability statement. Winning content has to be retrievable, permissioned, versioned and traceable — which is why the ontology and context work sits upstream of this stage rather than after it.
Business semantics work →The unglamorous work is where the margin is
Pipeline hygiene is pure operations: matching CRM records to the contract system, identifying opportunities with no next step, flagging close dates that have no supporting activity, chasing missing documentation, reconciling the forecast that was built by hand last month. It is repetitive, high-volume, embarrassing to do manually, and it changes the number the whole function is measured on.
- Measured on — forecast accuracy variance, number of opportunities with no next step, data completeness at stage entry, admin hours per manager per week.
- Escalates on — any forecast change that reduces the committed number, and any opportunity where the CRM record and the contract record disagree about terms.
Start here, not in the demo
Most enterprises begin their sales agentic ambition with the customer-facing experience, because that is the visible part. We usually start with pipeline hygiene, because it is where the delegation is unambiguous, the cost is measurable, and nothing irreversible reaches a customer.
Four levels of organisational readiness
Use this as a structured self-assessment, not a marketing score. Most organisations that believe they are at level three are at level one. The honest answer in each column is worth more to your board than any maturity label.
| Dimension | L1 — Tooling | L2 — Governed assistance | L3 — Delegated workflow | L4 — Agentic operating model |
|---|---|---|---|---|
| Work design | Tools adopted per team, work left as written | Selected tasks documented, still framed as tasks | Work decomposed into units with mandates | Work is designed around what humans and agents each hold |
| Authority | None; suggestions only, inconsistently followed | Advisory tier, no formal tier model | Tiers defined per unit, with expiry and revocation | Authority allocation is a portfolio decision reviewed by a committee |
| Supervision | Ad hoc; users escalate informally | Support queue, no calibrated review sampling | Sampled review tuned to risk, exception routing designed | Supervision is staffed as a function with its own tooling and metrics |
| Business semantics | Terms differ by team; no shared definition | Glossary exists, not enforced or used by systems | Governed ontology consumed by agent tools and evaluation | Semantic layer is a maintained asset with change governance |
| Evaluation | Ad hoc impressions of quality | Sample-based review, no regression set | Regression set runs on every release, bad cases feed back | Continuous quality inspection on live traffic with trend reporting |
| Measurement | Adoption and usage counts | Time saved estimated by the team itself | Agent outcome measured against the same baseline as the human role | Human and agent performance reported on one scorecard |
| Capability | Learning is optional and self-directed | Central team holds all the knowledge | Business teams own their agents, central team enables | Internal roles build, evaluate and govern without external dependency |
| Governance | No inventory; agents are invisible to IT | Inventory exists, no classification or evidence | Inventory, tiering and evidence, reviewed on a cycle | Management system certified or audit-ready, incidents rehearsed |
L1, with an L2 pilot running
Broad tool adoption, one promising pilot, no authority model. The pilot either gets a business owner or it quietly stops after renewal.
L4 claimed, L2 delivered
An ambition statement describes an agentic enterprise; the estate is a handful of assistants with no mandate, no evaluation and no inventory.
Two levels in one process
One workflow running at T1 with real authority and measured outcomes, and one at T0 with human decisions. That combination builds belief faster than any roadmap.
One scorecard, two populations
Human Machine Capital only means something if it produces a single view of capability. We put agent outcomes and human outcomes on the same rows, reviewed by the same committee on the same cycle, and we add two measures that only agents carry: their escalation precision and their confidence calibration.
The reason is practical. If a leadership team reviews two separate dashboards, the comparison is never made and every conversation becomes a debate about which number looks better. One table makes the trade-off visible: where the agent is genuinely better than the person, where it is worse, and where the honest answer is a different split of the work.
The uncomfortable metric we insist on. A redeployment rate. If agents free up human capacity that is never redeployed into better work, the programme has cut cost and created nothing. It is easy to omit and hard to defend against — which is why it is in the standard report.
Agent and human on the same measure
| Measure | Agent | Human team | Read together |
|---|---|---|---|
| Outcome quality | Cases resolved to standard | Cases resolved to standard | Where the gap sits |
| Cycle time | Median and 90th percentile | Median and 90th percentile | Tail behaviour matters more |
| Cost per unit of work | Inference, retrieval, review time | Handling time, rework | Net of supervision cost |
| Escalation precision | Share of escalations worth a human's time | — | Agent-only measure |
| Confidence calibration | Stated confidence vs correctness | Self-rated certainty vs outcome | Agent-only, comparable in spirit |
| Exception rate and trend | Share of work escalated | Share of work needing rework | Direction over level |
| Redeployment rate | Capacity released, redeployed | Hours released, redeployed | Whether value was created |
Four levels of organisational readiness
We assess this honestly, including when the answer is zero. An organisation that believes it is at level 2 when it is at level 1 will buy the wrong first project.
| Level | What it looks like | What is missing | Time to next level |
|---|---|---|---|
| 0 — Exploring | P curiosity, isolated pilots, no production workflow | Nobody owns the estate; no workflow is documented well enough to delegate | 6–12 months |
| 1 — Structured | A diagnostic has been run; candidate workflows are ranked and scored | The evaluation harness and the learning system have not been built | 3–6 months |
| 2 — Operating | One workflow in production with a named approver and a decision log | The ontology is thin; the evidence model is manual; partners are not enabled | 3–6 months |
| 3 — Institutional | Multiple workflows, an owned portfolio, evidence accepted by risk, partners trained | Cross-region consistency and retirement discipline | Continuous |
Who owns the agent portfolio, and who is allowed to decide what
Governance fails when everyone is responsible. We design three bodies with non-overlapping mandates, a standing agenda, and a decision log — and we insist the business owner is not the technology owner, because a team that both builds and authorises an agent will authorise it too generously.
| Decision | Decides | Consulted | Informed |
|---|---|---|---|
| Which workflow enters the estate | Business owner | Central AI enablement team | Steering committee |
| Authority tier for a unit | Business owner, ratified by steering committee | Risk and control function | Technology operations |
| Re-authorisation and expiry | Steering committee | Business owner, risk function | Agent builder |
| Build or buy for a capability | Central AI enablement team | Business owner, architecture | Procurement |
| Release of a new agent version | Agent owner | Evaluation, security | Business owner |
| Suspension of a live agent | Any single named authority, unilaterally | — | All parties, same day |
| Escalation of a recurring failure | Steering committee | Business owner, technology operations | Executive sponsor |
Three bodies, three mandates
Executive sponsors from the business, risk, technology and finance. It owns the estate: what may be built, at what authority tier, with what capital. It meets on a fixed cycle and keeps a decision log. It is deliberately small, and it is chaired by a business executive rather than by IT.
A small standing team that owns the shared assets: the semantic layer, the evaluation harness, the context pipeline, the tool catalogue, the deployment patterns and the training curriculum. It publishes the paved road so business teams build on it rather than around it.
Named individuals inside the business who own an agent's outcome day to day. They review exceptions, tune thresholds, approve releases and are the human escalation destination. Without this role, supervision degrades into an unstaffed inbox.
Segregation of duties
The team that configures an agent must not be the team that authorises it, and neither should be the team that audits its evidence.
Re-authorisation as a ritual
Every mandate has a review date, and lapsing it drops the tier. Re-authorisation becomes a scheduled committee item rather than a decision nobody remembers making.
A kill switch that is tested
Suspension is exercised during a rehearsal each quarter, and the exercise is recorded. An untested kill switch is a hypothesis, not a control.
Who owns the agent portfolio, and who is allowed to decide what
An agent with no named owner is a liability nobody can escalate to. This is the governance map we insist on before the first agent is built.
| Task | R | A | C | I |
|---|---|---|---|---|
| Set the agentic strategy | Consulting + executive sponsor | Board or C-suite | Function heads, risk | All |
| Decide which workflows to delegate | Function owner | Function head | Risk, technology | Executive |
| Own an individual agent in production | Named agent owner | Function head | Technology, risk | Supervisors |
| Approve an agent's authority tier | Agent owner | Function head | Risk | Technology |
| Approve live actions | Operations supervisor | Operations manager | Agent owner | Risk |
| Define the ontology | Data & ontology team | Chief data or technology officer | Business owners | All |
| Own the learning and skills system | L&D lead | CHRO | Function heads, supervisors | All |
| Approve go-live | Agent owner + risk | Function head | Technology, L&D | Executive |
| Retire an agent | Agent owner | Function head | Risk, technology | All |
A sequence that produces evidence before it produces ambition
The order matters more than the content. We front-load the things that make the business safer and the organisation more capable, because they are what makes the third and fourth agents possible at all.
| Period | Outcome | What is actually done | How we judge it |
|---|---|---|---|
| Months 1–2 | The work is understood | Two weeks inside a candidate process; the delegable units, their inputs, action sets and acceptance tests are written down; the estate is inventoried. | One page per unit that a supervisor could hand to a builder without a meeting. |
| Months 3–4 | Authority and semantics exist | Tier model applied to each unit; the business terms the workflow depends on are modelled and owned; data contracts are agreed with the data team. | Every unit has a tier, an owner, a mandate and an expiry date. |
| Months 5–7 | One agent in shadow, then live | The agent runs in shadow alongside people for a defined period, its decisions compared case by case; thresholds are tuned; it then goes live at a tier its evidence supports. | Agreement rate with the human decision, escalation precision, and a clean audit trail from day one. |
| Months 8–9 | Supervision is staffed | Supervisors appointed, exception queue and response expectation operating, review sampling tuned, kill switch rehearsed and recorded. | Every exception has an owner and a response time; the rehearsal is documented. |
| Months 10–11 | Capability transfers | Internal roles trained on build, evaluation and operation; the semantic layer, evaluation harness and context pipeline are handed to the internal team with runbooks. | Internal team ships a change to the agent without us in the loop. |
| Months 12 | The committee decides the next portfolio | A full operating review on the shared scorecard, a re-authorisation of every mandate, and a decision on the next two or three units — taken by the business, not by us. | Two or three more units enter the estate with a named owner and a measured baseline. |
Weekly during the pilot
Exception review with the agent owner, thresholds adjusted in the open, bad cases pulled into the evaluation set the same week they are found.
Monthly across the estate
The shared scorecard, one table, agent and human results side by side, with the committee in the room and a decision log at the end.
Quarterly on authority
Every mandate re-authorised or allowed to lapse. Killing a mandate is treated as a normal governance act, not an incident.
A sequence that produces evidence before it produces ambition
This is the difference between an agentic business and an agentic pilot portfolio. The order is deliberate and we will not reorder it for a faster start.
One workflow, in production, with a decision log and a measured result. This is the step every organisation skips and every organisation that skips it regrets.
Now that you have one real agent, the ontology gaps become concrete and arguable. Before this, they are theoretical and get deferred.
Operators are already working with the agent. Now build the curriculum that scales that to the next twenty people — and to partners.
Multiple workflows, a governed portfolio, authority tiers reviewed rather than assumed. This is where the board's interest becomes well-founded.
The organisation can originate, scope and govern its own agentic work. Our role shrinks deliberately. That is the goal, not a loss of account.
Where agentic business is easiest to establish
Our own 0–3 assessment by industry. Readiness dominates — an industry with documented work converts faster regardless of how regulated it is.
| Board mandate | Workflow documentation | Data readiness | Risk appetite | Learning capacity | |
|---|---|---|---|---|---|
| Financial Services | |||||
| Banking & Digital Banking | |||||
| Insurance & Insurtech | |||||
| Asset & Wealth Management | |||||
| Manufacturing & Smart Factory | |||||
| Automotive & Mobility | |||||
| Logistics, Supply Chain & Freight | |||||
| Retail & Consumer Goods | |||||
| Healthcare & Hospitals | |||||
| Pharmaceuticals & Life Sciences | |||||
| Telecoms & Technology | |||||
| Media, Entertainment & Publishing | |||||
| Energy & Utilities | |||||
| Water, Waste & Environmental Services | |||||
| Construction & Real Estate | |||||
| Public Sector & Government | |||||
| Education & Higher Education | |||||
| Professional Services & Legal | |||||
| Hospitality, Travel & Leisure | |||||
| Accounting, Audit & Advisory |
What we do not do, and will not sell you
Consulting firms earn trust by being specific about the work they will not take. These are the boundaries we hold, and they are enforced on ourselves as much as on the client.
- We do not run a programme with no business owner. If no named person owns the outcome, we decline. A steering committee is not a substitute for an accountable individual, and a pilot without an owner is an unfunded demonstration.
- We do not treat buying a platform as the transformation. We are authorised to deploy the technology, but the platform is the least durable part of the work. If the engagement is scoped as installation plus enablement, we will say so rather than sell a strategy we do not believe in.
- We do not promise autonomy on a fixed date. Authority tier is earned by evidence, and the schedule to earn it is set by the evaluation, not by the project plan. We will not commit a tier in advance of the evidence that supports it.
- We do not take on the work itself permanently. Every engagement is designed to leave internal roles able to build, evaluate, operate and govern without us. We will tell you when a dependency on us is still real, and we will name the date it should end.
- We do not extend beyond our competence. We do not provide legal advice, we do not certify, and we do not sign off regulatory conclusions. We build the systems and the evidence that your counsel and auditors need — and we say clearly where that line sits.
This is the spine; the rest is the apparatus
Business semantics
Modelling the entities, states, rules and decisions the delegated units depend on.
Runtime and context
The retrieval, tool, evaluation and observability layer that makes delegation dependable.
Capability inside the organisation
Building the roles that design, evaluate and supervise the agent estate.
Governance and evidence
Making the authority model defensible to an auditor, a regulator and a customer.
Training and capability
The curriculum that turns the people running the business into the people running the agents.
What we do not do, and will not sell you
We are not a platform reseller and we will not recommend a platform based on a partner agreement. We do not run an AI transformation programme that leaves your supervisors unable to do their job. And we will not take on a twenty-workflow transformation as a first engagement, because the first workflow has to teach you something and twenty will not.
Four levels of readiness is the honest answer to "are we ready?"
Four yes-or-no answers. All four yes means start with agentic operations. Any no means the first project is about readiness, not agents — and we will say so.
- You can name a workflow, not a department.
- That workflow is written down as policy rather than tribal knowledge.
- A human can be named who would be accountable if it were wrong.
- You have historical cases to build an evaluation set from.
What this page answers
Every topic below is covered in the sections above, and each links to the page that goes deeper on it.
Bring us one workflow you think is not delegable
A 45-minute working session with a practice lead. We will map it against the selection rule, tell you plainly whether it is a good first agent, and what it would take to run one safely. No slides, no assessment score, no obligation.